Physics-informed neural networks for spacecraft trajectory optimization with exact boundary conditions

arXiv cs.LG Papers

Summary

该论文提出一种简单的重参数化变换,将边界条件直接嵌入物理信息神经网络(PINN)的解空间,从而把航天器轨迹优化中的约束问题转化为无约束的单目标优化,显著改善收敛性与求解精度。

arXiv:2611.21933v1 Announce Type: new Abstract: We use neural networks to optimize trajectories by modeling the position as a function of time and minimizing the thrust with respect to the network parameters. This corresponds to an unsupervised training procedure in which physical laws are incorporated directly into the optimization objective. In standard physics-informed neural networks, initial and terminal states are enforced through additional weighted loss terms, resulting in a multi-objective optimization problem that is highly sensitive to the choice of loss weights and can lead to poor convergence and reduced accuracy. To address this challenge, we construct a simple transformation, exploiting the mathematical structure of trajectory optimization problems, to enforce boundary conditions exactly a-priori. This transformation can be interpreted as a reparameterization of the solution space that embeds the boundary conditions directly into the function. This converts the original constrained optimization problem into an unconstrained one over admissible trajectories. I.e. the transformation eliminates the boundary condition penalty terms and reduces the problem to a single-objective optimization. Crucially, trajectory optimization comes with a low-dimensional boundary, which allows particularly simple and stable constructions of distance functions compared to general domains in partial differential equations. Numerical experiments demonstrate that the proposed approach meaningfully improves convergence behavior and solution accuracy.
Original Article
View Cached Full Text

Cached at: 11/19/26, 03:35 AM

# Physics-informed neural networks for spacecraft trajectory optimization with exact boundary conditions
Source: [https://arxiv.org/html/2611.21933](https://arxiv.org/html/2611.21933)
Sascha RanftlNote:Fellow, Division of Applied Mathematics, 182 George Street\.Note:Assistant Professor, School of Mechanical Engineering, 585 Purdue Mall\.Note:Corresponding author: sranftl@purdue\.eduAffiliation:Brown University, Providence, Rhode Island, 02912Affiliation:Purdue University, West Lafayette, Indiana, 47907

###### Abstract

We use neural networks to optimize trajectories by modeling the position as a function of time and minimizing the thrust with respect to the network parameters\. This corresponds to an unsupervised training procedure in which physical laws are incorporated directly into the optimization objective\. In standard physics\-informed neural networks, initial and terminal states are enforced through additional weighted loss terms, resulting in a multi\-objective optimization problem that is highly sensitive to the choice of loss weights and can lead to poor convergence and reduced accuracy\. To address this challenge, we construct a simple transformation, exploiting the mathematical structure of trajectory optimization problems, to enforce boundary conditions exactly a\-priori\. This transformation can be interpreted as a reparameterization of the solution space that embeds the boundary conditions directly into the function\. This converts the original constrained optimization problem into an unconstrained one over admissible trajectories\. I\.e\. the transformation eliminates the boundary condition penalty terms and reduces the problem to a single\-objective optimization\. Crucially, trajectory optimization comes with a low\-dimensional boundary, which allows particularly simple and stable constructions of distance functions compared to general domains in partial differential equations\. Numerical experiments demonstrate that the proposed approach meaningfully improves convergence behavior and solution accuracy\.

