面向工业过程的以目标为中心的因果发现合成基准

arXiv cs.LG 论文

摘要

本文提出了一种面向工业批次过程的目标中心因果发现合成基准,该基准基于一个两层结构因果模型构建,其各个组成部分允许逐项调节条件以控制难度。研究评估了八种方法,误差分析表明,随着多种困难因素叠加,这些方法的性能退化方式与其各自的假设一致。

arXiv:2611.17556v1 Announce Type: new Abstract: Causal discovery promises to find which recorded variables drive a process target such as product quality, yield or energy use. Process data with known causes are rare, and in existing benchmarks the difficulty of identifying the causes of a target is either fixed by the recorded data or set for the whole graph. We present a synthetic benchmark for target-centric causal discovery in industrial processes. Its generator is a two-level structural causal model inspired by batch production, assembled from components that each model one phenomenon of process data, such as a control loop or a delayed effect. The graph is built around the target: without sensor error, every non-cause can be distinguished from the direct causes by conditioning on them or by time order. Each component is set independently, so the difficulty can be varied one condition at a time. We evaluate eight methods, from association measures to effect estimation, with an error analysis by the role of the selected candidate; all degrade as the difficulties combine, each in a way consistent with its assumptions.
查看原文
查看缓存全文

缓存时间: 2026/11/16 08:44

# A Synthetic Benchmark for Target-Centric Causal Discoveryin Industrial Processes
Source: [https://arxiv.org/html/2611.17556](https://arxiv.org/html/2611.17556)
###### Abstract

Causal discovery promises to find which recorded variables drive a process target such as product quality, yield or energy use\. Process data with known causes are rare, and in existing benchmarks the difficulty of identifying the causes of a target is either fixed by the recorded data or set for the whole graph\. We present a synthetic benchmark for target\-centric causal discovery in industrial processes\. Its generator is a two\-level structural causal model inspired by batch production, assembled from components that each model one phenomenon of process data, such as a control loop or a delayed effect\. The graph is built around the target: without sensor error, every non\-cause can be distinguished from the direct causes by conditioning on them or by time order\. Each component is set independently, so the difficulty can be varied one condition at a time\. We evaluate eight methods, from association measures to effect estimation, with an error analysis by the role of the selected candidate; all degrade as the difficulties combine, each in a way consistent with its assumptions\.

###### keywords

causal discovery; time series; benchmark; root\-cause analysis; batch processes; structural causal models

## 1Introduction

Many industrial processes run as repeated episodes\. In batch production, a vessel is filled, heated, held while the reaction runs, cooled and emptied, and the same recipe is run hundreds of times; in every batch, tens of process variables are logged minute by minute, together with a target of interest, such as the product quality, the yield or the energy use of the batch\. After an off\-target batch, the engineer needs the recorded variables that act on the target directly, since a variable merely correlated with the target is no lever for intervention\. This is local causal discovery\([Aliferis et al\., 2010](https://arxiv.org/html/2611.17556#bib.bib23);[Yu et al\., 2020](https://arxiv.org/html/2611.17556#bib.bib24);[Ling et al\., 2025](https://arxiv.org/html/2611.17556#bib.bib46)\)on time series observed over many episodes; root cause analysis is a main application of causal discovery in manufacturing\([Vuković and Thalmann, 2022](https://arxiv.org/html/2611.17556#bib.bib22)\)\. Causal discovery from time series, however, is usually evaluated on the graph over all variables; how well a method finds the direct causes of one target is a different question, and current benchmarks are not built to answer it\.

Benchmarks built on real data or fitted to it\([Cheng et al\., 2024](https://arxiv.org/html/2611.17556#bib.bib1);[Stein et al\., 2025](https://arxiv.org/html/2611.17556#bib.bib2)\)are realistic, but few plants release data, and those from industrial processes\([Menegozzo et al\., 2022](https://arxiv.org/html/2611.17556#bib.bib6);[Mogensen et al\., 2024](https://arxiv.org/html/2611.17556#bib.bib5);[Mehling et al\., 2026](https://arxiv.org/html/2611.17556#bib.bib45)\)come with expert graphs, some only between subsystems, whereas simulated benchmarks\([Herdeanu et al\., 2025](https://arxiv.org/html/2611.17556#bib.bib8);[Göbler et al\., 2024](https://arxiv.org/html/2611.17556#bib.bib44)\)know their graph exactly\. Any of them allows a target to be chosen and its direct causes scored, but the conditions that make those causes hard to find, how strong they are, what disturbs them and whether a controller acts on them, are fixed in the recorded datasets and, in the generators, set for the whole graph instead of for the chosen target\.

We present a synthetic benchmark built around the target\. Its generator is a structural causal model inspired by batch production, assembled from components; each is a causal mechanism whose effect on the task is known and is set by one parameter, so the difficulty of finding the causes of the target can be varied one condition at a time \(Section[4](https://arxiv.org/html/2611.17556#S4)\)\. Our contributions are \(i\) a*target\-centric benchmark*: the task, the roles of the candidates \(direct cause, ancestor, sibling or descendant\) and the difficulty are all defined around one target \(Sections[3](https://arxiv.org/html/2611.17556#S3)and[5](https://arxiv.org/html/2611.17556#S5)\); \(ii\) a*configurable generator*of batch plants \(Section[4](https://arxiv.org/html/2611.17556#S4)\); and \(iii\) an*evaluation*of eight methods, from association measures to effect estimation, with an error analysis \(Sections[5](https://arxiv.org/html/2611.17556#S5)to[6](https://arxiv.org/html/2611.17556#S6)\)\.

## 2Related work

Surveys review temporal causal discovery\([Gong et al\., 2024](https://arxiv.org/html/2611.17556#bib.bib12);[Runge et al\., 2019a](https://arxiv.org/html/2611.17556#bib.bib13)\);[Assaad et al\. \(2022\)](https://arxiv.org/html/2611.17556#bib.bib11)also evaluate methods on simulated systems and a few real datasets\. CausalTime fits neural generative models to real series and derives the graph from the fitted model\([Cheng et al\., 2024](https://arxiv.org/html/2611.17556#bib.bib1)\); CausalRivers provides river\-discharge networks with graphs built from metadata\([Stein et al\., 2025](https://arxiv.org/html/2611.17556#bib.bib2)\); CausalDynamics\([Herdeanu et al\., 2025](https://arxiv.org/html/2611.17556#bib.bib8)\)generates coupled differential equations with noise, confounding and lags; causal chambers are a physical testbed with controllable inputs\([Gamella et al\., 2025](https://arxiv.org/html/2611.17556#bib.bib9)\); CausalProfiler samples tabular benchmarks from explicit assumptions about models, queries and data\([Panayiotou et al\., 2026](https://arxiv.org/html/2611.17556#bib.bib4)\); causalAssembly generates semi\-synthetic per\-part data of an assembly line from a graph set by engineers and mechanisms fitted to recorded data, scored on the full graph\([Göbler et al\., 2024](https://arxiv.org/html/2611.17556#bib.bib44)\)\.[Lawrence et al\. \(2020\)](https://arxiv.org/html/2611.17556#bib.bib7)argue, as we do, for a data\-generating process with explicit knobs\. For process data, Tennessee Eastman\([Downs and Vogel, 1993](https://arxiv.org/html/2611.17556#bib.bib18);[Reinartz et al\., 2021](https://arxiv.org/html/2611.17556#bib.bib19)\)and IndPenSim\([Goldrick et al\., 2015](https://arxiv.org/html/2611.17556#bib.bib20);[Goldrick et al\., 2019](https://arxiv.org/html/2611.17556#bib.bib21)\)are the standard simulated continuous and fed\-batch processes, CIPCaD\-Bench pairs data simulated from the former with real data from a food\-processing line\([Menegozzo et al\., 2022](https://arxiv.org/html/2611.17556#bib.bib6)\), the cryoplant of the European Spallation Source provides real plant data with an expert graph between subsystems\([Mogensen et al\., 2024](https://arxiv.org/html/2611.17556#bib.bib5)\), causRCA provides lathe recordings with an expert graph for root\-cause analysis\([Mehling et al\., 2026](https://arxiv.org/html/2611.17556#bib.bib45)\), and root\-cause benchmarks exist for software systems\([Hardt et al\., 2024](https://arxiv.org/html/2611.17556#bib.bib14)\); each provides a fixed set of systems and graphs\. Our benchmark differs in three respects: the task is defined for one target, difficulty is set through variance\-based controls relative to that target, and every candidate of every plant carries its role with respect to that target, in addition to the full graph\.

## 3The task

A process is observed overBBepisodes, repeated runs of the same process such as the batches of a batch plant\. In each episode,ppvariablesX1,…,XpX\_\{1\},\\dots,X\_\{p\}and a targetYYare recorded atTTtime steps; the other variables of the process are latent, as a plant logs its process variables but not every condition that acts on them\. The task is to find the*direct causes*DDofYYamong the candidates𝒞\\mathcal\{C\}: the candidates with an edge intoYYin the latent projection of the causal graph onto the recorded variables, in which an edge joins two variables when one acts on the other directly or only through latent variables\. Their effect onYYmay be delayed and nonlinear\. A method returns a ranking of all candidates and a selected list, which may be empty\.

Each candidate has one role for the targetYY, read on the generating graph𝒢\\mathcal\{G\}\(Section[4\.1](https://arxiv.org/html/2611.17556#S4.SS1)\)\. It is a*direct cause*; an*ancestor*, which influencesYYonly through other recorded variables; a*descendant*, whichYYinfluences; or a*sibling*, which shares a latent cause with a direct cause and is correlated withYYwithout affecting it\. The roles partition the candidates by their ancestral relation toYY, which the latent projection preserves, so they are defined by𝒢\\mathcal\{G\}and the set of recorded variables, as direct causation is in any causal graph over a variable set\([Spirtes et al\., 2000](https://arxiv.org/html/2611.17556#bib.bib35)\)\. In our generator, every candidate that is not a direct cause also shares a latent cause withYY, so every wrong candidate is associated withYY\. We count only the direct causes as correct answers; ancestors, although they influenceYY, count as wrong in our evaluation\. We release the role of every candidate, so that other scorings are possible\. Without sensor error, every wrong candidate is either separated from the target given its direct causes or distinguished by time order \(Appendix[B](https://arxiv.org/html/2611.17556#A2)\)\. These labels describe the generating graph; they do not guarantee recovery of the direct causes from observational data\.

## 4The generator

The generator is a structural causal model \(SCM\) whose variables are defined at two levels, one value per batch and one value per time step, and whose mechanisms are built from components inspired by batch production \(Table[1](https://arxiv.org/html/2611.17556#S4.T1), Figure[1](https://arxiv.org/html/2611.17556#S4.F1)\)\. To fix ideas, we illustrate the components with a generic batch process: raw materials are charged into a vessel, transformed under controlled conditions such as temperature and mixing, and the product is then separated and dried downstream; the target is an outcome of the batch, e\.g\. the quality of the product, the yield or the energy use\.

Figure 1:How the generator produces data\. \(1\) The causal graph of a plant and the role of each recorded variable forYY\. \(2\) The profile ofX3X\_\{3\}in three batches, built from its profile parametersθ⁡\(X3\)\\theta\(X\_\{3\}\)\. \(3\) One batch: time\-step links, control and the target\.Table 1:Core components of the generator\.ComponentIn the generatorIn a batch plant*Once per batch*batch profileθ⁡\(Xi\)\\theta\(X\_\{i\}\)2–4 motifs per variable \(step, ramp, spike, plateau\); their onsets, durations, sizes and the level formθ⁡\(Xi\)\\theta\(X\_\{i\}\)a temperature or flow rises, holds and falls in every batch, at different times and by different amountsbatch conditionsUjU\_\{j\}act on the profiles of many variablesconditions that differ between batches and affect many sensors, e\.g\. coolant or outside temperatureconfoundersZkZ\_\{k\},Zk​\(t\)Z\_\{k\}\(t\)move a direct cause and its siblings together, also within the batchan unmeasured property of a raw material, e\.g\. its purity, that moves several sensors togetherbatch\-level linksbetween profile parameters, without lag; seen only across batchesone batch quantity sets another, e\.g\. a larger charge gives a higher temperature peak*At every time step within a batch*time\-step linksFτ,ϑ\\mathrm\{F\}\_\{\\tau,\\vartheta\}first\-order response with a dead timea change propagates to another variable after a delay, e\.g\. the pressure in the vessel follows its temperaturecontrollerκ\\kappaholds a direct cause by moving an actuator \(not a candidate\)a control loop moves a valve, e\.g\. the coolant valve to hold the vessel temperaturetarget responsegkg\_\{k\},sints\_\{\\mathrm\{int\}\}linear, saturating, optimal or threshold; joint effects of pairs of causesquality improves and then levels off, is best at one temperature, or drops past an impurity limit; some causes matter only together, e\.g\. temperature and the amount of a raw materialunexplained partssigs\_\{\\mathrm\{sig\}\}a slow part ofYYno recorded variable explainscauses that are not measureddownstream measurementsrespond toYYwith a delaymeasured after the product is formed, e\.g\. a pressure in the separation step or the moisture after drying### 4\.1A two\-level structural causal model

The generator draws an SCM, a causal graph𝒢\\mathcal\{G\}with an assignment for each of its nodes, which is then fixed; we call one such SCM a*plant*\. Its structure comprises the motifs of every recorded variable, the graph and its weights, and the direct causes ofYYwith their response functions\. Its data areBBbatches, each a sample of the SCM with newly sampled exogenous noise\.

A batch plant produces data on two time scales\. Each batch is run from a recipe and leaves a batch record, which lists, once per batch, the amounts charged, the lots of the raw materials, and the setpoints and duration of each phase\. During the batch, sensors log trajectories at a fixed interval\. The recipe prescribes the same sequence of phases for every batch, e\.g\. a heat\-up, a hold and a cool\-down, so the trajectories recur from batch to batch; they differ in when each part starts, how long it lasts, how high it rises and sometimes whether it occurs at all\. Causal effects act on both time scales\. Some act from one moment to the next: the pressure in the vessel follows its temperature within minutes, in every batch\. Others act between properties of the whole batch, e\.g\. a larger charge can give a higher temperature peak; these appear most clearly when batches are compared, and batch data are commonly analysed by comparing batches\([Nomikos and MacGregor, 1994](https://arxiv.org/html/2611.17556#bib.bib42)\)\.

The generator reproduces this structure with an SCM whose variables are defined at two levels\.*Batch\-level*variables take one value per batch: the latent batch conditionsUjU\_\{j\}, the confoundersZkZ\_\{k\}, the phase timingsΦq\\Phi\_\{q\}, and the profile parametersθ⁡\(Xi\)\\theta\(X\_\{i\}\)of every recorded variable, i\.e\. the onset, duration and size of each motif of its profile and its level\.*Time\-step*variables take one value per time stept=1,…,T\{t=1,\\dots,T\}of the batch: the recorded variablesXi​\(t\)X\_\{i\}\(t\), the targetY⁡\(t\)Y\(t\), the actuator of the controller and the within\-batch partZk​\(t\)Z\_\{k\}\(t\)of each confounder\.

The generator follows the order of a plant: in each batch, the batch level is drawn first, with new noise\. A batch is generated in five steps \(Figure[1](https://arxiv.org/html/2611.17556#S4.F1)\): \(i\) the batch conditions and confounders are drawn; \(ii\) they set the profile parameters, each of which sets the timing or size of one motif of its own variable, e\.g\. the start of the heating ramp; \(iii\) the profiles are built from these parameters and the within\-batch part of each confounder is added; \(iv\) the time\-step links and the controller act within the batch; \(v\) the target is computed from its direct causes, followed by the downstream measurements\.

The two levels give two kinds of causal links between recorded variables\. A*batch\-level link*acts when the profiles are set: a parameter of one variable, e\.g\. the size of the charging step of the feed, is a parent of a parameter of another, e\.g\. the height of the temperature peak\. Each batch gives a single value of each parameter, so the link appears only across batches\. A*time\-step link*acts within the batch: the effect follows the cause after a dead timeϑ\\vartheta, in every batch, and the order in time shows the direction\.

The two kinds of links leave different evidence\.YYresponds to its direct causes through time\-step links, which can be read from the time series\. The siblings, and the candidates linked to a direct cause at the batch level, are instead tied toYYthrough shared batch conditions, confounders and batch\-level links, and these ties appear mostly when batches are compared\. Methods that pool many batches see both\.

The graph𝒢\\mathcal\{G\}, unrolled over the time steps of a batch, is acyclic: the controller and the operator reactions act with a delay of at least one step, so their feedback loops become chains in time, and one batch can be generated by evaluating the nodes in that order \(Algorithm[A](https://arxiv.org/html/2611.17556#A1.SS0.SSS0.Px3)\)\. Batches are independent draws of the SCM unless the additional lots and drift are switched on; these add batch\-level variables linked from one batch to the next \(Section[4\.4](https://arxiv.org/html/2611.17556#S4.SS4)\)\. A recorded variable is the group of its nodes,\{θ⁡\(Xi\)\}∪\{Xi​\(t\)\}t\\\{\\theta\(X\_\{i\}\)\\\}\\cup\\\{X\_\{i\}\(t\)\\\}\_\{t\}, and its role forYY\(Section[3](https://arxiv.org/html/2611.17556#S3)\) is read on𝒢\\mathcal\{G\}; the methods receive only the trajectoriesXi​\(t\)X\_\{i\}\(t\)andY⁡\(t\)Y\(t\)\. Every plant is released with its data, candidate roles, separation or time\-order labels for non\-causes, the unique share of every direct cause and its full graph𝒢\\mathcal\{G\}\(Appendix[B](https://arxiv.org/html/2611.17556#A2)\)\.

### 4\.2Mechanisms

#### Batch level\.

*Batch conditions\.*Each plant hasmmlatent batch conditionsU1,…,UmU\_\{1\},\\dots,U\_\{m\}\(e\.g\. the temperature of the coolant or of the surroundings\), linked by a random directed tree in which each condition acts on one other\.

*Confounders\.*Each direct causeXkX\_\{k\}shares a latent confounderZkZ\_\{k\}\(e\.g\. the purity of a raw material\), drawn independently in each batch, with the non\-causes assigned to it \(every non\-cause is assigned to one direct cause\)\. It acts in two stages\. Once per batch,ZkZ\_\{k\}enters their profile parameters, so the motifs of a cause and its siblings start, last and rise together\. Its slow componentZk​\(t\)Z\_\{k\}\(t\)is then added to their trajectories, so they also move together from one time step to the next\.ZkZ\_\{k\}reachesYYonly throughXkX\_\{k\}\. The non\-causes that reachYYby no other route, the*siblings*, are therefore confounded withYY: they move withXkX\_\{k\}without affectingYY\.

*Profile parameters\.*Each parameterθ\\thetais the mean of its parents plus Gaussian noise, mapped into its admissible range \(Equation[4](https://arxiv.org/html/2611.17556#A1.E4), Appendix[A](https://arxiv.org/html/2611.17556#A1)\)\. Its parents are batch conditions, confounders, phase timings \(Section[4\.4](https://arxiv.org/html/2611.17556#S4.SS4)\) and parameters of other recorded variables; the weights of the confounders, phases and batch\-level links are instead set by their shares \(Section[4\.3](https://arxiv.org/html/2611.17556#S4.SS3)\)\.

#### Time\-step level\.

*Profiles\.*Each recorded variableXiX\_\{i\}follows a profile that recurs in every batch, built from a small library of*motifs*: a step, a ramp, a spike and a plateau\. A variable has two to four motifs, each up or down; their shapes, directions and order are drawn once per plant and then fixed\. The batch is divided into consecutive slots, one per motif, and in each batch the motif starts at an onset within its slot, lasts a duration and has a size; these, with the level, are the profile parameters\. The profile is the level plus the sum of the motifs, plus a smooth random texture and white noise \(Appendix[A](https://arxiv.org/html/2611.17556#A1)\)\. Other families of profiles could replace the motifs without changing the rest of the generator, e\.g\. compositions of Gaussian\-process kernels as in CauKer\([Xie et al\., 2026](https://arxiv.org/html/2611.17556#bib.bib10)\)\.

*Time\-step links\.*Within a batch, a variable responds to its upstream variables with a dead time and a lag, e\.g\. the pressure in the vessel follows its temperature\. We use the first\-order\-plus\-dead\-time response that is standard in process control: the input is delayed by the dead timeϑ\\varthetaand then smoothed exponentially with time constantτ\\tau,Fτ,ϑ​\[x\]​\(t\)=y⁡\(t\)\\mathrm\{F\}\_\{\\tau,\\vartheta\}\[x\]\(t\)=y\(t\)withy⁡\(t\)=y⁡\(t−1\)\+\(x⁡\(t−ϑ\)−y⁡\(t−1\)\)/τy\(t\)=y\(t\-1\)\+\\big\(x\(t\-\\vartheta\)\-y\(t\-1\)\\big\)/\\tau\. A variable with upstream variables receives the sum of their delayed responses, each with its own gain, mixed with its own profile in the proportion set by the upstream sharesups\_\{\\mathrm\{up\}\}\. Recorded variables are ordered \(upstream variables, direct causes, other variables\), and all links, batch\-level and time\-step, point forward in this order; a variable may carry links of both kinds\. This givesYYancestors that act only through its direct causes, and non\-causes that follow the direct causes\.

*Controller\.*The main variables of a plant are under feedback control, which holds them near a setpoint and hides their variation\. The generator puts one direct causeXcX\_\{c\}, e\.g\. the vessel temperature, under control: a recorded actuatorXaX\_\{a\}, e\.g\. the coolant valve, is moved by a control rule so thatXcX\_\{c\}follows its setpoint, the trajectory the recipe prescribes over the batch \(the profile ofXcX\_\{c\}\)\. A recorded upstream variableXdX\_\{d\}, the*disturbance of the controlled cause*, e\.g\. a colder raw material entering the vessel, pushesXcX\_\{c\}away from it\. The rule is integral, with tightnessκ∈\[0,1\]\\kappa\\in\[0,1\]: its correction accumulates the gap between setpoint andXcX\_\{c\}and is added to the actuator’s own profile \(Equation[6](https://arxiv.org/html/2611.17556#A1.E6)\)\. Atκ=0\{\\kappa=0\}the controller is off andXdX\_\{d\}is an ordinary ancestor ofYY, scored as wrong like any ancestor\. Asκ\\kappagrows, the slow part of the disturbance moves fromXcX\_\{c\}to the actuator:XcX\_\{c\}still causesYYbut varies less, so its effect is harder to detect, andXdX\_\{d\}remains a cause ofYY, but the loop cancels its effect, so the data may show no dependence; with the controller on, it is therefore left out of the candidates\. The actuator, a controller output known from the plant’s diagram, is never a candidate\.

#### Target and downstream measurements\.

Each direct causeXkX\_\{k\}acts onYYthrough a delayed response followed by a response function drawn per cause from a library of four,

S⁡\(t\)=1\|D\|​∑Xk∈Dwk​gk​\(Fτk,ϑk​\[X~k\]​\(t\)\),gk∈\{x,tanh⁡x,x2−1,max⁡\(0,x−1\)\},S\(t\)=\\frac\{1\}\{\\sqrt\{\|D\|\}\}\\sum\_\{X\_\{k\}\\in D\}w\_\{k\}\\,g\_\{k\}\\\!\\left\(\\mathrm\{F\}\_\{\\tau\_\{k\},\\vartheta\_\{k\}\}\[\\tilde\{X\}\_\{k\}\]\(t\)\\right\),\\qquad g\_\{k\}\\in\\\{\\,x,\\ \\tanh x,\\ x^\{2\}\-1,\\ \\max\(0,x\-1\)\\,\\\},\(1\)wherez~=\(z−μz\)/σz\{\\tilde\{z\}=\(z\-\\mu\_\{z\}\)/\\sigma\_\{z\}\}uses reference meanμz\\mu\_\{z\}and standard deviationσz\\sigma\_\{z\}, fixed under intervention, andwkw\_\{k\}is the weight of the cause\. The library holds the simplest form of each way a process variable can act on quality: proportional, saturating, optimum\-shaped \(best at one value and worse on either side, so that the cause is nearly uncorrelated with the target\) and thresholded \(no effect until a tolerance is exceeded\); other forms can be added\.SScan also contain interaction terms, products of two direct causes, e\.g\. the vessel temperature acting only together with the amount of a raw material; the interaction share of Section[4\.3](https://arxiv.org/html/2611.17556#S4.SS3)sets how much ofSSthey carry\. The target mixes this signal with a partε\\varepsilonthat no recorded variable explains,

Y⁡\(t\)=ssig​S~​\(t\)\+1−ssig​ε~​\(t\),Y\(t\)=\\sqrt\{s\_\{\\mathrm\{sig\}\}\}\\,\\tilde\{S\}\(t\)\+\\sqrt\{1\-s\_\{\\mathrm\{sig\}\}\}\\,\\tilde\{\\varepsilon\}\(t\),\(2\)whereS~\\tilde\{S\}andε~\\tilde\{\\varepsilon\}are the standardised signal and noise, the signal sharessigs\_\{\\mathrm\{sig\}\}sets the part ofYYexplained by its direct causes \(Section[4\.3](https://arxiv.org/html/2611.17556#S4.SS3)\), andε\\varepsilonvaries on the time scale ofSS\. Downstream measurements influence no direct cause; each receives the delayed responseFτ,ϑ​\[Y\]\\mathrm\{F\}\_\{\\tau,\\vartheta\}\[Y\], mixed with its own profile in the proportion set by the downstream sharesdescs\_\{\\mathrm\{desc\}\}\.

### 4\.3Variance\-based difficulty controls

The generator controls route strength through variance\-based weightss∈\[0,1\]s\\in\[0,1\], called*shares*below\. For trajectory mixtures,a~\\tilde\{a\}andr~\\tilde\{r\}denote the standardised route contributionaaand remainderrr, mixed as

variable=s​a~\+1−s​r~\.\\text\{variable\}=\\sqrt\{s\}\\,\\tilde\{a\}\+\\sqrt\{1\-s\}\\,\\tilde\{r\}\.\(3\)For uncorrelated unit\-variance contributions,ssequals the route’s variance fraction; otherwise it specifies their relative weighting\.

The shares are set along the process, from the inputs of a batch to the downstream measurements \(Figure[1](https://arxiv.org/html/2611.17556#S4.F1), panel 1\)\. A latent confounder, e\.g\. the purity of a raw material, moves a direct cause together with its siblings: the*confounder share*sconfs\_\{\\mathrm\{conf\}\}sets its share of their profile parameters and, relative to the reference, the strength of its within\-batch component \(Appendix[A](https://arxiv.org/html/2611.17556#A1.SS0.SSS0.Px5)\)\. The*upstream share*sups\_\{\\mathrm\{up\}\}sets the part of a variable, e\.g\. a direct cause, explained by its recorded parents\. The*signal share*ssigs\_\{\\mathrm\{sig\}\}weights the causes’ contribution toYY; the*interaction share*sints\_\{\\mathrm\{int\}\}weights products of pairs of causes within that contribution\. Finally,YYdrives the downstream measurements with the*downstream share*sdescs\_\{\\mathrm\{desc\}\}\. Onlyssigs\_\{\\mathrm\{sig\}\}andsints\_\{\\mathrm\{int\}\}shapeYY; the others set how strongly siblings, ancestors and descendants are associated withYY\. In a linear model with independent errors and a single connecting path, correlations multiply along that path\. Additional paths, temporal filtering and nonlinearities affect the observed associations\. Section[5](https://arxiv.org/html/2611.17556#S5)varies these shares, the controller tightnessκ\\kappaand the parameters below, individually or in combination; measured on the generated plants, each share matches its setting \(Table[2](https://arxiv.org/html/2611.17556#S5.T2); Appendix[A](https://arxiv.org/html/2611.17556#A1.SS0.SSS0.Px5)\)\.

### 4\.4Additional components

The remaining components add difficulties typical of batch plants; Appendix[A](https://arxiv.org/html/2611.17556#A1)gives their equations\. Common causes across batches:

- •*Recipe phases*\(sphases\_\{\\mathrm\{phase\}\}\): a latent timing per phase window, drawn once per batch, shifts the onset of every motif in the window, so all variables shift together at the phase boundaries\.
- •*Raw\-material lots*and*campaign drift*\(slots\_\{\\mathrm\{lot\}\},sdrifts\_\{\\mathrm\{drift\}\}\): a latent lot process over batches and a linear trend shift the batch mean of every variable, so batches become dependent and every pair of variables shares these common causes\.

*Imperfections*, which change how the relations are observed or how they act:

- •*Sensor error*\(serrs\_\{\\mathrm\{err\}\}\): each reading mixes the true value with independent, slowly drifting error, weighted byserrs\_\{\\mathrm\{err\}\}\. A method conditions on the reading of a cause, not on its true value, so the path from a sibling through the cause toYYis no longer fully blocked\.
- •*Product grades*\(cc\): a latent grade per batch scales the effect of each direct cause in proportion1\+c1\+cto1−c1\-cbetween the two grades, so atc=1\{c=1\}each cause acts in one grade only\.
- •*Variable delays*\(jj\): the dead time of each direct cause grows in each batch by a random amount up tojjtimes its time scale, as mixing takes longer in a fuller vessel\.
- •*Cumulative effects*\(kavgk\_\{\\mathrm\{avg\}\}\):kavgk\_\{\\mathrm\{avg\}\}weights each cause’s recent average against its current response, as when the product depends on the heat received over the last hour\.
- •*Operator reactions*\(sops\_\{\\mathrm\{op\}\}\): after a deviation ofYY, the reference value of one direct cause is moved to counter it, one decorrelation time ofYYlater, so the graph unrolled over time stays acyclic\.

## 5Experimental setup

#### Plants and configurations\.

At the reference, each plant hasp=20\{p=20\}recorded variables, between one and four direct causes \(\|D\|\|D\|\) and between 8 and 11 batch conditions \(mm\), and generatesB=200\{B=200\}batches ofT=512\{T=512\}steps; the size configurations raiseppto 40 or lowerBBto 25, 50 or 100 \(Table[2](https://arxiv.org/html/2611.17556#S5.T2)\)\. The candidates are thep−1p\-1recorded variables other than the actuator, orp−2p\-2when the controller is on \(Section[4\.2](https://arxiv.org/html/2611.17556#S4.SS2.SSS0.Px2)\)\. The reference configuration has no additional component and moderate shares; the other 118 configurations \(Table[2](https://arxiv.org/html/2611.17556#S5.T2)\) move one share of Section[4\.3](https://arxiv.org/html/2611.17556#S4.SS3)or one component of Section[4\.4](https://arxiv.org/html/2611.17556#S4.SS4)over a range, with all else at the reference, or combine several\. The*combined difficulties*, used throughout Section[6](https://arxiv.org/html/2611.17556#S6), move five shares at once:ssig=0\.3\{s\_\{\\mathrm\{sig\}\}=0\.3\},sconf=sup=0\.6\{s\_\{\\mathrm\{conf\}\}=s\_\{\\mathrm\{up\}\}=0\.6\},sdesc=0\.5\{s\_\{\\mathrm\{desc\}\}=0\.5\}andsint=1\{s\_\{\\mathrm\{int\}\}=1\}, that is, a faint signal, strong confounding and upstream links, strong downstream measurements and causes that act only jointly\. All ranges and the reference values are listed in Appendix[A](https://arxiv.org/html/2611.17556#A1)\. Each configuration is evaluated on 10 plants\. We audit the roles of one plant per configuration by intervention: replayed with its noise held fixed,do⁡\(X:=X\+ξ\)\{\\mathrm\{do\}\(X:=X\+\\xi\)\}on a direct cause or an ancestor changesYY, on a sibling or a descendant leaves it exactly unchanged\.

Table 2:The released benchmark: 119 configurations of the generator\. Candidates per plant: mean±\\pms\.d\. over all plants\. Shares: each setting, with all else at the reference, and its value measured on the generated plants \(median over 10\)\.
#### Baselines\.

We compare eight methods from four families: the association measures lagged correlation and Granger causality\([Granger, 1969](https://arxiv.org/html/2611.17556#bib.bib31)\), which do not adjust for the other candidates; the conditional\-independence tests PC\([Spirtes et al\., 2000](https://arxiv.org/html/2611.17556#bib.bib35)\)and PCMCI\+\([Runge et al\., 2019b](https://arxiv.org/html/2611.17556#bib.bib32);[Runge, 2020](https://arxiv.org/html/2611.17556#bib.bib33)\); the score\-based and functional models DYNOTEARS\([Pamfil et al\., 2020](https://arxiv.org/html/2611.17556#bib.bib34)\), VARLiNGAM\([Hyvärinen et al\., 2010](https://arxiv.org/html/2611.17556#bib.bib38)\)and SpaceTime\([Mameche et al\., 2025](https://arxiv.org/html/2611.17556#bib.bib36)\); and DoubleML\([Chernozhukov et al\., 2018](https://arxiv.org/html/2611.17556#bib.bib37)\)with LightGBM nuisance models\([Ke et al\., 2017](https://arxiv.org/html/2611.17556#bib.bib43)\), which estimates the effect of each candidate’s past onYY\. Hyperparameters are the same in every configuration and listed in Appendix[E](https://arxiv.org/html/2611.17556#A5)\. As sortability checks, we also rank the candidates by their variance and by how well the other variables predict them\([Reisach et al\., 2021](https://arxiv.org/html/2611.17556#bib.bib15);[Reisach et al\., 2023](https://arxiv.org/html/2611.17556#bib.bib16);[Lohse and Wahl, 2025](https://arxiv.org/html/2611.17556#bib.bib17)\)\. All methods receive the data at the same resolution and use the same look\-back, since results depend on the lag and the sampling rate\([Shojaie and Fox, 2022](https://arxiv.org/html/2611.17556#bib.bib30)\)\. The methods receive the mean of everyffsteps, withffset from the smallest target decorrelation time across plants in each configuration \(f=32\{f=32\}at the reference; Appendix[B](https://arxiv.org/html/2611.17556#A2)\), and look back twice the largest\. This matches what a process historian normally stores: averages over fixed intervals\. The data are released at full resolution\.