## Nomenclature

𝒩𝜽\\mathcal\{N\}\_\{\\boldsymbol\{\\theta\}\}=neural network𝜽\\boldsymbol\{\\theta\}=neural\-network parameters𝐫\\mathbf\{r\},𝐯\\mathbf\{v\}=spacecraft position vector and velocity vector𝐫𝜽\\mathbf\{r\}\_\{\\boldsymbol\{\\theta\}\}=neural\-network trajectoryρ𝜽\\rho\_\{\\boldsymbol\{\\theta\}\}=transformed neural network trajectory satisfying exact boundary conditionsTT=time of flight \(TOF\), terminal timett=physical timet0t\_\{0\}=initial timeτ\\tau=normalized time,t/Tt/T𝐅\\mathbf\{F\}=thrust vectorRR,VV=interpolation term satisfying position and velocity boundary conditions, respectivelyϕ⁡\(τ\)\\phi\(\\tau\)=distance function for velocity constraintsψ⁡\(τ\)\\psi\(\\tau\)=distance function for position constraintsΦ\\Phi=gravitational potentialmsm\_\{s\}=spacecraft massMqM\_\{q\}=mass of gravitating bodyqq𝐏q\\mathbf\{P\}\_\{q\}=position of gravitating bodyqqNMN\_\{M\}=number of gravitating bodiesNN=number of collocation points; with subscripts BC, IC, “data", or “physics"𝒟\\mathcal\{D\}=differential operatoruu=solution of governing differential equationΩ\\Omega=computational domain∂Ω\\partial\\Omega=boundary of computational domainℒ\\mathcal\{L\}=loss functionSubscripts:BC =boundary conditionIC =intial conditionopt =optimal value00=initial valueTT=terminal value
## 1Introduction

Machine learning has gained recent interest as a tool in astronautics and aerospace engineering \(AAE\), e\.g\. for data\-driven modeling, control\[[49](https://arxiv.org/html/2611.21933#bib.bib40),[64](https://arxiv.org/html/2611.21933#bib.bib41)\], or design of complex systems\[[9](https://arxiv.org/html/2611.21933#bib.bib29)\]\. Applications range from aerodynamic shape optimization\[[58](https://arxiv.org/html/2611.21933#bib.bib28)\]and turbulence modeling\[[45](https://arxiv.org/html/2611.21933#bib.bib32)\]to design\[[36](https://arxiv.org/html/2611.21933#bib.bib30)\]and satellite fault diagnosis\[[31](https://arxiv.org/html/2611.21933#bib.bib31)\]\. Data\-driven surrogates have been particularly effective in reducing the computational burden of high\-fidelity numerical simulations, e\.g\. for model optimization, parameter estimation or uncertainty quantification\[[47](https://arxiv.org/html/2611.21933#bib.bib18)\]\.

Machine learning has also gained interest for trajectory optimization specifically\. Already Ampatzis and Izzo\[[2](https://arxiv.org/html/2611.21933#bib.bib33)\]demonstrated the use of machine learning for surrogate modeling of objective functions in interplanetary trajectory design\. Deep learning and/or reinforcement learning has gained particular interest recently and has been used for interplanetary missions under uncertainty\[[67](https://arxiv.org/html/2611.21933#bib.bib34)\], low\-thrust orbital transfer\[[57](https://arxiv.org/html/2611.21933#bib.bib64),[8](https://arxiv.org/html/2611.21933#bib.bib65)\], intercept guidance\[[23](https://arxiv.org/html/2611.21933#bib.bib66)\], planetary landing\[[24](https://arxiv.org/html/2611.21933#bib.bib67)\], uncontrolled re\-entry prediction\[[51](https://arxiv.org/html/2611.21933#bib.bib62)\]and autonomous spacecraft guidance during proximity operations\[[21](https://arxiv.org/html/2611.21933#bib.bib35)\]\. Beyond deep or reinforcement learning, many other models and algorithms have been proposed, e\.g\. hybrid models for collision avoidance\[[27](https://arxiv.org/html/2611.21933#bib.bib63)\]or Gaussian mixture models \(GMMs\)\[[22](https://arxiv.org/html/2611.21933#bib.bib68)\]to equip the trajectory optimization with a notion of uncertainty\. Data\-driven optimization has also expanded to the aviation domain, where machine learning techniques assist in designing fuel\-efficient aircraft trajectories subject to operational constraints\[[65](https://arxiv.org/html/2611.21933#bib.bib36)\]\. In proximity operations, learning\-based trajectory planning has been used through deep neural networks\[[5](https://arxiv.org/html/2611.21933#bib.bib37),[66](https://arxiv.org/html/2611.21933#bib.bib38)\], incl\. e\.g\. transformers for rendezvous scenarios\[[28](https://arxiv.org/html/2611.21933#bib.bib39)\]\.