#### Metrics\.

Each method returns a selected list of candidates, possibly empty, and a ranking of all candidates\. The selected list is what a method would hand to a process engineer, and we score it by its*first pick*, whether its first entry is a direct cause \(an empty list counts as wrong\); its*precision*, over the plants where it is not empty; its*lift*, the precision divided by that of a random list \(\|D\|/\|𝒞\|\|D\|/\|\\mathcal\{C\}\|\); its*length*; and how often it is*empty*\. The ranking does not depend on where a method cuts its list; we score it with*R\-precision*, the share of direct causes among the first\|D\|\|D\|candidates, mean average precision, mean reciprocal rank and nDCG@3\([Manning et al\., 2008](https://arxiv.org/html/2611.17556#bib.bib3)\); ties are ordered at random\.

## 6Experiments

Figure[2](https://arxiv.org/html/2611.17556#S6.F2)shows the selected lists on the reference configuration and under the combined difficulties, Table[3](https://arxiv.org/html/2611.17556#S6.T3)the metrics of each method over all 119 configurations, and Appendix[C](https://arxiv.org/html/2611.17556#A3)every configuration and metric\. Each method selects candidates at a fixed threshold: PC and PCMCI\+ at a significance level of 0\.05; correlation, Granger causality and DoubleML at a Benjamini–Hochberg adjustedpp\-value below 0\.05\([Benjamini and Hochberg, 1995](https://arxiv.org/html/2611.17556#bib.bib47)\); the others by the rules in Appendix[E](https://arxiv.org/html/2611.17556#A5)\.

![Refer to caption](https://arxiv.org/html/2611.17556v1/fig_money.png)Figure 2:The selected lists of the eight methods on the 10 plants of the reference configuration \(top\) and of the combined difficulties \(bottom\)\. Green beyond the black tick is better than random; grey lift: fewer than three non\-empty lists\.#### Reference and combined difficulties\.

In the reference configuration \(Figure[2](https://arxiv.org/html/2611.17556#S6.F2), top\), every method except PC ranks a direct cause first on most plants; the methods differ in their lists\. Correlation and Granger causality select 18\.2 and 17\.1 of the 19 candidates, with lifts near one, while PCMCI\+, DoubleML and SpaceTime, which adjust for the other candidates, select one to three with lifts of 5\.7, 4\.7 and 6\.3\.

Performance deteriorates when the difficulties are combined \(Figure[2](https://arxiv.org/html/2611.17556#S6.F2), bottom; Table[9](https://arxiv.org/html/2611.17556#A3.T9)\)\. Two difficulty measures computed from the generated data confirm that the combined configurations are harder and place them alongside three public industrial process datasets \(Appendix[D](https://arxiv.org/html/2611.17556#A4)\)\. Correlation, VARLiNGAM and PC name a descendant ofYYfirst on 10, 9 and 8 of 10 plants; PCMCI\+, DoubleML and DYNOTEARS select nothing on 4, 7 and 10; SpaceTime names a direct cause first on 6 plants with a list of 1\.5 candidates, Granger on 4 with one of 3\.9\.

#### One difficulty at a time\.

Because the generator sets each component independently, through the shares of Section[4\.3](https://arxiv.org/html/2611.17556#S4.SS3)and the additional components of Section[4\.4](https://arxiv.org/html/2611.17556#S4.SS4), we can simulate configurations that depart from the reference in one component only \(Table[9](https://arxiv.org/html/2611.17556#A3.T9); Figure[4](https://arxiv.org/html/2611.17556#A3.F4)shows two\)\. A faint signal \(ssig=0\.1\{s\_\{\\mathrm\{sig\}\}=0\.1\}\) makes the downstream measurements, which respond to all ofYYincluding the part no recorded variable explains, more strongly associated withYYthan its causes: correlation names a descendant first on 10 plants, while Granger and PCMCI\+, which use the delay, still name a direct cause first on most; SpaceTime, which accepts a parent only if it shortens the description length ofYY, selects nothing on 6 plants\. Joint effects \(sint=1\{s\_\{\\mathrm\{int\}\}=1\}\) give each cause an average linear effect near zero: DoubleML, whose partially linear model estimates that quantity, abstains on 6 plants, while SpaceTime, whose Gaussian\-process mechanisms represent joint effects, names a direct cause first on 9/10\. Stronger confounding pulls the siblings ahead of the weaker causes: atsconf=0\.8\{s\_\{\\mathrm\{conf\}\}=0\.8\}, correlation, VARLiNGAM, DYNOTEARS and SpaceTime name a direct cause first on 7, 5, 3 and 6 of 10 plants, while PCMCI\+ and DoubleML, which condition on the direct causes, keep 10 and 10\. Upstream links, downstream measurements and the controller leave PCMCI\+, DoubleML and SpaceTime close to their reference results \(Table[9](https://arxiv.org/html/2611.17556#A3.T9), rowssups\_\{\\mathrm\{up\}\},sdescs\_\{\\mathrm\{desc\}\}andκ\\kappa\)\.

The batch\-level components and the imperfections generally leave the first picks of PCMCI\+, DoubleML and SpaceTime intact\. Sensor error \(error share 0\.5\) lowers the precision of the methods that condition on the candidates, since a noisy reading does not fully block a path \(PCMCI\+: 0\.79→\\to0\.47\)\. Unrecorded product grades weaken the pooled association of each cause withYYbut not that of its descendants: correlation names a descendant first on 4 plants\. Lots and drift move every variable together from batch to batch, which misleads correlation \(10/10→\\to7/10 at lot share 0\.9\) but not the methods that condition on the direct causes\. Recipe phases align non\-causes withYYin time and add wrong candidates to the lists of PCMCI\+ \(precision 0\.79→\\to0\.66\)\. Variable delays and operator reactions change little at the evaluation resolution\. The number of batches matters: with 25 instead of 200, the first picks of PCMCI\+ and DoubleML worsen, while those of Granger and SpaceTime change little\.

Table 3:Each method over all configurations \(top\) and over those with the combined difficulties \(bottom\)\. Subscripts: half\-width of the 95%tt\-interval over the 10 plant structures, each averaged over its configurations \(precision and lift: over the structures with a non\-empty list\)\.
#### Over all configurations\.

The lift of the selected list separates the methods most clearly \(Table[3](https://arxiv.org/html/2611.17556#S6.T3)\): the lists of correlation and Granger causality are barely more precise than a random one \(lift 1\.1 and 1\.7\), while PCMCI\+, DoubleML and SpaceTime return one to three candidates with lifts of 5\.3, 5\.5 and 5\.9\. DYNOTEARS has the highest precision and lift at its paper’s thresholds, 1\.00 and 7\.3, but on an almost empty list: it selects nothing on 72% of the plants\. At thresholds ten times lower it selects on most plants and is about as good as PCMCI\+ \(Appendix[C](https://arxiv.org/html/2611.17556#A3)\)\. Under the combined difficulties, correlation’s R\-precision falls to 0\.14, about that of a random ranking\. Granger, SpaceTime and PCMCI\+ name a direct cause first on 48%, 41% and 41% of the plants and select nothing on 31%, 13% and 31%; DoubleML selects nothing on 67% and names a direct cause first on only 19%\. Sorting the candidates by variance or explained variance, a known artefact of simulated data\([Reisach et al\., 2021](https://arxiv.org/html/2611.17556#bib.bib15);[Reisach et al\., 2023](https://arxiv.org/html/2611.17556#bib.bib16);[Lohse and Wahl, 2025](https://arxiv.org/html/2611.17556#bib.bib17)\), picks a direct cause first on 9% and 6% of the plants, against 13% at random\.

#### Errors by method\.

Because the role of every candidate is known, each wrong selection can be broken down by role \(Table[8](https://arxiv.org/html/2611.17556#A3.T8), Appendix[C](https://arxiv.org/html/2611.17556#A3)\)\. Correlation and Granger causality select every role in about its random proportion but differ in their first pick: correlation ignores time order and ranks a descendant ofYYfirst on 33% of the plants, Granger on 1%, and most of Granger’s wrong first picks are ancestors one step upstream of a cause\. PC, without time order, selects descendants \(55% of its wrong selections, against 13% at random\)\. PCMCI\+, DoubleML and SpaceTime make fewer than one wrong selection per plant; their characteristic failure is to select nothing when difficulties combine, on 10%, 27% and 16% of the plants\.

## 7Conclusion

We presented a synthetic benchmark for target\-centric causal discovery in industrial processes\. Its generator is a two\-level structural causal model inspired by batch production\. Each component is switched on and set independently, and the ground truth gives candidate roles and separation or time\-order labels\. Eight methods, evaluated on 119 configurations, degrade as the difficulties combine\. Configurations can change one component at a time, so that a change in a method’s result can be attributed to one mechanism\.

#### Using the benchmark\.

First, researchers can test a new method on each difficulty separately and on their combinations: each configuration changes named components of the reference, and every wrong selection is labelled with the candidate’s role\. Second, engineers can generate plants that resemble their own in size and difficulties and test a method on them before applying it to the plant’s data\. Third, the generator can supply training data\. Foundation models for tabular prediction and causal effect estimation are pretrained on data drawn from structural causal models\([Hollmann et al\., 2025](https://arxiv.org/html/2611.17556#bib.bib41);[Qu et al\., 2026](https://arxiv.org/html/2611.17556#bib.bib39);[Robertson et al\., 2025](https://arxiv.org/html/2611.17556#bib.bib40)\); a generator of process data could serve the same purpose for industrial processes\.

#### Limitations\.

The benchmark is synthetic: the generator reproduces the causal structure of a batch process \(Section[4\.1](https://arxiv.org/html/2611.17556#S4.SS1)\), not its physics\. The profiles come from one family of motifs and the exogenous noise is Gaussian\. Each configuration is evaluated on 10 plants only, and the methods run at one resolution and look\-back\. The benchmark gives no finite\-sample recovery bound, so a failure of every method cannot be distinguished from an unrecoverable signal\.

#### Use of generative AI\.

Generative AI tools were used for language editing; the authors take full responsibility for the content\.

## References

- Aliferiset al\.\(2010\)C\. F\. Aliferis, A\. Statnikov, I\. Tsamardinos, S\. Mani, and X\. D\. KoutsoukosLocal causal and Markov blanket induction for causal discovery and feature selection for classification part I: algorithms and empirical evaluation\.Journal of Machine Learning Research11\(7\),pp\. 171–234\.External Links:[Link](https://jmlr.org/papers/v11/aliferis10a.html)Cited by:[§1](https://arxiv.org/html/2611.17556#S1.p1.1)\.
- Assaadet al\.\(2022\)C\. K\. Assaad, E\. Devijver, and E\. GaussierSurvey and evaluation of causal discovery methods for time series\.Journal of Artificial Intelligence Research73,pp\. 767–819\.External Links:[Document](https://dx.doi.org/10.1613/jair.1.13428)Cited by:[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Benjamini and Hochberg \(1995\)Y\. Benjamini and Y\. HochbergControlling the false discovery rate: a practical and powerful approach to multiple testing\.Journal of the Royal Statistical Society: Series B \(Methodological\)57\(1\),pp\. 289–300\.External Links:[Document](https://dx.doi.org/10.1111/j.2517-6161.1995.tb02031.x)Cited by:[§6](https://arxiv.org/html/2611.17556#S6.p1.1)\.
- Chenget al\.\(2024\)Y\. Cheng, Z\. Wang, T\. Xiao, Q\. Zhong, J\. Suo, and K\. HeCausalTime: realistically generated time\-series for benchmarking of causal discovery\.InInternational Conference on Learning Representations,B\. Kim, Y\. Yue, S\. Chaudhuri, K\. Fragkiadaki, M\. Khan, and Y\. Sun \(Eds\.\),Vol\.2024,pp\. 2923–2945\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2024/file/0c79d6ed1788653643a1ac67b6ea32a7-Paper-Conference.pdf)Cited by:[§1](https://arxiv.org/html/2611.17556#S1.p2.1),[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Chernozhukovet al\.\(2018\)V\. Chernozhukov, D\. Chetverikov, M\. Demirer, E\. Duflo, C\. Hansen, W\. Newey, and J\. RobinsDouble/debiased machine learning for treatment and structural parameters\.The Econometrics Journal21\(1\),pp\. C1–C68\.External Links:[Document](https://dx.doi.org/10.1111/ectj.12097)Cited by:[§5](https://arxiv.org/html/2611.17556#S5.SS0.SSS0.Px2.p1.1)\.
- Downs and Vogel \(1993\)J\. J\. Downs and E\. F\. VogelA plant\-wide industrial process control problem\.Computers & Chemical Engineering17\(3\),pp\. 245–255\.External Links:[Document](https://dx.doi.org/10.1016/0098-1354%2893%2980018-I)Cited by:[Appendix D](https://arxiv.org/html/2611.17556#A4.p4.1),[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Forré and Mooij \(2017\)P\. Forré and J\. M\. MooijMarkov properties for graphical models with cycles and latent variables\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.1710.08775)Cited by:[Appendix B](https://arxiv.org/html/2611.17556#A2.SS0.SSS0.Px5.p1.1)\.
- Forré and Mooij \(2018\)P\. Forré and J\. M\. MooijConstraint\-based causal discovery for non\-linear structural causal models with cycles and latent confounders\.InProceedings of the 34th Conference on Uncertainty in Artificial Intelligence,A\. Globerson and R\. Silva \(Eds\.\),Proceedings of Machine Learning Research, Vol\.R16,pp\. 268–277\.External Links:[Link](https://proceedings.mlr.press/r16/forre18a.html)Cited by:[Appendix B](https://arxiv.org/html/2611.17556#A2.SS0.SSS0.Px5.p1.1)\.
- Gamellaet al\.\(2025\)J\. L\. Gamella, J\. Peters, and P\. BühlmannCausal chambers as a real\-world physical testbed for AI methodology\.Nature Machine Intelligence7\(1\),pp\. 107–118\.External Links:[Document](https://dx.doi.org/10.1038/s42256-024-00964-x)Cited by:[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Göbleret al\.\(2024\)K\. Göbler, T\. Windisch, M\. Drton, T\. Pychynski, M\. Roth, and S\. SonntagcausalAssembly: generating realistic production data for benchmarking causal discovery\.InProceedings of the Third Conference on Causal Learning and Reasoning,F\. Locatello and V\. Didelez \(Eds\.\),Proceedings of Machine Learning Research, Vol\.236,pp\. 609–642\.External Links:[Link](https://proceedings.mlr.press/v236/gobler24a.html)Cited by:[§1](https://arxiv.org/html/2611.17556#S1.p2.1),[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Goldricket al\.\(2019\)S\. Goldrick, C\. A\. Duran\-Villalobos, K\. Jankauskas, D\. Lovett, S\. S\. Farid, and B\. LennoxModern day monitoring and control challenges outlined on an industrial\-scale benchmark fermentation process\.Computers & Chemical Engineering130,pp\. 106471\.External Links:[Document](https://dx.doi.org/10.1016/j.compchemeng.2019.05.037)Cited by:[Appendix D](https://arxiv.org/html/2611.17556#A4.p4.1),[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Goldricket al\.\(2015\)S\. Goldrick, A\. Ştefan, D\. Lovett, G\. Montague, and B\. LennoxThe development of an industrial\-scale fed\-batch fermentation simulation\.Journal of Biotechnology193,pp\. 70–82\.External Links:[Document](https://dx.doi.org/10.1016/j.jbiotec.2014.10.029)Cited by:[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Gonget al\.\(2024\)C\. Gong, C\. Zhang, D\. Yao, J\. Bi, W\. Li, and Y\. XuCausal discovery from temporal data: an overview and new perspectives\.ACM Computing Surveys57\(4\),pp\. 100:1–100:38\.External Links:[Document](https://dx.doi.org/10.1145/3705297)Cited by:[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Gonget al\.\(2017\)M\. Gong, K\. Zhang, B\. Schölkopf, C\. Glymour, and D\. TaoCausal discovery from temporally aggregated time series\.InProceedings of the 33rd Conference on Uncertainty in Artificial Intelligence,G\. Elidan and K\. Kersting \(Eds\.\),Proceedings of Machine Learning Research, Vol\.R15,pp\. 161–170\.External Links:[Link](https://proceedings.mlr.press/r15/gong17a.html)Cited by:[Appendix B](https://arxiv.org/html/2611.17556#A2.SS0.SSS0.Px6.p1.1)\.
- Granger \(1969\)C\. W\. J\. GrangerInvestigating causal relations by econometric models and cross\-spectral methods\.Econometrica37\(3\),pp\. 424–438\.External Links:[Document](https://dx.doi.org/10.2307/1912791)Cited by:[§5](https://arxiv.org/html/2611.17556#S5.SS0.SSS0.Px2.p1.1)\.
- Hardtet al\.\(2024\)M\. Hardt, W\. R\. Orchard, P\. Blöbaum, E\. Kirschbaum, and S\. KasiviswanathanThe PetShop dataset — finding causes of performance issues across microservices\.InProceedings of the Third Conference on Causal Learning and Reasoning,F\. Locatello and V\. Didelez \(Eds\.\),Proceedings of Machine Learning Research, Vol\.236,pp\. 957–978\.External Links:[Link](https://proceedings.mlr.press/v236/hardt24a.html)Cited by:[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Herdeanuet al\.\(2025\)B\. Herdeanu, J\. Nathaniel, C\. Roesch, J\. Buch, G\. Ramien, J\. Haux, and P\. GentineCausalDynamics: a large\-scale benchmark for structural discovery of dynamical causal models\.InAdvances in Neural Information Processing Systems,D\. Belgrave, C\. Zhang, H\. Lin, R\. Pascanu, P\. Koniusz, M\. Ghassemi, and N\. Chen \(Eds\.\),Vol\.38,pp\. 45204–45248\.External Links:[Document](https://dx.doi.org/10.52202/085713-1349),[Link](https://proceedings.neurips.cc/paper_files/paper/2025/file/39a3f7dc00e4723a5ca7808b04fc4d4f-Paper-Datasets_and_Benchmarks_Track.pdf)Cited by:[§1](https://arxiv.org/html/2611.17556#S1.p2.1),[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Hollmannet al\.\(2025\)N\. Hollmann, S\. Müller, L\. Purucker, A\. Krishnakumar, M\. Körfer, S\. B\. Hoo, R\. T\. Schirrmeister, and F\. HutterAccurate predictions on small data with a tabular foundation model\.Nature637\(8045\),pp\. 319–326\.External Links:[Document](https://dx.doi.org/10.1038/s41586-024-08328-6)Cited by:[§7](https://arxiv.org/html/2611.17556#S7.SS0.SSS0.Px1.p1.1)\.
- Hyvärinenet al\.\(2010\)A\. Hyvärinen, K\. Zhang, S\. Shimizu, and P\. O\. HoyerEstimation of a structural vector autoregression model using non\-Gaussianity\.Journal of Machine Learning Research11\(56\),pp\. 1709–1731\.External Links:[Link](https://jmlr.org/papers/v11/hyvarinen10a.html)Cited by:[§5](https://arxiv.org/html/2611.17556#S5.SS0.SSS0.Px2.p1.1)\.
- Keet al\.\(2017\)G\. Ke, Q\. Meng, T\. Finley, T\. Wang, W\. Chen, W\. Ma, Q\. Ye, and T\. LiuLightGBM: a highly efficient gradient boosting decision tree\.InAdvances in Neural Information Processing Systems,I\. Guyon, U\. V\. Luxburg, S\. Bengio, H\. Wallach, R\. Fergus, S\. Vishwanathan, and R\. Garnett \(Eds\.\),Vol\.30,pp\. 3146–3154\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2017/file/6449f44a102fde848669bdd9eb6b76fa-Paper.pdf)Cited by:[Appendix E](https://arxiv.org/html/2611.17556#A5.p3.1),[§5](https://arxiv.org/html/2611.17556#S5.SS0.SSS0.Px2.p1.1)\.
- Krapivsky and Redner \(2001\)P\. L\. Krapivsky and S\. RednerOrganization of growing random networks\.Physical Review E63\(6\),pp\. 066123\.External Links:[Document](https://dx.doi.org/10.1103/PhysRevE.63.066123)Cited by:[Table 4](https://arxiv.org/html/2611.17556#A1.T4.4.3.2.1.1)\.
- Lawrenceet al\.\(2020\)A\. R\. Lawrence, M\. Kaiser, R\. Sampaio, and M\. SiposData generating process to evaluate causal discovery techniques for time series data\.InCausal Discovery & Causality\-Inspired Machine Learning Workshop at NeurIPS,External Links:[Document](https://dx.doi.org/10.48550/arXiv.2104.08043)Cited by:[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Linget al\.\(2025\)Z\. Ling, H\. Peng, Y\. Zhang, D\. Cheng, X\. Wu, P\. Zhou, and K\. YuHybrid local causal discovery\.InProceedings of the Thirty\-Fourth International Joint Conference on Artificial Intelligence, IJCAI\-25,J\. Kwok \(Ed\.\),pp\. 5725–5733\.External Links:[Document](https://dx.doi.org/10.24963/ijcai.2025/637)Cited by:[§1](https://arxiv.org/html/2611.17556#S1.p1.1)\.
- Lohse and Wahl \(2025\)C\. Lohse and J\. WahlSortability of time series data\.Transactions on Machine Learning Research\.External Links:[Link](https://openreview.net/forum?id=OGvmCpcHdV)Cited by:[§5](https://arxiv.org/html/2611.17556#S5.SS0.SSS0.Px2.p1.1),[§6](https://arxiv.org/html/2611.17556#S6.SS0.SSS0.Px3.p1.1)\.
- Mamecheet al\.\(2025\)S\. Mameche, L\. Cornanguer, U\. Ninad, and J\. VreekenSpaceTime: causal discovery from non\-stationary time series\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.39,pp\. 19405–19413\.External Links:[Document](https://dx.doi.org/10.1609/aaai.v39i18.34136)Cited by:[§5](https://arxiv.org/html/2611.17556#S5.SS0.SSS0.Px2.p1.1)\.
- Manninget al\.\(2008\)C\. D\. Manning, P\. Raghavan, and H\. SchützeIntroduction to information retrieval\.Cambridge University Press\.External Links:[Document](https://dx.doi.org/10.1017/CBO9780511809071)Cited by:[§5](https://arxiv.org/html/2611.17556#S5.SS0.SSS0.Px3.p1.1)\.
- Mehlinget al\.\(2026\)C\. W\. Mehling, S\. Pieper, T\. Lüke, J\. Döbelt, and S\. IhlenfeldtEnabling joint benchmarking of automated root cause analysis and causal discovery in manufacturing using the causRCA dataset\.Procedia CIRP139,pp\. 114–120\.External Links:[Document](https://dx.doi.org/10.1016/j.procir.2025.09.010)Cited by:[§1](https://arxiv.org/html/2611.17556#S1.p2.1),[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Menegozzoet al\.\(2022\)G\. Menegozzo, D\. Dall’Alba, and P\. FioriniCIPCaD\-Bench: continuous industrial process datasets for benchmarking causal discovery methods\.In2022 IEEE 18th International Conference on Automation Science and Engineering \(CASE\),pp\. 2124–2131\.External Links:[Document](https://dx.doi.org/10.1109/CASE49997.2022.9926420)Cited by:[Appendix D](https://arxiv.org/html/2611.17556#A4.p4.1),[§1](https://arxiv.org/html/2611.17556#S1.p2.1),[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Mogensenet al\.\(2024\)S\. W\. Mogensen, K\. Rathsman, and P\. NilssonCausal discovery in a complex industrial system: a time series benchmark\.InProceedings of the Third Conference on Causal Learning and Reasoning,F\. Locatello and V\. Didelez \(Eds\.\),Proceedings of Machine Learning Research, Vol\.236,pp\. 1218–1236\.External Links:[Link](https://proceedings.mlr.press/v236/mogensen24a.html)Cited by:[§1](https://arxiv.org/html/2611.17556#S1.p2.1),[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Mooij and Claassen \(2020\)J\. M\. Mooij and T\. ClaassenConstraint\-based causal discovery using partial ancestral graphs in the presence of cycles\.InProceedings of the 36th Conference on Uncertainty in Artificial Intelligence \(UAI\),J\. Peters and D\. Sontag \(Eds\.\),Proceedings of Machine Learning Research, Vol\.124,pp\. 1159–1168\.External Links:[Link](https://proceedings.mlr.press/v124/m-mooij20a.html)Cited by:[Appendix B](https://arxiv.org/html/2611.17556#A2.SS0.SSS0.Px5.p1.1)\.
- Nomikos and MacGregor \(1994\)P\. Nomikos and J\. F\. MacGregorMonitoring batch processes using multiway principal component analysis\.AIChE Journal40\(8\),pp\. 1361–1375\.External Links:[Document](https://dx.doi.org/10.1002/aic.690400809)Cited by:[§4\.1](https://arxiv.org/html/2611.17556#S4.SS1.p2.1)\.
- Pamfilet al\.\(2020\)R\. Pamfil, N\. Sriwattanaworachai, S\. Desai, P\. Pilgerstorfer, K\. Georgatzis, P\. Beaumont, and B\. AragamDYNOTEARS: structure learning from time\-series data\.InProceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics,S\. Chiappa and R\. Calandra \(Eds\.\),Proceedings of Machine Learning Research, Vol\.108,pp\. 1595–1605\.External Links:[Link](https://proceedings.mlr.press/v108/pamfil20a.html)Cited by:[Appendix C](https://arxiv.org/html/2611.17556#A3.SS0.SSS0.Px3.p1.1),[§5](https://arxiv.org/html/2611.17556#S5.SS0.SSS0.Px2.p1.1)\.
- Panayiotouet al\.\(2026\)P\. Panayiotou, A\. Poinsot, A\. Leite, N\. Chesneau, M\. Schoenauer, and Ö\. ŞimşekCausalProfiler: generating synthetic benchmarks for rigorous and transparent evaluation of causal machine learning\.InProceedings of the 43rd International Conference on Machine Learning,T\. Zhang, M\. Dudik, M\. Jaggi, A\. Agarwal, S\. Li, D\. Schuurmans, J\. Zhu, F\. Berkenkamp, H\. Dong, and A\. Bietti \(Eds\.\),Proceedings of Machine Learning Research, Vol\.306,pp\. 96055–96110\.External Links:[Link](https://proceedings.mlr.press/v306/panayiotou26a.html)Cited by:[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Quet al\.\(2026\)J\. Qu, D\. Holzmüller, G\. Varoquaux, and M\. Le MorvanTabICLv2: a better, faster, scalable, and open tabular foundation model\.InProceedings of the 43rd International Conference on Machine Learning,T\. Zhang, M\. Dudik, M\. Jaggi, A\. Agarwal, S\. Li, D\. Schuurmans, J\. Zhu, F\. Berkenkamp, H\. Dong, and A\. Bietti \(Eds\.\),Proceedings of Machine Learning Research, Vol\.306,pp\. 102477–102555\.External Links:[Link](https://proceedings.mlr.press/v306/qu26a.html)Cited by:[§7](https://arxiv.org/html/2611.17556#S7.SS0.SSS0.Px1.p1.1)\.
- Reinartzet al\.\(2021\)C\. Reinartz, M\. Kulahci, and O\. RavnAn extended Tennessee Eastman simulation dataset for fault\-detection and decision support systems\.Computers & Chemical Engineering149,pp\. 107281\.External Links:[Document](https://dx.doi.org/10.1016/j.compchemeng.2021.107281)Cited by:[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Reisachet al\.\(2021\)A\. G\. Reisach, C\. Seiler, and S\. WeichwaldBeware of the simulated DAG\! Causal discovery benchmarks may be easy to game\.InAdvances in Neural Information Processing Systems,M\. Ranzato, A\. Beygelzimer, Y\. Dauphin, P\.S\. Liang, and J\. W\. Vaughan \(Eds\.\),Vol\.34,pp\. 27772–27784\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2021/file/e987eff4a7c7b7e580d659feb6f60c1a-Paper.pdf)Cited by:[§5](https://arxiv.org/html/2611.17556#S5.SS0.SSS0.Px2.p1.1),[§6](https://arxiv.org/html/2611.17556#S6.SS0.SSS0.Px3.p1.1)\.
- Reisachet al\.\(2023\)A\. G\. Reisach, M\. Tami, C\. Seiler, A\. Chambaz, and S\. WeichwaldA scale\-invariant sorting criterion to find a causal order in additive noise models\.InAdvances in Neural Information Processing Systems,Vol\.36,pp\. 785–807\.External Links:[Document](https://dx.doi.org/10.52202/075280-0036)Cited by:[§5](https://arxiv.org/html/2611.17556#S5.SS0.SSS0.Px2.p1.1),[§6](https://arxiv.org/html/2611.17556#S6.SS0.SSS0.Px3.p1.1)\.
- Robertsonet al\.\(2025\)J\. Robertson, A\. Reuter, S\. Guo, N\. Hollmann, F\. Hutter, and B\. SchölkopfDo\-PFN: in\-context learning for causal effect estimation\.InAdvances in Neural Information Processing Systems,D\. Belgrave, C\. Zhang, H\. Lin, R\. Pascanu, P\. Koniusz, M\. Ghassemi, and N\. Chen \(Eds\.\),Vol\.38,pp\. 193605–193642\.External Links:[Document](https://dx.doi.org/10.52202/085713-5814)Cited by:[§7](https://arxiv.org/html/2611.17556#S7.SS0.SSS0.Px1.p1.1)\.
- Rungeet al\.\(2019a\)J\. Runge, S\. Bathiany, E\. Bollt, G\. Camps\-Valls, D\. Coumou, E\. Deyle, C\. Glymour, M\. Kretschmer, M\. D\. Mahecha, J\. Muñoz\-Marí, E\. H\. van Nes, J\. Peters, R\. Quax, M\. Reichstein, M\. Scheffer, B\. Schölkopf, P\. Spirtes, G\. Sugihara, J\. Sun, K\. Zhang, and J\. ZscheischlerInferring causation from time series in Earth system sciences\.Nature Communications10\(1\),pp\. 2553\.External Links:[Document](https://dx.doi.org/10.1038/s41467-019-10105-3)Cited by:[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Rungeet al\.\(2019b\)J\. Runge, P\. Nowack, M\. Kretschmer, S\. Flaxman, and D\. SejdinovicDetecting and quantifying causal associations in large nonlinear time series datasets\.Science Advances5\(11\),pp\. eaau4996\.External Links:[Document](https://dx.doi.org/10.1126/sciadv.aau4996)Cited by:[§5](https://arxiv.org/html/2611.17556#S5.SS0.SSS0.Px2.p1.1)\.
- Runge \(2020\)J\. RungeDiscovering contemporaneous and lagged causal relations in autocorrelated nonlinear time series datasets\.InProceedings of the 36th Conference on Uncertainty in Artificial Intelligence \(UAI\),J\. Peters and D\. Sontag \(Eds\.\),Proceedings of Machine Learning Research, Vol\.124,pp\. 1388–1397\.External Links:[Link](https://proceedings.mlr.press/v124/runge20a.html)Cited by:[§5](https://arxiv.org/html/2611.17556#S5.SS0.SSS0.Px2.p1.1)\.
- Scheines and Ramsey \(2016\)R\. Scheines and J\. RamseyMeasurement error and causal discovery\.InProceedings of the UAI 2016 Workshop on Causation: Foundation to Application,F\. Eberhardt, E\. Bareinboim, M\. Maathuis, J\. Mooij, and R\. Silva \(Eds\.\),CEUR Workshop Proceedings, Vol\.1792,pp\. 1–7\.External Links:[Link](https://ceur-ws.org/Vol-1792/paper1.pdf)Cited by:[Appendix B](https://arxiv.org/html/2611.17556#A2.SS0.SSS0.Px5.p1.1)\.
- Shojaie and Fox \(2022\)A\. Shojaie and E\. B\. FoxGranger causality: a review and recent advances\.Annual Review of Statistics and Its Application9\(1\),pp\. 289–319\.External Links:[Document](https://dx.doi.org/10.1146/annurev-statistics-040120-010930)Cited by:[§5](https://arxiv.org/html/2611.17556#S5.SS0.SSS0.Px2.p1.1)\.
- Spirteset al\.\(2000\)P\. Spirtes, C\. Glymour, and R\. ScheinesCausation, prediction, and search\.2 edition,MIT Press\.External Links:[Document](https://dx.doi.org/10.7551/mitpress/1754.001.0001)Cited by:[Appendix B](https://arxiv.org/html/2611.17556#A2.SS0.SSS0.Px5.p1.1),[§3](https://arxiv.org/html/2611.17556#S3.p2.1),[§5](https://arxiv.org/html/2611.17556#S5.SS0.SSS0.Px2.p1.1)\.
- Steinet al\.\(2025\)G\. Stein, M\. Shadaydeh, J\. Blunk, N\. Penzel, and J\. DenzlerCausalRivers – scaling up benchmarking of causal discovery for real\-world time\-series\.InInternational Conference on Learning Representations,Y\. Yue, A\. Garg, N\. Peng, F\. Sha, and R\. Yu \(Eds\.\),Vol\.2025,pp\. 64516–64531\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2025/file/a205fda871b0f6c1e18a7ad7325eb6cf-Paper-Conference.pdf)Cited by:[§1](https://arxiv.org/html/2611.17556#S1.p2.1),[§2](https://arxiv.org/html/2611.17556#S2.p1.1)\.
- Vuković and Thalmann \(2022\)M\. Vuković and S\. ThalmannCausal discovery in manufacturing: a structured literature review\.Journal of Manufacturing and Materials Processing6\(1\),pp\. 10\.External Links:[Document](https://dx.doi.org/10.3390/jmmp6010010)Cited by:[§1](https://arxiv.org/html/2611.17556#S1.p1.1)\.
- Xieet al\.\(2026\)S\. Xie, V\. Feofanov, J\. Zhang, T\. Palpanas, and I\. RedkoCauKer: classification time series foundation models can be pretrained on synthetic data\.InInternational Conference on Learning Representations,C\. Vondrick, B\. Hariharan, C\. Raffel, L\. Pinto, D\. Yang, and A\. Faust \(Eds\.\),Vol\.2026,pp\. 68606–68634\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2026/file/702b67152ec4435795f681865b67999c-Paper-Conference.pdf)Cited by:[§4\.2](https://arxiv.org/html/2611.17556#S4.SS2.SSS0.Px2.p1.1)\.
- Yuet al\.\(2020\)K\. Yu, X\. Guo, L\. Liu, J\. Li, H\. Wang, Z\. Ling, and X\. WuCausality\-based feature selection: methods and evaluations\.ACM Computing Surveys53\(5\),pp\. 111:1–111:36\.External Links:[Document](https://dx.doi.org/10.1145/3409382)Cited by:[§1](https://arxiv.org/html/2611.17556#S1.p1.1)\.