In parallel, physics\-informed neural networks \(PINNs\) have emerged as an interesting paradigm for solving forward and inverse problems governed by differential equations \(DEs\)\[[35](https://arxiv.org/html/2611.21933#bib.bib43),[15](https://arxiv.org/html/2611.21933#bib.bib42),[46](https://arxiv.org/html/2611.21933#bib.bib1)\]\. PINNs embed the governing laws of physics directly into the training objective of a neural network by penalizing DE residuals alongside data mismatch losses, and have since been extended into broader frameworks of physics\-informed machine learning\[[34](https://arxiv.org/html/2611.21933#bib.bib2)\]\. Applications range from scientific computing\[[12](https://arxiv.org/html/2611.21933#bib.bib23),[20](https://arxiv.org/html/2611.21933#bib.bib22)\]and fluid mechanics\[[10](https://arxiv.org/html/2611.21933#bib.bib21)\]to optimization tasks in engineering systems, and, in particular, also includes PINNs for trajectory optimization\[[53](https://arxiv.org/html/2611.21933#bib.bib5)\]\.

Despite their promise, PINNs for trajectory optimization face a fundamental challenge: the need to balance multiple competing objectives\. In standard formulations, the governing equations, boundary conditions, and possibly data are enforced via weighted loss terms\. In particular, boundary conditions are typically imposed through penalty terms, leading to a multi\-objective optimization problem whose performance is highly sensitive to the choice of loss weights\. This can result in convergence issues, reduced accuracy, and significant tuning effort\.

In this work, we address this challenge by proposing a simple transformation\- or reparametrization principle of the PINN based on distance functions\[[56](https://arxiv.org/html/2611.21933#bib.bib6)\]that enforces boundary conditions exactly*a priori*, i\.e\. before training\. For this, we identify and exploit a structural property of trajectory optimization problems that allows boundary conditions to be embedded directly into the function class, thereby converting a multi\-objective learning problem into a better\-conditioned single\-objective optimization problem\. This eliminates the need for boundary condition penalty terms altogether\. In contrast to more general settings, we exploit the typically relatively simple geometric structure of boundary conditions in trajectory optimization specifically\. We demonstrate that the imposition of exact boundary conditions via this transformation can improve both convergence behavior and solution accuracy for trajectory optimization problems\.

## 2Related Work and Contribution

### 2\.1PINNs for trajectory optimization

Already\[[53](https://arxiv.org/html/2611.21933#bib.bib5)\]applied PINNs, in their basic form, to trajectory optimization, and found that PINNs compare favorably to reinforcement learning on the same task\. Since then, a number of studies utilized PINNs for trajectory optimization specifically, incl\. in an aerospace engineering context\. Beyond AAE, applications span intelligent vehicle tracking control\[[11](https://arxiv.org/html/2611.21933#bib.bib51)\], online trajectory regeneration for multirotors\[[25](https://arxiv.org/html/2611.21933#bib.bib52)\], and vessel trajectory prediction\[[1](https://arxiv.org/html/2611.21933#bib.bib56)\]\. These works demonstrate that PINNs can learn nonlinear kinematic and dynamic constraints directly from governing equations, but they also inherit the challenges of multi\-objective formulations, where the physics\-informed residuals must be balanced with trajectory boundary conditions, path constraints or possibly data\.

In light of this trade\-off, recent works such as AW\-EL\-PINNs\[[38](https://arxiv.org/html/2611.21933#bib.bib59)\]and proportional\-derivative PINNs\[[25](https://arxiv.org/html/2611.21933#bib.bib52)\]develop structured approaches to constraint enforcement by tailoring the PINN architecture or loss formulation to specific control\-theoretic structures\.