## Appendix AThe generator in detail

This appendix details the generator of Section[4](https://arxiv.org/html/2611.17556#S4): one entry per component with its mechanism and equation, then the order in which a batch is generated, the parameter ranges and the interventional audit\.

#### Core components\.

- •*Batch conditions\.*Themmlatent conditionsU1,…,UmU\_\{1\},\\dots,U\_\{m\}are linked in a random tree \(Table[4](https://arxiv.org/html/2611.17556#A1.T4)\)\. A condition without parents is standard Gaussian; any other is a weighted sum of its parents plus Gaussian noise\. The conditions enter the profile parameters with random coefficients that no share sets, so in Equation[3](https://arxiv.org/html/2611.17556#S4.E3)they belong to the remainderrr\.
- •*Confounders\.*After the direct causes are drawn, every other recorded variable, the actuator included, is assigned to one of them: thep−\|D\|p\-\|D\|variables are sorted by name and split into consecutive blocks of⌈\(p−\|D\|\)/\|D\|⌉\\lceil\(p\-\|D\|\)/\|D\|\\rceil, one per direct cause, the last block taking the remainder \(Table[4](https://arxiv.org/html/2611.17556#A1.T4)\)\. Each direct causeXkX\_\{k\}then receives a latent confounderZkZ\_\{k\}, a standard Gaussian drawn anew in each batch, which enters the profile parameters ofXkX\_\{k\}and of every variable in its block with the confounder sharesconfs\_\{\\mathrm\{conf\}\}\. A variable in the block ofXkX\_\{k\}therefore shares a latent cause withXkX\_\{k\}\. It is a sibling ofYYunless it is also an ancestor or a descendant ofYY, in which case it keeps that role\. Its within\-batch componentZk​\(t\)Z\_\{k\}\(t\), white noise smoothed with a Gaussian filter, is added to the trajectories ofXkX\_\{k\}and of its block\.
- •*Profile parameters\.*Each parameterθ\\thetais set from its parentsPa⁡\(θ\)\\mathrm\{Pa\}\(\\theta\)by θ=ℓθ\+\(hθ−ℓθ\)​ς​\(1\|Pa⁡\(θ\)\|​∑q∈Pa⁡\(θ\)q\+σθ​εθ\),\\theta=\\ell\_\{\\theta\}\+\(h\_\{\\theta\}\-\\ell\_\{\\theta\}\)\\,\\varsigma\\Big\(\\textstyle\\frac\{1\}\{\|\\mathrm\{Pa\}\(\\theta\)\|\}\\sum\_\{q\\in\\mathrm\{Pa\}\(\\theta\)\}q\+\\sigma\_\{\\theta\}\\,\\varepsilon\_\{\\theta\}\\Big\),\(4\)whereς\\varsigmais the logistic function,\[ℓθ,hθ\]\[\\ell\_\{\\theta\},h\_\{\\theta\}\]the admissible range ofθ\\theta\(Table[4](https://arxiv.org/html/2611.17556#A1.T4)\) andεθ\\varepsilon\_\{\\theta\}standard Gaussian noise\. The parents are batch conditions, confounders, phase timings and parameters of other recorded variables \(the batch\-level links\); the weights are equal except on the routes that carry a share\.
- •*Profiles\.*A recorded variableXiX\_\{i\}hasKi∈\{2,3,4\}K\_\{i\}\\in\\\{2,3,4\\\}motifs, each with a directiondk∈\{−1,\+1\}d\_\{k\}\\in\\\{\-1,\+1\\\}\(up or down\); their shapesμk\\mu\_\{k\}, directions and order are drawn once per plant and then fixed\. In each batch, Xi​\(t\)=λi\+∑k=1Kiϕk​dk​Ak​ψμk​\(t,ok,δk\)\+ηi​\(t\)\+νi​\(t\),X\_\{i\}\(t\)=\\lambda\_\{i\}\+\\sum\_\{k=1\}^\{K\_\{i\}\}\\phi\_\{k\}\\,d\_\{k\}\\,A\_\{k\}\\,\\psi\_\{\\mu\_\{k\}\}\(t;o\_\{k\},\\delta\_\{k\}\)\+\\eta\_\{i\}\(t\)\+\\nu\_\{i\}\(t\),\(5\)where the levelλi\\lambda\_\{i\}, the sizesAkA\_\{k\}, the onsetsoko\_\{k\}\(each within the slot of its motif\) and the durationsδk\\delta\_\{k\}are the profile parameters of the batch,ψμk\\psi\_\{\\mu\_\{k\}\}is the shape of motifkk,ϕk∈\{0,1\}\\phi\_\{k\}\\in\\\{0,1\\\}indicates whether the motif occurs in this batch,ηi\\eta\_\{i\}is a smooth random texture \(a Gaussian process\) andνi\\nu\_\{i\}is white noise\. The four shapes are two changes that persist \(a smooth step and a linear ramp\) and two that return \(a spike and a plateau with short edges\), each with onsetooand durationδ\\delta: step \(persists\):ψ\(t;o,δ\)=\(1\+e−\(t−o−δ/2\)/\(0\.15δ\)\)−1,\\displaystyle\\psi\(t;o,\\delta\)=\\big\(1\+e^\{\-\(t\-o\-\\delta/2\)/\(0\.15\\,\\delta\)\}\\big\)^\{\-1\},ramp \(persists\):ψ⁡\(t,o,δ\)=min⁡\(1,max⁡\(0,\(t−o\)/δ\)\),\\displaystyle\\psi\(t;o,\\delta\)=\\min\\\!\\big\(1,\\max\(0,\(t\-o\)/\\delta\)\\big\),spike \(returns\):ψ⁡\(t,o,δ\)=exp⁡\(−12​\(\(t−o−δ/2\)/\(δ/4\)\)2\),\\displaystyle\\psi\(t;o,\\delta\)=\\exp\\\!\\big\(\-\\tfrac\{1\}\{2\}\\big\(\(t\-o\-\\delta/2\)/\(\\delta/4\)\\big\)^\{2\}\\big\),plateau \(returns\):ψ⁡\(t,o,δ\)=min⁡\(h⁡\(t−o\),h⁡\(o\+δ−t\)\),\\displaystyle\\psi\(t;o,\\delta\)=\\min\\\!\\big\(h\(t\-o\),\\,h\(o\+\\delta\-t\)\\big\),withh⁡\(u\)=min⁡\(1,max⁡\(0,u/\(0\.15​δ\)\)\)h\(u\)=\\min\\\!\\big\(1,\\max\(0,u/\(0\.15\\,\\delta\)\)\\big\)\. In a batch plant, a heat\-up looks like a step, a dosing phase like a ramp, an exotherm \(the temperature spike from the heat released by the reaction\) like a spike and a holding phase like a plateau; the generator draws the shapes at random\. In each batch a motif occurs with a fixed probability, and its onset and size are jittered\.
- •*Time\-step links\.*A variableXjX\_\{j\}with upstream variables receives the sum of their delayed responses, mixed with its trajectory before the links,Xj0X\_\{j\}^\{0\}, as in Equation[3](https://arxiv.org/html/2611.17556#S4.E3): Xj​\(t\)=sup​a~j​\(t\)\+1−sup​X~j0​\(t\),aj​\(t\)=∑iβi​j​Fτi​j,ϑi​j​\[Xi\]​\(t\),X\_\{j\}\(t\)=\\sqrt\{s\_\{\\mathrm\{up\}\}\}\\,\\tilde\{a\}\_\{j\}\(t\)\+\\sqrt\{1\-s\_\{\\mathrm\{up\}\}\}\\,\\tilde\{X\}\_\{j\}^\{0\}\(t\),\\qquad a\_\{j\}\(t\)=\\sum\_\{i\}\\beta\_\{ij\}\\,\\mathrm\{F\}\_\{\\tau\_\{ij\},\\vartheta\_\{ij\}\}\[X\_\{i\}\]\(t\),withF\\mathrm\{F\}the first\-order\-plus\-dead\-time response of Section[4\.2](https://arxiv.org/html/2611.17556#S4.SS2.SSS0.Px2)\. Variables are coupled in causal order, so each receives its upstream variables already coupled; the standardising scales are fixed under intervention\. The gainβi​j\\beta\_\{ij\}, time constantτi​j\\tau\_\{ij\}and dead timeϑi​j\\vartheta\_\{ij\}of each link are drawn once per plant \(Table[4](https://arxiv.org/html/2611.17556#A1.T4)\)\.
- •*Controller\.*The controller holdsXcX\_\{c\}at the setpointρ⁡\(t\)\\rho\(t\)against the delayed disturbanceXdX\_\{d\}by moving the actuatorXaX\_\{a\}, Xc​\(t\)\\displaystyle X\_\{c\}\(t\)=ρ⁡\(t\)\+Fτd,ϑd​\[Xd\]​\(t\)\+Xa​\(t−1\),\\displaystyle=\\rho\(t\)\+\\mathrm\{F\}\_\{\\tau\_\{d\},\\vartheta\_\{d\}\}\[X\_\{d\}\]\(t\)\+X\_\{a\}\(t\-1\),\(6\)u⁡\(t\)\\displaystyle u\(t\)=u\(t−1\)\+κ\(ρ\(t\)−Xc\(t\)\),Xa\(t\)=Xa0\(t\)\+u\(t\),\\displaystyle=u\(t\-1\)\+\\kappa\\big\(\\rho\(t\)\-X\_\{c\}\(t\)\\big\),\\qquad X\_\{a\}\(t\)=X\_\{a\}^\{0\}\(t\)\+u\(t\),withu⁡\(t\)u\(t\)the accumulated correction andXa0X\_\{a\}^\{0\}the actuator’s own profile\.
- •*Target\.*The interaction terms are products of pairs of direct causes, chosen so that every direct cause is in at least one pair; each product has its own gain\. The unexplained partε\\varepsilonofYY\(Equation[2](https://arxiv.org/html/2611.17556#S4.E2)\) stands for causes ofYYthat are not recorded\. Like the recorded causes, these change slowly, soε\\varepsilonis drawn as a slowly varying random process: an AR\(1\) whose autocorrelation falls to1/e1/eat the decorrelation time ofSS\. Ifε\\varepsilonwere white noise, the block means the methods receive would average it away, and the direct causes would explain more ofYYthanssigs\_\{\\mathrm\{sig\}\}sets\.
- •*Downstream measurements\.*Each receives the delayed response toYY, with a dead time set from the decorrelation time ofYY, mixed with its own profile by the downstream sharesdescs\_\{\\mathrm\{desc\}\}\.

#### Additional components\.

- •*Recipe phases\.*The batch is divided into four equal phase windows\. Each windowqqhas a latent timingΦq\\Phi\_\{q\}, drawn once per batch, that states how early or late the batch runs through that phase; it is a parent, in Equation[4](https://arxiv.org/html/2611.17556#A1.E4), of the onset of every motif whose slot lies in the window, so the motifs of all variables in that window shift together\. The weight ofΦq\\Phi\_\{q\}is set so that it carries the sharesphases\_\{\\mathrm\{phase\}\}of each such onset’s variance: before the data are drawn, the generator simulates 3,000 batches of the profile parameters, measures the variance that the onset’s other parents contribute and solves for the weight\. The equation is linear, so the share holds exactly\. The weights of the confounders are set in the same way\.
- •*Lots and drift\.*Both are set on each variable’s*batch\-to\-batch variation*, the variance over batches of its batch meanX¯i,b\\bar\{X\}\_\{i,b\}; this includes everything that changes between batches \(the profile parameters, whether a motif occurs, its timing and size\)\. The lot processLbL\_\{b\}is an AR\(1\) over the batch index and the driftRbR\_\{b\}a linear trend in it; the lot models consecutive batches sharing a raw\-material lot \(persistent but random, with a random signζi\\zeta\_\{i\}per variable\), the drift a slow change over the campaign such as fouling or catalyst ageing \(the same direction for every variable\); each variable’s trajectory is shifted by a constant per batch such that, in standardised units, X¯i,b→1−slot−sdrift​X¯~i,b\+slot​ζi​L~b\+sdrift​R~b,\\bar\{X\}\_\{i,b\}\\;\\to\\;\\sqrt\{1\-s\_\{\\mathrm\{lot\}\}\-s\_\{\\mathrm\{drift\}\}\}\\;\\tilde\{\\bar\{X\}\}\_\{i,b\}\+\\sqrt\{s\_\{\\mathrm\{lot\}\}\}\\,\\zeta\_\{i\}\\,\\tilde\{L\}\_\{b\}\+\\sqrt\{s\_\{\\mathrm\{drift\}\}\}\\,\\tilde\{R\}\_\{b\},\(7\)with a signζi\\zeta\_\{i\}per variable, as in Equation[3](https://arxiv.org/html/2611.17556#S4.E3)\. The shape of the trajectory within the batch is unchanged, and the shift is applied before the time\-step links, the controller and the target, so lots and drift reachYYonly through its direct causes\.
- •*Sensor error\.*The methods receive readings, not true values: each variable is recorded asXirec=1−serr​X~i\+serr​e~iX\_\{i\}^\{\\mathrm\{rec\}\}=\\sqrt\{1\-s\_\{\\mathrm\{err\}\}\}\\,\\tilde\{X\}\_\{i\}\+\\sqrt\{s\_\{\\mathrm\{err\}\}\}\\,\\tilde\{e\}\_\{i\}, so the error carries the shareserrs\_\{\\mathrm\{err\}\}of the reading’s variance\. The erroreie\_\{i\}varies slowly, like a drifting sensor: an AR\(1\) with the decorrelation time ofXiX\_\{i\}, so that the block means do not average it away\. The plant itself runs on the true values; the links, the controller and the target never see the error\.
- •*Product grades\.*The gradeγ∈\{−1,\+1\}\\gamma\\in\\\{\-1,\+1\\\}of each batch is drawn with equal probability and is not recorded\. It multiplies the weight of each direct cause by\(1\+c​ζk​γ\)/1\+c2\(1\+c\\,\\zeta\_\{k\}\\gamma\)/\\sqrt\{1\+c^\{2\}\}, wherec∈\[0,1\]c\\in\[0,1\]is the grade contrast \(Section[4\.4](https://arxiv.org/html/2611.17556#S4.SS4)\) and the signζk∈\{−1,\+1\}\\zeta\_\{k\}\\in\\\{\-1,\+1\\\}of each cause selects the grade in which that cause acts more strongly\. Atc=0c=0the two grades are identical; atc=1c=1each cause acts in one grade only\. The squared multiplier averages to one over the two grades, so the average strength of each cause does not depend oncc\.
- •*Variable delays\.*In each batch, the dead timeϑk\\vartheta\_\{k\}of each direct cause is increased by a hold\-up drawn uniformly between zero andjjtimes the decorrelation time of the cause, so the response ofYYto the cause falls at different lags from batch to batch\.
- •*Cumulative effects\.*Some causes act through their accumulated exposure, e\.g\. the heat a product has received over the last hour, and not only through their current value\. The response ofYYto a cause is therefore a mix of two parts, both standardised and combined as in Equation[3](https://arxiv.org/html/2611.17556#S4.E3)with weightkavgk\_\{\\mathrm\{avg\}\}: the cause’s current delayed response, and its running mean over the last two decorrelation times of the cause\. Atkavg=0k\_\{\\mathrm\{avg\}\}=0only the current value acts; atkavg=1k\_\{\\mathrm\{avg\}\}=1only the accumulated one\. The two parts are correlated, sokavgk\_\{\\mathrm\{avg\}\}sets their relative weight but is not the share of the response’s variance that each carries \(Section[4\.3](https://arxiv.org/html/2611.17556#S4.SS3)\)\.
- •*Operator reactions\.*An operator watchesYYand corrects it through one direct cause, chosen among those that act onYYmonotonically \(linear, saturating or thresholded\), so that countering a deviation has a direction\. The operator seesYYthrough a first\-order lag with time constantτY\\tau\_\{Y\}, delayed byτY\\tau\_\{Y\}, the decorrelation time ofYY, and moves the reference value of the cause against it: the cause becomes1−sop\\sqrt\{1\-s\_\{\\mathrm\{op\}\}\}times its own trajectory plussop\\sqrt\{s\_\{\\mathrm\{op\}\}\}times this reaction, both standardised as in Equation[3](https://arxiv.org/html/2611.17556#S4.E3)\. Over the variables,Y→X→YY\\to X\\to Yis a cycle; over time it is not, since the cause at timettdepends only onYYup tot−τYt\-\\tau\_\{Y\}\. The generator therefore computes the batch forward in steps ofτY\\tau\_\{Y\}, each fixed by the previous ones; Appendix[B](https://arxiv.org/html/2611.17556#A2)treats the cycle in the separation labels\.