### 2\.2Constraint enforcement in PINNs

Fundamental challenges of PINNs include stiffness and gradient pathologies\[[62](https://arxiv.org/html/2611.21933#bib.bib15)\], convergence sensitivity to loss scaling\[[60](https://arxiv.org/html/2611.21933#bib.bib17)\], and difficulties with training stability in high\-dimensional or complex domains\[[63](https://arxiv.org/html/2611.21933#bib.bib16),[61](https://arxiv.org/html/2611.21933#bib.bib19)\]\. Various approaches have been proposed to mitigate these issues, including improved architectures, training strategies, domain decomposition, Gaussian process\-based formulations\[[13](https://arxiv.org/html/2611.21933#bib.bib3),[33](https://arxiv.org/html/2611.21933#bib.bib4)\], and adaptive loss balancing\[[6](https://arxiv.org/html/2611.21933#bib.bib20)\]\.

A particular class of methods focuses on the exact imposition of constraints\. Standard PINNs typically impose constraints approximately via penalty terms, which can lead to inaccurate constraint satisfaction, especially in stiff or ill\-conditioned problems\[[63](https://arxiv.org/html/2611.21933#bib.bib16)\]\. This has motivated approaches that enforce constraints exactly by construction, e\.g\. via stochastic process formulations\[[48](https://arxiv.org/html/2611.21933#bib.bib11)\]or by constraining underlying fields or potentials\[[59](https://arxiv.org/html/2611.21933#bib.bib26),[43](https://arxiv.org/html/2611.21933#bib.bib27)\]\.

Several works specifically address boundary conditions: distance function approaches\[[56](https://arxiv.org/html/2611.21933#bib.bib6)\], unified hard\-constraint frameworks\[[40](https://arxiv.org/html/2611.21933#bib.bib7)\], penalty\-free neural methods\[[54](https://arxiv.org/html/2611.21933#bib.bib8)\], Fourier feature embeddings for Neumann conditions\[[55](https://arxiv.org/html/2611.21933#bib.bib13)\], and decoupled formulations\[[42](https://arxiv.org/html/2611.21933#bib.bib12)\]\. Beyond neural networks, constrained Gaussian processes have been used to encode boundary conditions exactly in linear DEs\[[13](https://arxiv.org/html/2611.21933#bib.bib3),[33](https://arxiv.org/html/2611.21933#bib.bib4),[37](https://arxiv.org/html/2611.21933#bib.bib9),[30](https://arxiv.org/html/2611.21933#bib.bib10)\]\. These approaches demonstrate that exact constraint imposition can improve accuracy, convergence, and robustness by avoiding the trade\-offs inherent in penalty\-based formulations\[[50](https://arxiv.org/html/2611.21933#bib.bib24),[62](https://arxiv.org/html/2611.21933#bib.bib15),[29](https://arxiv.org/html/2611.21933#bib.bib25)\]\.

### 2\.3PINNs in astronautics and aerospace applications

In trajectory optimization within AAE, PINNs have been applied to orbital transfers, rendezvous guidance, and ascent trajectory design\. Schiassi et al\.\[[52](https://arxiv.org/html/2611.21933#bib.bib47)\]demonstrated PINNs for optimal planar orbit transfers, while subsequent work extended to asteroid proximity maneuvers\[[19](https://arxiv.org/html/2611.21933#bib.bib50)\], low\-thrust rendezvous of CubeSats\[[14](https://arxiv.org/html/2611.21933#bib.bib45)\], and orbital pursuit\-evasion problems\[[68](https://arxiv.org/html/2611.21933#bib.bib46)\]\. Physics\-Informed Pontryagin Neural Networks \(PIPNNs\) incorporate optimality conditions directly\[[18](https://arxiv.org/html/2611.21933#bib.bib44),[17](https://arxiv.org/html/2611.21933#bib.bib48),[7](https://arxiv.org/html/2611.21933#bib.bib54)\], while hybrid methods combine PINNs with sampling\-based planning\[[16](https://arxiv.org/html/2611.21933#bib.bib53)\]\.

Several works also aim to improve constraint handling, including unsupervised PINNs for low\-thrust transfers\[[26](https://arxiv.org/html/2611.21933#bib.bib49)\], multi\-objective planning approaches\[[69](https://arxiv.org/html/2611.21933#bib.bib55)\], and applications in flight dynamics and uncertainty quantification\[[44](https://arxiv.org/html/2611.21933#bib.bib58)\]\. A limited number of works enforce constraints structurally, e\.g\. via Pontryagin\-based formulations\[[17](https://arxiv.org/html/2611.21933#bib.bib48)\], trajectory\-embedded transformations\[[68](https://arxiv.org/html/2611.21933#bib.bib46)\], or constraint\-regularized transfer learning\[[4](https://arxiv.org/html/2611.21933#bib.bib57)\]\. PINNs notably have also been used successfully for orbit determination\[[3](https://arxiv.org/html/2611.21933#bib.bib61)\]\.

### 2\.4Contribution

These works confirm that boundary condition enforcement remains a central challenge in PINNs for trajectory optimization\. Notably, among the existing literature, only\[[18](https://arxiv.org/html/2611.21933#bib.bib44),[17](https://arxiv.org/html/2611.21933#bib.bib48)\]impose certain constraints exactly using the Theory of Functional Connections, while most approaches rely on soft constraint formulations\. In contrast to the Theory of Functional Connections, which constructs problem\-specific basis functions with potentially high implementation complexity, our approach operates directly on standard neural network parametrizations via simple multiplicative distance functions\. In contrast,\[[4](https://arxiv.org/html/2611.21933#bib.bib57)\]argue that exact boundary conditions are often not available in certain applications\.

In this work, we make the following contributions:

- •Methodological contribution:We propose a simple and systematic method for embedding exact kinematic constraints into PINNs for trajectory optimization*a priori*, i\.e\. before training, via a linear transformation of the network\. In contrast to\[[68](https://arxiv.org/html/2611.21933#bib.bib46),[4](https://arxiv.org/html/2611.21933#bib.bib57),[18](https://arxiv.org/html/2611.21933#bib.bib44),[14](https://arxiv.org/html/2611.21933#bib.bib45)\], this approach avoids the need for and limitations of specialized frameworks such as the Theory of Functional Connections or Extreme Learning Machines\.
- •Conceptual contribution:The proposed transformation, based on distance functions, enforces initial and terminal conditions exactly by construction and eliminates the boundary condition loss term, thereby reducing the original multi\-objective problem to a single\-objective optimization, which is generally easier to solve\. The proposed construction applies to trajectory optimization problems with kinematic constraints and extends naturally to alternative coordinate systems and conceptually to higher\-order constraints, and phase\-space formulations\.
- •Computational implications:By removing the need for balancing loss terms, the method removes sensitivity to loss weights associated with boundary conditions, which is a major source of instability in standard PINNs, and simplifies the training process\.
- •Empirical validation:Through a series of trajectory optimization problems of increasing complexity, including multi\-body swing\-by and low\-thrust orbit transfer, we demonstrate improved convergence behavior, robustness, and solution quality compared to standard PINNs and competitive performance relative to a baseline defined by the classical numerical solverOpenGoddard\[[32](https://arxiv.org/html/2611.21933#bib.bib60)\]\. The code of our PINN\-implementation with exact boundary conditions is available at:[https://github\.com/s\-schoenherr/SpacePINN](https://github.com/s-schoenherr/SpacePINN)

## 3Background

We first introduce basic PINNs in Sec\.[3\.1](https://arxiv.org/html/2611.21933#S3.SS1)and then derive the appropriate loss terms for trajectory optimization problems in Sec\.[3\.2](https://arxiv.org/html/2611.21933#S3.SS2)\. Based on this, we will later introduce the transformation of the PINN to eliminate the boundary condition\-term \(BC\) a priori in Sec\.[4](https://arxiv.org/html/2611.21933#S4)\.

### 3\.1Physics\-informed neural networks

The central idea of physics\-informed neural networks \(PINNs\) is to embed prior knowledge of the underlying physical process into a machine learning model\[[34](https://arxiv.org/html/2611.21933#bib.bib2),[46](https://arxiv.org/html/2611.21933#bib.bib1)\]\. This prior knowledge is typically expressed in the form of an ordinary or partial differential equation \(PDE\) along with corresponding boundary and initial conditions\.