#### Order of generation\.

Algorithm[A](https://arxiv.org/html/2611.17556#A1.SS0.SSS0.Px3)gives the order in which one batch is generated\.

\{algorithm2e\}

\[h\]Generation of one batch of a plant\.\\DontPrintSemicolonDraw the batch conditionsUjU\_\{j\}, the confoundersZkZ\_\{k\}and the phase timingsΦq\\Phi\_\{q\}\\Foreach profile parameterθ\\theta, in causal orderθ←\\theta\\leftarrowEquation[4](https://arxiv.org/html/2611.17556#A1.E4), from its parents among theUjU\_\{j\},ZkZ\_\{k\},Φq\\Phi\_\{q\}and the parameters of other variables\\Foreach recorded variableXiX\_\{i\}Xi​\(t\)←X\_\{i\}\(t\)\\leftarrowits profile with timing and sizes fromθ⁡\(Xi\)\\theta\(X\_\{i\}\), plus texture \(Equation[5](https://arxiv.org/html/2611.17556#A1.E5)\) Add the within\-batch partZk​\(t\)Z\_\{k\}\(t\)of each confounder to its direct cause and to the variables of its block Shift each variable by its lot and drift contribution for this batch \(Equation[7](https://arxiv.org/html/2611.17556#A1.E7)\)\\Foreach recorded variable with upstream variables, in causal order mix in the delayed responses of its upstream variables \(time\-step links\) Close the control loop \(Equation[6](https://arxiv.org/html/2611.17556#A1.E6)\)Y⁡\(t\)←Y\(t\)\\leftarrowresponse to the direct causes, with product grades, variable delays and cumulative effects if set, plus the unexplained part \(Equations[1](https://arxiv.org/html/2611.17556#S4.E1)and[2](https://arxiv.org/html/2611.17556#S4.E2)\)\\Ifoperator reactions are set move the reference value of one direct cause against the past ofYYand recomputeYY, forward in steps ofτY\\tau\_\{Y\}Add the delayed response toYYto each downstream measurement\\Ifsensor error is set record each variable as a mix of its true value and its sensor error

#### Parameters and seeds\.

Table[4](https://arxiv.org/html/2611.17556#A1.T4)lists the parameter ranges used in all experiments\. Values are drawn once per plant unless marked per batch; time is in steps\. Each plant has two seeds: a structure seed, which draws the plant, and a data seed, which draws only its batches, so datasets drawn with different data seeds share the plant’s structure\.

Table 4:Parameter ranges of the generator in all experiments\.
#### Realised shares\.

Table[5](https://arxiv.org/html/2611.17556#A1.T5)compares every share with its value measured on the generated plants, at the resolution and look\-back the methods use, as an out\-of\-sampleR2R^\{2\}\(five folds split by batch; ridge regression on the standardised lags0,…,L0,\\dots,L\), on the 10 plants of each setting with all else at the reference\. The signal share is theR2R^\{2\}ofYYfrom its direct causes, with the squares of their lags and the products of pairs, so that nonlinear and joint effects count; the upstream share is theR2R^\{2\}of a variable from the variables upstream of it; the downstream share is theR2R^\{2\}of a downstream measurement fromYY; the interaction share is the part of the causes’R2R^\{2\}that only the products carry\. The confounder share acts in two stages\. Its share of the profile parameters is set exactly by calibration, but the parameters are not recorded\. Its within\-batch componentZk​\(t\)Z\_\{k\}\(t\)is added to the trajectory of the cause and of each sibling with a strength drawn per variable \(Table[4](https://arxiv.org/html/2611.17556#A1.T4)\), multiplied byg=s/\(1−s\)/s0/\(1−s0\)g=\\sqrt\{s/\(1\-s\)\}\\,/\\sqrt\{s\_\{0\}/\(1\-s\_\{0\}\)\}withs0=0\.3s\_\{0\}=0\.3, so that the within\-batch oddsq/\(1−q\)q/\(1\-q\)of each variable change by the factorg2g^\{2\}from the reference, whereqq, the share of the variable’s within\-batch variance due toZk​\(t\)Z\_\{k\}\(t\), is measured as1−corr21\-\\mathrm\{corr\}^\{2\}between the variable with and withoutZk​\(t\)Z\_\{k\}\(t\), all else fixed \(10 plants per setting\)\. Three measured values fall below their settings, each for a known reason\. A faint signal \(ssig≤0\.1s\_\{\\mathrm\{sig\}\}\\leq 0\.1\) is estimated low, as an out\-of\-sampleR2R^\{2\}is when the signal is weak\. At high interaction shares, the squares of the causes in the regression absorb part of the joint effect, so the share that only the products carry reads low\. Atsconf=0\.8s\_\{\\mathrm\{conf\}\}=0\.8, part of the within\-batch component is passed on to other variables by the time\-step links or countered by the controller, so less of it remains in the variable than was added\. Siblings are also associated with their cause through the shared batch conditions: with both stages of the confounder removed, theR2R^\{2\}a sibling shares with its cause is 0\.06, against 0\.10 at the reference \(median over 10 plants\)\.

Table 5:Configured and realised shares: each setting, with all else at the reference \(in bold\), against its value measured on the generated plants \(median and interquartile range\)\. For the confounder share, the measured quantity is the factor by which the within\-batch odds of a cause or sibling change from the reference, designed to beg2g^\{2\}\.ShareSettingExpectedMeasured, median \[IQR\]ssigs\_\{\\mathrm\{sig\}\}0\.050\.050\.02 \[0\.01, 0\.04\]0\.10\.10\.07 \[0\.06, 0\.10\]0\.20\.20\.18 \[0\.16, 0\.21\]0\.40\.40\.39 \[0\.36, 0\.42\]0\.70\.70\.69 \[0\.68, 0\.71\]0\.90\.90\.88 \[0\.87, 0\.89\]sups\_\{\\mathrm\{up\}\}0\.10\.10\.11 \[0\.10, 0\.14\]0\.30\.30\.29 \[0\.26, 0\.34\]0\.60\.60\.59 \[0\.57, 0\.61\]0\.80\.80\.79 \[0\.78, 0\.80\]sdescs\_\{\\mathrm\{desc\}\}0\.10\.10\.11 \[0\.09, 0\.15\]0\.250\.250\.25 \[0\.24, 0\.32\]0\.50\.50\.50 \[0\.48, 0\.56\]sints\_\{\\mathrm\{int\}\}000\.00 \[\-0\.01, 0\.00\]0\.250\.250\.21 \[0\.14, 0\.24\]0\.50\.50\.42 \[0\.31, 0\.47\]0\.750\.750\.65 \[0\.50, 0\.71\]110\.82 \[0\.66, 0\.95\]sconfs\_\{\\mathrm\{conf\}\}0\.1,g=0\.51g=0\.51g2=0\.26g^\{2\}=0\.260\.26 \[0\.26, 0\.27\]0\.3,g=1\.00g=1\.00g2=1\.00g^\{2\}=1\.001 \(by definition\)0\.6,g=1\.87g=1\.87g2=3\.50g^\{2\}=3\.503\.37 \[3\.21, 3\.48\]0\.8,g=3\.06g=3\.06g2=9\.33g^\{2\}=9\.338\.59 \[7\.69, 9\.21\]
#### Interventional audit\.

The roles are read on the generating graph; the audit checks them against the generator’s behaviour on one plant of every configuration \(119 plants, 10%\)\. \(i\) The plant is rebuilt from its configuration and seeds and replayed; the replayedYYmust equal the released one\. \(ii\) For every recorded variable, one intervention at a time with all exogenous noise held fixed: its trajectory is shifted,do⁡\(Xi:=Xi\+ξ\)\\mathrm\{do\}\(X\_\{i\}:=X\_\{i\}\+\\xi\)withξ\\xiwhite noise of the variable’s standard deviation, and each of its profile parameters is shifted by one standard deviation and propagated through the SCM\. The effect is the root\-mean\-square change ofYYin units of its standard deviation\. \(iii\) A direct cause or an ancestor must have a non\-zero effect; a sibling or a descendant must have an effect of exactly zero \(below10−1210^\{\-12\}\); and every profile parameter must change the trajectory of its own variable, since a parameter without effect would be a false edge\.

## Appendix BGround truth and separation labels

#### The full graph\.

The generator builds and exports the full causal graph of every plant at three levels: the fine\-grained structural causal model, with every latent batch condition, confounder, phase timing, profile parameter and trajectory as a node \(about 210 nodes for 20 recorded variables\); its latent projection onto the recorded variables, with a directed edge where one variable acts on another directly or through latent nodes only and a bidirected edge where two share a latent cause; and the links between recorded variables, split into batch\-level links and time\-step links with their gains, time constants and dead times\.

#### Latent variables\.

The batch conditions and confounders are left out of the data by choice: a plant rarely records ambient conditions, utilities, raw\-material properties or the crew on duty, although they act on every batch\. Because the batch conditions, linked by their tree, act on the profiles of every recorded variable, almost every two recorded variables share an unrecorded cause: of the 190 pairs formed by the 20 recorded variables of a reference plant, at least 189 do in every plant\. This is the trade\-off of a realistic plant: the full graph is hard to recover from such data, while every wrong candidate remains separable from the direct causes of the target without sensor error\.

#### The benchmark is target\-centric\.

Separability depends on the structure of the plant\. Figure[3](https://arxiv.org/html/2611.17556#A2.F3)shows two examples of structures that break it: an unrecorded variable acting onYYdirectly, which leaves a path from a sibling toYYopen given the direct causes, and feedback within one time step, which makes a variable at once a cause and a consequence ofYY\. The generator is built so that every wrong candidate is separable fromYY, given the direct causes or by time order, and the label of each candidate, computed on the graph of its plant, verifies this\. In the reference configuration all 78 siblings over 10 plants are confounded with the target, yet all 164 wrong candidates have one of these labels: 141 by conditioning and 23 by time order\. Sensor error can prevent separation given the recorded causes\. The difficulty is set around the target as well: the strength of its causes, the confounders they share with siblings, the controller on one of them and the measurements downstream of it\. The full graph is released with every dataset and can be used to evaluate full\-graph causal discovery, with a metric that accounts for hidden common causes\.

Figure 3:Two structures that break separability \(left in each panel\) and the corresponding structures of the generator \(right\)\. \(a\) An unrecorded variableUUacting on the target directly: conditioning on the direct causeXXdoes not separate the siblingLLfromYY\. In the generator, unrecorded variables reachYYonly through its direct causes\. \(b\)XXacting onYYandYYback onXXwithin one time step, as a controller or an operator reacting at once would:XXis both a cause and a consequence ofYYand has no role\. In the generator, every feedback loop acts with a delay of at least one step, so that, unrolled in time,X⁡\(t\)X\(t\)is a cause ofY⁡\(t\)Y\(t\)andY⁡\(t\)Y\(t\)a cause ofX⁡\(t\+1\)X\(t\+1\)\.
#### Resolution and look\-back\.

Both are set from the data alone, through the memory ofYY: for each plant,τY\\tau\_\{Y\}is the lag at which the autocorrelation ofYYwithin a batch, normalised by its lag\-1 value, falls below1/e1/e\. The methods receive the mean of everyffsteps, withffthe largest power of two not above the smallestτY\\tau\_\{Y\}in the configuration, and we verify that every descendant ofYYresponds to it with a delay of at leastffsteps, so that after averaging its response still followsYYby at least one averaged point\. The methods look backL=2​τYL=2\\tau\_\{Y\}of the slowest plant, and every cause acts within that window\. Both hold in every configuration\. The rule forffgivesf=32f=32in all configurations except two with 50 batches, where the estimate of the smallestτY\\tau\_\{Y\}falls below 32 andf=16f=16; the look\-back is 3 to 8 blocks, 4 at the reference\. Averaging leaves 9% of the wrong candidates not separable from the direct causes by a linear test, against 2% at a resolution of 4 steps\.

#### Separation labels\.

For every wrong candidate, we record whether a method could rule it out in principle, computed on the true graph\. It is labelled*by conditioning*when it isdd\-separated fromYYgiven the direct causesDD\([Spirtes et al\., 2000](https://arxiv.org/html/2611.17556#bib.bib35)\): once the direct causes are known, the candidate carries no further information aboutYY\. It is labelled*by time order*when it is a descendant that responds toYYat least one time step later at the evaluation resolution \(Section[5](https://arxiv.org/html/2611.17556#S5)\), and*neither criterion*otherwise\. Outside the configurations with sensor error, every label is by conditioning or by time order\. The test runs on the summary graph, with one node per variable grouping its trajectory and profile parameters\. A controller closes a loop in this graph, anddd\-separation can then give wrong answers; we useσ\\sigma\-separation, its counterpart for cyclic causal models\([Forré and Mooij, 2017](https://arxiv.org/html/2611.17556#bib.bib25);[Forré and Mooij, 2018](https://arxiv.org/html/2611.17556#bib.bib26)\), computed asdd\-separation on an acyclic version of the graph in which the variables of each loop share common causes\([Mooij and Claassen, 2020](https://arxiv.org/html/2611.17556#bib.bib27)\)\. An operator who reacts toYYcloses a loop only over time, since the reaction follows past values ofYY; the separation test therefore conditions on the direct causes and on the past ofYY, as every lag\-based method does, and the loop leaves the labels unchanged\. Under sensor error, a method conditions on the readings, not on the true values, and conditioning on a noisy reading does not remove all of the information the true value carries\([Scheines and Ramsey, 2016](https://arxiv.org/html/2611.17556#bib.bib28)\)\. We therefore also compute the labels on the graph in which each true value is unrecorded and only its reading is recorded; separation given the true causes need not hold given their readings\. Both versions are recorded\.