LetΩ⊂ℝd\\Omega\\subset\\mathbb\{R\}^\{d\}denote a spatial domain with boundary∂Ω\\partial\\Omega, with timettwhereT\>0T\>0denotes the terminal time\. Formally, we seek a solutionuuin a suitable function space𝒱\\mathcal\{V\}of sufficiently smooth functions satisfying the governing equation and boundary conditions\. The governing equation is a differential equation of the form

𝒟⁡\[u⁡\(𝐱,t\)\]=f⁡\(𝐱,t\),\(𝐱,t\)∈Ω×\(0,T\],\\displaystyle\\mathcal\{D\}\[u\(\\mathbf\{x\},t\)\]=f\(\\mathbf\{x\},t\),\\quad\(\\mathbf\{x\},t\)\\in\\Omega\\times\(0,T\],\(1\)where𝒟⁡\[⋅\]\\mathcal\{D\}\[\\cdot\]is a \(possibly nonlinear\) differential operator andffis an inhomogeneity\. This PDE is complemented by initial and boundary conditions

u⁡\(𝐱,0\)=u0​\(𝐱\),𝐱∈Ω,\\displaystyle u\(\\mathbf\{x\},0\)=u\_\{0\}\(\\mathbf\{x\}\),\\quad\\mathbf\{x\}\\in\\Omega,\(2\)u⁡\(𝐱,t\)=g⁡\(𝐱,t\),\(𝐱,t\)∈∂Ω×\(0,T\]\.\\displaystyle u\(\\mathbf\{x\},t\)=g\(\\mathbf\{x\},t\),\\quad\(\\mathbf\{x\},t\)\\in\\partial\\Omega\\times\(0,T\]\.\(3\)
For instance, depending on the regularity requirements of the differential operatorD⁡\[⋅\]D\[\\cdot\], one may consider function spaces𝒱\\mathcal\{V\}such asu∈Ck​\(Ω×\(0,T\]\)u\\in C^\{k\}\(\\Omega\\times\(0,T\]\), or an appropriate Sobolev spaceu∈Hk​\(Ω×\(0,T\]\)u\\in H^\{k\}\(\\Omega\\times\(0,T\]\)\.

A PINN approximates the solutionuuby a neural networkuθ​\(𝐱,t\)u\_\{\\mathbf\{\\theta\}\}\(\\mathbf\{x\},t\)that is parametrized by weights and biases summarized asθ\\mathbf\{\\theta\}\. The PINN ansatzuθu\_\{\\mathbf\{\\theta\}\}then defines a parametrized subset𝒱θ⊂𝒱\\mathcal\{V\}\_\{\\theta\}\\subset\\mathcal\{V\}\. By substitutinguθ​\(𝐱,t\)u\_\{\\mathbf\{\\theta\}\}\(\\mathbf\{x\},t\)into Eqs\. \(1\)\-\(3\), we obtain residuals each for the physical law, the initial and boundary conditions, respectively\. Then, training is performed by minimizing weighted sum of these residuals, evaluated at a selected number of collocation points\. Formally, the optimal weights and biasesθopt\\mathbf\{\\theta\}\_\{\\text\{opt\}\}are given by

θopt=arg⁡min⁡ℒ⁡\(θ\)\\displaystyle\\mathbf\{\\theta\}\_\{\\text\{opt\}\}=\\arg\\min\\mathcal\{L\}\(\\mathbf\{\\theta\}\)\(4\)whereℒ\\mathcal\{L\}is the so\-called loss function\. For example, the loss function can be defined as the mean squared error111Also other definitions and norms are possiblebased on theL2L\_\{2\}\-norm as