#### Separability at the evaluation resolution\.

The methods see block means offfsteps \(Section[5](https://arxiv.org/html/2611.17556#S5)\)\. Averaging hides the movement of a direct cause within a block, so a wrong candidate that drives or follows it closely can still carry information aboutYYgiven the averaged causes\([Gong et al\., 2017](https://arxiv.org/html/2611.17556#bib.bib29)\)\. We therefore test every wrong candidate on every plant on the data the methods see: does it improve the prediction ofYYbeyond the direct causes? We fit two linear regressions ofY⁡\(t\)Y\(t\)on the block means, pooled over the batches of the plant: the first on the past ofYY\(lags 1 toLL\) and on the direct causes \(lags 0 toLL\), withLLthe common look\-back of Section[5](https://arxiv.org/html/2611.17556#S5); the second adds the candidate at the same lags\. AnFF\-test compares the two fits\. A candidate is*separable at the evaluation resolution*if it adds no significant improvement \(p≥0\.01p\\geq 0\.01\), so that a linear conditional\-independence test, like those of the baselines, would rule it out given the direct causes\. Each candidate that is not separable is tested again at a resolution of 4 steps, close to the full resolution, with the same look\-back: if it is separable there, averaging is the reason\. The remaining candidates fail the linear test for other reasons, such as a nonlinear response, sensor error, or chance\. Table[6](https://arxiv.org/html/2611.17556#A2.T6)gives the shares by role over the 23,467 wrong candidates, and Table[7](https://arxiv.org/html/2611.17556#A2.T7)the share of each method’s wrong selections that are not separable, with the list rescored without them\. The label is released with every plant\.

Table 6:Wrong candidates not separable from the direct causes at the evaluation resolution, by role, over all plants\. The same test at a resolution of 4 steps, close to the full resolution, splits them into two groups:*by averaging*, candidates separable at 4 steps \(p≥0\.05p\\geq 0\.05\), and*also at 4 steps*, candidates not separable there either\.Table 7:Wrong selections of each method that are not separable at the evaluation resolution\. Left: their share of the method’s wrong selections, split as in Table[6](https://arxiv.org/html/2611.17556#A2.T6)into*by averaging*and*also at 4 steps*, and their share of all its selections \(*share of list*\)\. Right: first pick, precision and lift as in Table[3](https://arxiv.org/html/2611.17556#S6.T3), then recomputed with the non\-separable candidates removed from the candidates of each plant\.

## Appendix CAdditional results

#### Errors by role\.

Table[8](https://arxiv.org/html/2611.17556#A3.T8)breaks the wrong selections of each method down by the role of the selected candidate, over all plants of all configurations, and compares each share with the share of that role among all wrong candidates\. Section[6](https://arxiv.org/html/2611.17556#S6)reads the table\.

Table 8:Errors of each method over the 1,190 plants of all 119 configurations: the share of each role among the method’s wrong selections, against the share among all wrong candidates \(*random choice*\)\. A cell is shaded when the method picks that role more often than a random choice would, darker the more often\.*Nothing selected*: share of plants on which the method selects nothing\. The disturbance of the controlled cause \(Section[4\.2](https://arxiv.org/html/2611.17556#S4.SS2.SSS0.Px2)\) is an ordinary ancestor with the controller off, as in all configurations where it is scored\.
#### All configurations\.

Figure[4](https://arxiv.org/html/2611.17556#A3.F4)shows the selected lists for two single difficulties, as Figure[2](https://arxiv.org/html/2611.17556#S6.F2)does for the reference and the combined difficulties\. Table[9](https://arxiv.org/html/2611.17556#A3.T9)gives, for every one of the 119 configurations, the first pick, the mAP of the ranking and the precision of the selected list of every method; the configurations are grouped by the difficulty they vary, with the symbols of Sections[4\.3](https://arxiv.org/html/2611.17556#S4.SS3)and[4\.4](https://arxiv.org/html/2611.17556#S4.SS4)\.

![Refer to caption](https://arxiv.org/html/2611.17556v1/fig_money_appx.png)Figure 4:The selected lists, as in Figure[2](https://arxiv.org/html/2611.17556#S6.F2), for two single difficulties: a faint signal \(signal share 0\.1, top\) and causes that act only jointly \(interaction share 1, bottom\)\.Table 9:Every configuration and method, grouped by the share or component that the configuration changes:*first pick*, plants \(of 10\) whose first selected candidate is a direct cause;*mAP*, mean average precision of the ranking;*precision*, of the selected list over the plants where the method selects something \(– when it never does\)\. Each row departs from the reference configuration in the parameters named;*combined difficulties*:ssigs\_\{\\mathrm\{sig\}\}0\.3,sconf=sups\_\{\\mathrm\{conf\}\}=s\_\{\\mathrm\{up\}\}0\.6,sdescs\_\{\\mathrm\{desc\}\}0\.5 andsints\_\{\\mathrm\{int\}\}1 \(Figure[2](https://arxiv.org/html/2611.17556#S6.F2), bottom\)\. Darker shading means a higher value\.first pickmAPprecisionConfigurationCorrelation

Granger

VARLiNGAM

DYNOTEARS

PC

PCMCI\+

SpaceTime

DoubleML

Correlation

Granger

VARLiNGAM

DYNOTEARS

PC

PCMCI\+

SpaceTime

DoubleML

Correlation

Granger

VARLiNGAM

DYNOTEARS

PC

PCMCI\+

SpaceTime

DoubleML

Reference\(5\)sints\_\{\\mathrm\{int\}\}0 \(reference\)10101063108100\.790\.780\.840\.940\.310\.960\.770\.940\.140\.150\.331\.000\.350\.790\.850\.70100 batches991062108100\.830\.760\.870\.930\.310\.960\.710\.880\.150\.170\.271\.000\.370\.780\.940\.8350 batches9810839890\.820\.750\.830\.950\.440\.890\.630\.870\.160\.250\.461\.000\.500\.880\.940\.8225 batches999627960\.800\.700\.670\.880\.390\.810\.730\.610\.200\.480\.420\.920\.450\.651\.000\.6040 variables88104110890\.640\.630\.780\.960\.220\.920\.650\.770\.070\.080\.251\.000\.290\.710\.800\.77Signal sharessigs\_\{\\mathrm\{sig\}\}\(reference 0\.7\)\(9\)ssigs\_\{\\mathrm\{sig\}\}0\.910101073108100\.810\.780\.890\.940\.320\.970\.780\.940\.140\.150\.241\.000\.360\.730\.850\.78ssigs\_\{\\mathrm\{sig\}\}0\.48108049590\.720\.790\.630\.940\.300\.870\.600\.840\.150\.160\.26–0\.350\.810\.710\.75ssigs\_\{\\mathrm\{sig\}\}0\.2496039180\.540\.770\.530\.920\.270\.870\.390\.780\.160\.230\.27–0\.300\.720\.200\.93ssigs\_\{\\mathrm\{sig\}\}0\.1084038050\.400\.740\.440\.860\.290\.790\.290\.690\.180\.280\.40–0\.320\.650\.001\.00100 batches073028210\.410\.680\.440\.770\.240\.700\.350\.510\.210\.470\.36–0\.290\.670\.330\.7550 batches385015030\.470\.790\.420\.560\.220\.710\.380\.570\.240\.520\.45–0\.230\.57–0\.6725 batches012025000\.340\.370\.420\.410\.170\.410\.250\.180\.240\.500\.29–0\.140\.940\.000\.0040 variables173025430\.420\.560\.360\.770\.160\.720\.410\.490\.130\.180\.25–0\.210\.470\.500\.75ssigs\_\{\\mathrm\{sig\}\}0\.05062024030\.340\.640\.300\.560\.230\.620\.260\.520\.220\.550\.26–0\.250\.440\.000\.92Confounder sharesconfs\_\{\\mathrm\{conf\}\}\(reference 0\.3\)\(7\)sconfs\_\{\\mathrm\{conf\}\}0\.11091063109100\.800\.760\.810\.940\.320\.970\.760\.950\.140\.150\.231\.000\.360\.740\.940\.74sconfs\_\{\\mathrm\{conf\}\}0\.689933108100\.670\.700\.800\.940\.290\.930\.700\.940\.140\.150\.281\.000\.350\.800\.850\.73sconfs\_\{\\mathrm\{conf\}\}0\.879534106100\.570\.670\.540\.900\.320\.920\.590\.900\.140\.150\.191\.000\.390\.810\.620\.69100 batches796348680\.580\.670\.560\.880\.340\.830\.570\.800\.140\.160\.331\.000\.470\.640\.580\.8250 batches697347680\.570\.660\.610\.880\.330\.850\.570\.750\.130\.190\.391\.000\.430\.700\.500\.7225 batches677453730\.550\.640\.700\.820\.370\.610\.620\.590\.160\.290\.401\.000\.530\.440\.700\.3940 variables687138780\.420\.520\.620\.770\.230\.870\.530\.750\.060\.070\.221\.000\.350\.710\.510\.84Upstream sharesups\_\{\\mathrm\{up\}\}\(reference 0\.3\)\(7\)sups\_\{\\mathrm\{up\}\}0\.19101063108100\.760\.810\.910\.940\.330\.980\.790\.950\.150\.150\.361\.000\.380\.750\.940\.77sups\_\{\\mathrm\{up\}\}0\.6109754108100\.790\.760\.650\.910\.300\.960\.730\.920\.140\.140\.221\.000\.330\.650\.850\.81sups\_\{\\mathrm\{up\}\}0\.8995439790\.730\.720\.540\.890\.300\.950\.680\.810\.140\.150\.221\.000\.340\.700\.720\.60100 batches998437790\.760\.720\.640\.880\.280\.860\.640\.750\.140\.170\.241\.000\.320\.780\.700\.6750 batches978539670\.730\.670\.620\.920\.330\.870\.630\.670\.150\.210\.381\.000\.390\.870\.550\.8325 batches1086546750\.700\.700\.570\.890\.360\.710\.680\.480\.180\.350\.261\.000\.470\.560\.720\.6740 variables876319780\.570\.590\.510\.950\.160\.810\.530\.660\.070\.080\.171\.000\.240\.570\.700\.67Downstream sharesdescs\_\{\\mathrm\{desc\}\}\(reference 0\.25\)\(2\)sdescs\_\{\\mathrm\{desc\}\}0\.110101063108100\.800\.780\.880\.940\.300\.960\.760\.930\.140\.140\.291\.000\.350\.790\.850\.66sdescs\_\{\\mathrm\{desc\}\}0\.5810763108100\.690\.780\.720\.940\.300\.960\.770\.940\.140\.150\.261\.000\.350\.740\.850\.71Interaction sharesints\_\{\\mathrm\{int\}\}\(reference 0\)\(8\)sints\_\{\\mathrm\{int\}\}0\.251010104410990\.790\.770\.810\.890\.270\.880\.790\.870\.150\.150\.241\.000\.310\.810\.950\.64sints\_\{\\mathrm\{int\}\}0\.5999149990\.750\.730\.760\.870\.300\.800\.830\.770\.150\.180\.291\.000\.320\.890\.950\.80sints\_\{\\mathrm\{int\}\}0\.75898047980\.670\.820\.690\.860\.330\.750\.830\.790\.170\.230\.31–0\.370\.700\.950\.73sints\_\{\\mathrm\{int\}\}1\.0282018930\.380\.680\.410\.670\.210\.750\.760\.530\.200\.290\.29–0\.140\.700\.950\.75100 batches141005820\.360\.580\.360\.540\.150\.590\.690\.450\.180\.360\.12–0\.110\.620\.850\.5650 batches233126820\.360\.440\.400\.440\.190\.600\.700\.370\.200\.190\.351\.000\.100\.580\.800\.4025 batches332123900\.400\.440\.420\.390\.250\.400\.750\.290\.230\.390\.331\.000\.170\.311\.000\.5040 variables1640061020\.250\.450\.420\.500\.090\.550\.830\.300\.080\.160\.32–0\.060\.370\.950\.50Controller tightnessκ\\kappa\(8\)κ\\kappa0\.39983310880\.730\.830\.750\.940\.280\.920\.670\.850\.150\.190\.311\.000\.310\.730\.800\.75κ\\kappa0\.69993310890\.740\.810\.800\.940\.280\.920\.670\.860\.150\.180\.341\.000\.320\.830\.800\.81κ\\kappa1\.09993310880\.740\.810\.770\.940\.280\.900\.670\.850\.160\.190\.331\.000\.320\.830\.800\.77ssigs\_\{\\mathrm\{sig\}\}0\.2 andκ\\kappa1284028170\.470\.750\.470\.850\.260\.740\.380\.710\.180\.280\.43–0\.270\.720\.330\.86100 batches685037140\.530\.710\.500\.830\.270\.690\.310\.600\.200\.490\.54–0\.260\.730\.200\.9250 batches585018060\.490\.750\.700\.690\.260\.780\.430\.620\.230\.600\.54–0\.280\.83–0\.7925 batches326015110\.400\.490\.560\.500\.200\.480\.280\.320\.280\.670\.61–0\.200\.560\.200\.2040 variables365048360\.390\.510\.440\.790\.150\.780\.340\.530\.090\.230\.32–0\.220\.560\.380\.74Number of direct causes \(reference 1–4\)\(1\)4–6 direct causes89924109100\.610\.630\.740\.790\.380\.830\.610\.820\.290\.280\.421\.000\.460\.860\.900\.82Combined difficulties\(15\)combined, without joint effects483027680\.500\.710\.410\.900\.270\.780\.570\.750\.140\.160\.21–0\.320\.730\.600\.84100 batches593027460\.540\.750\.390\.920\.290\.870\.440\.590\.150\.230\.25–0\.370\.670\.500\.8350 batches564036310\.500\.630\.440\.790\.280\.740\.410\.480\.160\.410\.29–0\.390\.860\.280\.2525 batches534024300\.540\.630\.490\.810\.300\.580\.420\.350\.190\.400\.36–0\.390\.470\.380\.1740 variables372016770\.380\.500\.310\.770\.190\.710\.460\.610\.070\.090\.18–0\.270\.520\.650\.86combined, without joint effects,ssigs\_\{\\mathrm\{sig\}\}0\.15 and 0\.8263015230\.390\.640\.360\.670\.210\.660\.350\.570\.150\.180\.24–0\.270\.550\.240\.62100 batches270005320\.400\.670\.230\.710\.210\.650\.350\.520\.150\.310\.18–0\.230\.720\.280\.5850 batches351004300\.430\.620\.260\.680\.170\.530\.340\.300\.160\.360\.21–0\.170\.580\.440\.0025 batches423113300\.410\.520\.380\.550\.180\.450\.330\.200\.210\.280\.221\.000\.150\.500\.310\.0040 variables252012220\.280\.400\.260\.510\.100\.580\.240\.410\.070\.080\.13–0\.150\.310\.110\.62combined difficulties040014610\.250\.610\.240\.710\.180\.600\.500\.360\.110\.200\.16–0\.050\.580\.530\.17100 batches020012310\.290\.530\.220\.530\.180\.520\.410\.380\.120\.170\.10–0\.050\.250\.441\.0050 batches121023210\.300\.390\.270\.340\.200\.410\.360\.310\.100\.170\.19–0\.120\.620\.290\.7525 batches121021500\.290\.360\.400\.270\.230\.400\.540\.300\.080\.440\.340\.000\.200\.170\.64–40 variables050001610\.180\.360\.190\.380\.120\.240\.470\.230\.060\.160\.17–0\.070\.230\.630\.25Recipe phasessphases\_\{\\mathrm\{phase\}\}\(7\)sphases\_\{\\mathrm\{phase\}\}0\.25109963109100\.810\.770\.870\.940\.310\.940\.790\.910\.140\.140\.351\.000\.360\.790\.850\.72sphases\_\{\\mathrm\{phase\}\}0\.5109963109100\.810\.770\.820\.940\.310\.930\.790\.920\.140\.140\.311\.000\.360\.760\.850\.69sphases\_\{\\mathrm\{phase\}\}0\.751091063108100\.810\.760\.830\.940\.320\.930\.750\.940\.140\.140\.281\.000\.360\.690\.850\.7850 batches10810539790\.810\.730\.730\.920\.420\.930\.630\.850\.180\.320\.491\.000\.510\.840\.720\.7340 variables98103198100\.650\.630\.770\.920\.220\.920\.650\.830\.070\.080\.271\.000\.300\.810\.800\.88with the combined difficulties030023710\.240\.540\.280\.450\.200\.600\.560\.310\.070\.170\.18–0\.130\.330\.600\.25sphases\_\{\\mathrm\{phase\}\}0\.9109106310890\.800\.750\.870\.940\.320\.940\.750\.920\.150\.150\.331\.000\.370\.660\.800\.71Raw\-material lots and campaign drift\(10\)slots\_\{\\mathrm\{lot\}\}0\.251010105498100\.770\.770\.810\.940\.320\.970\.770\.920\.140\.150\.271\.000\.350\.790\.850\.67slots\_\{\\mathrm\{lot\}\}0\.51010944109100\.730\.760\.790\.930\.330\.980\.830\.920\.140\.150\.311\.000\.380\.770\.940\.71slots\_\{\\mathrm\{lot\}\}0\.759101024109100\.710\.750\.770\.930\.350\.990\.820\.890\.140\.150\.321\.000\.380\.720\.940\.72slots\_\{\\mathrm\{lot\}\}0\.979934109100\.670\.710\.760\.920\.350\.990\.790\.890\.140\.150\.251\.000\.390\.870\.940\.74sdrifts\_\{\\mathrm\{drift\}\}0\.1101096398100\.780\.780\.870\.940\.290\.950\.770\.930\.140\.140\.251\.000\.340\.720\.850\.73sdrifts\_\{\\mathrm\{drift\}\}0\.38996310890\.710\.760\.860\.920\.340\.910\.770\.900\.140\.140\.261\.000\.370\.710\.850\.69slots\_\{\\mathrm\{lot\}\}0\.5,sdrifts\_\{\\mathrm\{drift\}\}0\.289944108100\.700\.740\.820\.920\.370\.940\.750\.920\.140\.150\.251\.000\.390\.760\.850\.7550 batches997539760\.730\.730\.660\.920\.330\.860\.630\.760\.150\.290\.411\.000\.390\.950\.720\.8040 variables999459890\.610\.660\.730\.960\.270\.910\.620\.760\.070\.070\.281\.000\.360\.940\.800\.84with the combined difficulties050014410\.260\.570\.240\.580\.180\.500\.460\.340\.090\.280\.23–0\.070\.670\.500\.33Sensor errorserrs\_\{\\mathrm\{err\}\}\(7\)sensor errorserrs\_\{\\mathrm\{err\}\}0\.11010943108100\.780\.780\.830\.930\.290\.970\.770\.870\.140\.150\.311\.000\.330\.640\.940\.67sensor errorserrs\_\{\\mathrm\{err\}\}0\.210991310790\.770\.760\.790\.930\.280\.950\.740\.900\.140\.150\.291\.000\.330\.620\.940\.64sensor errorserrs\_\{\\mathrm\{err\}\}0\.310910029690\.770\.760\.800\.920\.230\.910\.670\.860\.140\.150\.37–0\.260\.530\.810\.5650 batches889119860\.740\.670\.770\.930\.280\.880\.650\.700\.170\.380\.411\.000\.320\.751\.000\.6540 variables998129890\.640\.630\.620\.930\.190\.810\.550\.750\.070\.170\.251\.000\.210\.440\.890\.69with the combined difficulties040013510\.260\.600\.170\.780\.170\.450\.520\.440\.140\.250\.11–0\.070\.330\.710\.25sensor errorserrs\_\{\\mathrm\{err\}\}0\.510990293100\.760\.740\.770\.840\.200\.880\.710\.860\.150\.170\.34–0\.210\.470\.750\.57Product gradescc\(7\)gradescc0\.251010104310890\.780\.770\.870\.920\.290\.970\.770\.860\.140\.150\.291\.000\.330\.700\.850\.76gradescc0\.58991491090\.740\.740\.720\.910\.300\.900\.800\.850\.150\.160\.351\.000\.330\.621\.000\.73gradescc0\.757891410870\.680\.690\.720\.910\.280\.920\.720\.740\.150\.170\.381\.000\.310\.621\.000\.6150 batches877237750\.710\.640\.670\.840\.330\.800\.550\.570\.170\.400\.351\.000\.360\.670\.930\.7340 variables599038970\.560\.610\.600\.930\.200\.780\.690\.650\.080\.100\.29–0\.310\.570\.900\.82with the combined difficulties050023420\.240\.560\.200\.640\.190\.480\.420\.390\.120\.320\.12–0\.080\.620\.390\.38gradescc1\.06890410870\.630\.680\.780\.900\.270\.930\.750\.720\.160\.160\.46–0\.300\.631\.000\.80Variable delaysjj\(7\)delaysjj0\.2510101073109100\.790\.840\.900\.940\.300\.970\.820\.960\.140\.140\.331\.000\.350\.730\.850\.68delaysjj0\.510101083109100\.790\.870\.870\.940\.280\.980\.800\.950\.140\.140\.281\.000\.320\.740\.850\.75delaysjj0\.7510101084109100\.790\.880\.870\.940\.280\.980\.810\.940\.140\.140\.291\.000\.330\.770\.850\.7650 batches910109310780\.800\.890\.840\.920\.350\.980\.570\.860\.170\.300\.381\.000\.420\.790\.880\.8340 variables810972108100\.640\.810\.780\.950\.150\.970\.570\.900\.070\.080\.301\.000\.230\.780\.800\.78with the combined difficulties051126510\.240\.530\.250\.470\.180\.580\.480\.340\.100\.380\.171\.000\.070\.520\.450\.12delaysjj1\.01010994109100\.780\.880\.830\.940\.260\.990\.750\.920\.140\.150\.371\.000\.300\.650\.850\.84Cumulative effectskavgk\_\{\\mathrm\{avg\}\}\(7\)cumulativekavgk\_\{\\mathrm\{avg\}\}0\.2581094310790\.820\.790\.810\.920\.330\.940\.670\.930\.140\.160\.251\.000\.350\.930\.880\.78cumulativekavgk\_\{\\mathrm\{avg\}\}0\.581095310690\.820\.820\.780\.920\.320\.940\.650\.930\.140\.160\.231\.000\.350\.931\.000\.80cumulativekavgk\_\{\\mathrm\{avg\}\}0\.759109549490\.840\.880\.850\.960\.320\.900\.570\.890\.140\.160\.311\.000\.350\.731\.000\.8450 batches899529560\.810\.750\.760\.920\.320\.920\.560\.700\.180\.280\.341\.000\.340\.861\.000\.7940 variables71064310690\.620\.720\.690\.960\.190\.890\.490\.770\.070\.100\.291\.000\.240\.730\.860\.88with the combined difficulties070026310\.290\.640\.220\.710\.180\.580\.410\.390\.170\.480\.20–0\.090\.600\.311\.00cumulativekavgk\_\{\\mathrm\{avg\}\}1\.09101094105100\.820\.900\.910\.940\.290\.990\.660\.890\.140\.150\.331\.000\.330\.821\.000\.82Operator reactionssops\_\{\\mathrm\{op\}\}\(7\)operatorsops\_\{\\mathrm\{op\}\}0\.1910106399100\.810\.800\.820\.940\.320\.990\.870\.940\.140\.140\.271\.000\.370\.790\.850\.80operatorsops\_\{\\mathrm\{op\}\}0\.299963109100\.840\.780\.830\.940\.340\.970\.870\.940\.140\.140\.281\.000\.380\.750\.850\.79operatorsops\_\{\\mathrm\{op\}\}0\.3101096399100\.870\.800\.800\.940\.310\.980\.860\.960\.140\.140\.311\.000\.360\.750\.850\.7750 batches1089829980\.850\.700\.880\.930\.390\.980\.780\.790\.160\.230\.461\.000\.420\.890\.940\.7740 variables8101040101090\.600\.680\.780\.960\.230\.920\.740\.820\.070\.080\.271\.000\.300\.711\.000\.77with the combined difficulties040014610\.300\.570\.270\.590\.180\.540\.480\.430\.200\.390\.17–0\.160\.470\.560\.25operatorsops\_\{\\mathrm\{op\}\}0\.5101096389100\.860\.790\.800\.940\.310\.970\.850\.940\.140\.140\.291\.000\.360\.750\.900\.68All five imperfections\(3\)all five imperfections799127070\.700\.810\.830\.840\.250\.770\.580\.810\.170\.220\.521\.000\.260\.520\.000\.8540 variables6107019050\.530\.680\.620\.800\.140\.660\.490\.630\.090\.210\.45–0\.160\.50–0\.68with the combined difficulties050015320\.320\.520\.210\.660\.200\.590\.480\.340\.260\.510\.14–0\.230\.510\.500\.50Edge types \(ablation\)\(2\)time\-step links only10101062108100\.770\.790\.870\.940\.310\.970\.790\.940\.140\.150\.331\.000\.350\.780\.800\.72batch\-level links only10101052107100\.750\.790\.930\.980\.351\.000\.620\.910\.160\.170\.401\.000\.390\.770\.780\.86
#### Sensitivity to the selection threshold\.

Every method selects a list from its ranking with a fixed rule: correlation and Granger causality keep the candidates whose Benjamini–Hochberg adjustedpp\-value is below 0\.05, and DYNOTEARS those with a contemporaneous weight of at leastτW=0\.3\\tau\_\{W\}=0\.3or a lagged weight of at leastτA=0\.1\\tau\_\{A\}=0\.1, the thresholds of its paper\([Pamfil et al\., 2020](https://arxiv.org/html/2611.17556#bib.bib34)\)\. Table[10](https://arxiv.org/html/2611.17556#A3.T10)gives the results at other thresholds; the ranking metrics do not depend on the threshold\. The long lists of correlation and Granger causality come from this threshold: keeping only their first three candidates raises their precision from 0\.15 and 0\.21 to 0\.42 and 0\.49\. DYNOTEARS selects nothing on 71% of the plants at its paper’s thresholds; atτW=0\.03\\tau\_\{W\}=0\.03,τA=0\.01\\tau\_\{A\}=0\.01it selects nothing on 8%, its first pick is right on 85% of the plants and its precision is 0\.73 \(PCMCI\+ at 0\.05: 78% and 0\.69\)\. For PCMCI\+ we report 0\.05, the level of our other tests; the experiments of its paper use 0\.01\. A stricter level shortens its list and raises its precision, from 0\.69 at 0\.05 to 0\.89 at 0\.001, but leaves it empty on 30% of the plants instead of 10%, with fewer first picks right \(66% instead of 78%\)\.

Table 10:Selected lists under other thresholds, over all 1,190 plants \(columns as in Table[3](https://arxiv.org/html/2611.17556#S6.T3)\)\. BH: Benjamini–Hochberg adjustedpp\-values;*first 3 by score*: the three highest\-ranked candidates\.

## Appendix DDifficulty measured on the data

Each configuration sets its difficulty through the shares\. Figure[5](https://arxiv.org/html/2611.17556#A4.F5)measures the resulting difficulty on the generated data with two quantities per plant, computed at the resolution and look\-back the methods use: the test strength of the direct causes and the share of wrong candidates that are look\-alikes\.

![Refer to caption](https://arxiv.org/html/2611.17556v1/fig_map.png)Figure 5:Difficulty measured on the data\. Horizontal: test strength of the median direct cause \(logarithmic scale\)\. Vertical: share of wrong candidates that are look\-alikes\. \(a\) Our configurations and three public process datasets\. \(b\) Share of plants whose first pick is a direct cause, averaged over the eight methods; the public datasets are not evaluated \(hollow\)\.*Test strength\.*For each direct cause, its partialR2R^\{2\}given the past ofYYand the other direct causes, from out\-of\-sample fits across batches, as the effect sizef2=R2/\(1−R2\)f^\{2\}=R^\{2\}/\(1\-R^\{2\}\), times the number of observations: the expected statistic of a test for that cause\. The figure shows the median over a plant’s direct causes, so one faint cause does not set the value\.

*Look\-alikes\.*A wrong candidate is a look\-alike if its past carries more information aboutYYthan the past of the median direct cause, measured bykk\-nearest\-neighbour mutual information\. The definition uses only the direct causes, so it applies to datasets whose other roles are unknown; in our plants, look\-alikes are siblings, ancestors and descendants\. The vertical axis is the share of wrong candidates that are look\-alikes\.

Each configuration is summarised over its 10 plants\. The public problems are Tennessee Eastman\([Downs and Vogel, 1993](https://arxiv.org/html/2611.17556#bib.bib18)\), with the fault runs as episodes and every variable with a single parent in the CIPCaD graph\([Menegozzo et al\., 2022](https://arxiv.org/html/2611.17556#bib.bib6)\)as a target; IndPenSim\([Goldrick et al\., 2019](https://arxiv.org/html/2611.17556#bib.bib21)\), with the production rate as target and one problem per control regime; and the food\-processing line of CIPCaD\-Bench, with the quality variable as target and the last stage’s variables as its direct causes\.

The combined configurations lie at lower test strength and with more look\-alikes than those with one difficulty, and the methods’ mean accuracy falls in the same direction \(Figure[5](https://arxiv.org/html/2611.17556#A4.F5)b\)\. The public datasets lie within the range of test strength of our configurations, with more look\-alikes than most of them\. Both quantities summarise difficulty for the target only\. A component aimed at one assumption of a method, such as downstream measurements for methods that ignore time order, can change the results of those methods with little change in either quantity; the results are therefore reported per method\.

## Appendix EMethod settings and reproducibility

Lagged correlation ranks the candidates by their correlation withYYover lags; Granger causality tests whether a candidate’s past improves the prediction ofYYbeyond its own past\. PC tests conditional independence between values at the same time step; PCMCI\+ tests lagged and contemporaneous links conditional on the estimated parents of both variables\. DYNOTEARS fits a structural vector autoregression under a continuous acyclicity constraint; VARLiNGAM orients effects within a time step through non\-Gaussian noise; SpaceTime scores graphs by minimum description length with every batch as a context\. DoubleML estimates the effect of a candidate’s past onYYwith flexible models of all other candidates and selects the candidates whose effect is significant\. DYNOTEARS uses the edge thresholds of its paper \(0\.3 for contemporaneous and 0\.1 for lagged edges\)\.

The common look\-back serves as Granger’s maximum lag, PCMCI\+’sτmax\\tau\_\{\\max\}, VARLiNGAM’s lag order and the window of lagged correlation, whose score is the largest absolute correlation over lags00toLL, tested with Fisher’szz\-transform onn−Ln\-Leffective samples\. PCMCI\+ uses partial correlation, at most 5 conditioning variables and only links into the target\. PC runs on the series at the evaluation resolution, without lags, at a significance level of 0\.05 and in its original variant with Fisher’szztest; a candidate is selected when PC links it toYYin either direction or leaves the edge undirected\. DYNOTEARS runs on standardised series, with its lag window set to the common look\-back, regularisationλW=λA=0\.05\\lambda\_\{W\}=\\lambda\_\{A\}=0\.05and at most 100 iterations\. VARLiNGAM runs on the batches concatenated in order; its list is the candidates that keep a non\-zero coefficient intoYYafter the adaptive\-lasso pruning of thelingamlibrary, and its ranking the largest absolute coefficient over lags\.

DoubleML fits one partially linear model per candidate with the partialling\-out score\. The data form a table: each row is one step of a batch at the evaluation resolution, from stepL\+1L\+1on, and lags never cross a batch \(2,400 rows at the reference\)\. The outcome isY⁡\(t\)Y\(t\)\. The treatments are the candidate at lags1,…,L1,\\dots,L, and each lag’s coefficient is estimated with the other lags among the controls\. The controls areYYat lags1,…,L1,\\dots,Land every other recorded variable at lags0,…,L0,\\dots,L; the lag\-0 values are included because one averaged step contains much of a cause’s effect\. The nuisance models are two LightGBM regressors\([Ke et al\., 2017](https://arxiv.org/html/2611.17556#bib.bib43)\)with fixed parameters: 300 trees, learning rate 0\.05, 31 leaves, at least 20 samples per leaf, and row and column subsampling of 0\.8\. Cross\-fitting uses three folds and one repetition\. Folds contain whole batches and standard errors are clustered by batch, which is how autocorrelation within a batch is handled\. Thepp\-values are adjusted with Holm over a candidate’s lags, the smallest is kept, and Benjamini–Hochberg at 0\.05 runs across variables\. The ranking is the largest\|t\|\|t\|over lags\.

SpaceTime scores graphs with Gaussian\-process mechanisms\. Every batch is a separate context, in which the mechanisms may differ, and each batch has a single regime, with no change of mechanism over time\. The search is restricted to the parents ofYY, as the task asks, andYYis constrained to have no children, so that every edge found atYYpoints into it\. Its list is the parents it names at its default acceptance level\. SpaceTime returns a set rather than a ranking, so we rank the parents it names first and the other candidates by their pairwise gain, the improvement in description length when each alone is a parent ofYY, which its search computes in its first step\.

The sortability checks rank the candidates in ascending order of their variance or of theR2R^\{2\}of a linear regression on the other candidates, so the least predictable candidate comes first\. All random seeds are fixed\. Every method sees the same table: the block means of every batch at the evaluation resolution\. Lags are constructed within batches, except for VARLiNGAM, which uses the concatenated batches\.

相似文章

基于因果效应约束的可解释因果发现

arXiv cs.LG

本文提出了一种用于条件因果发现的贝叶斯方法,其中因果图和参数的后验以用户指定的因果效应约束(例如,较大的因果效应)为条件。他们采用稀有事件估计技术来处理后验质量较小的事件,并在合成数据和Sachs蛋白质数据集上验证了该方法。

CausaLab: 面向AI科学家的可扩展交互式因果发现环境

Hugging Face Daily Papers

CausaLab 是一个可扩展的环境,用于评估LLM智能体在交互式因果发现中的表现,同时衡量预测准确性和对潜在因果机制的忠实复现。实验揭示了预测与机制复现之间的差距,突显了当前LLM智能体作为实验性因果推理者的局限性。