ℒ⁡\(θ\)\\displaystyle\\mathcal\{L\}\(\\mathbf\{\\theta\}\)=ℒdata\+λphysics​ℒphysics\+λBC​ℒBC\+λIC​ℒIC\\displaystyle=\\mathcal\{L\}\_\{\\text\{data\}\}\+\\lambda\_\{\\text\{physics\}\}\\mathcal\{L\}\_\{\\text\{physics\}\}\+\\lambda\_\{\\text\{BC\}\}\\mathcal\{L\}\_\{\\text\{BC\}\}\+\\lambda\_\{\\text\{IC\}\}\\mathcal\{L\}\_\{\\text\{IC\}\}=1Ndata​∑i=1Ndata\|uθ​\(𝐱i,ti\)−ui\|2\\displaystyle=\\frac\{1\}\{N\_\{\\text\{data\}\}\}\\sum\_\{i=1\}^\{N\_\{\\text\{data\}\}\}\\left\|u\_\{\\mathbf\{\\theta\}\}\(\\mathbf\{x\}\_\{i\},t\_\{i\}\)\-u\_\{i\}\\right\|^\{2\}\+λphysicsNphysics∑j=1Nphysics\|𝒟\[uθ\]\(𝐱j,tj\)−f\(𝐱j,tj\)\|2\\displaystyle\\quad\+\\frac\{\\lambda\_\{\\text\{physics\}\}\}\{N\_\{\\text\{physics\}\}\}\\sum\_\{j=1\}^\{N\_\{\\text\{physics\}\}\}\\left\|\\mathcal\{D\}\[u\_\{\\mathbf\{\\theta\}\}\]\(\\mathbf\{x\}\_\{j\},t\_\{j\}\)\-f\(\\mathbf\{x\}\_\{j\},t\_\{j\}\)\\right\|^\{2\}\+λBCNBC∑m=1NBC\|uθ\(𝐱m,tm\)−g\(𝐱m,tm\)\|2\\displaystyle\\quad\+\\frac\{\\lambda\_\{\\text\{BC\}\}\}\{N\_\{\\text\{BC\}\}\}\\sum\_\{m=1\}^\{N\_\{\\text\{BC\}\}\}\\left\|u\_\{\\mathbf\{\\theta\}\}\(\\mathbf\{x\}\_\{m\},t\_\{m\}\)\-g\(\\mathbf\{x\}\_\{m\},t\_\{m\}\)\\right\|^\{2\}\+λICNIC∑n=1NIC\|uθ\(𝐱n,0\)−u0\(𝐱n\)\|2\.\\displaystyle\\quad\+\\frac\{\\lambda\_\{\\text\{IC\}\}\}\{N\_\{\\text\{IC\}\}\}\\sum\_\{n=1\}^\{N\_\{\\text\{IC\}\}\}\\left\|u\_\{\\mathbf\{\\theta\}\}\(\\mathbf\{x\}\_\{n\},0\)\-u\_\{0\}\(\\mathbf\{x\}\_\{n\}\)\\right\|^\{2\}\.\(5\)withNdataN\_\{\\text\{data\}\},NphysicsN\_\{\\text\{physics\}\},NBCN\_\{\\text\{BC\}\}, andNICN\_\{\\text\{IC\}\}denote the number of data pointsui≡u⁡\(𝐱i,ti\)u\_\{i\}\\equiv u\(\\mathbf\{x\}\_\{i\},t\_\{i\}\)and collocation points𝐱k,tk\\mathbf\{x\}\_\{k\},t\_\{k\}with predictionsuθ​\(𝐱k,tk\)u\_\{\\mathbf\{\\theta\}\}\(\\mathbf\{x\}\_\{k\},t\_\{k\}\)for each componentk=\{j,m,n\}k=\\\{j,m,n\\\}, while the coefficientsλphysics\\lambda\_\{\\text\{physics\}\},λBC\\lambda\_\{\\text\{BC\}\}, andλIC\\lambda\_\{\\text\{IC\}\}are tunable parameters that control their relative contributions during training\. The data term \(first term\) enforces consistency with observed or simulated data, the PDE term \(second term\) penalizes violations of the governing equation, and the boundary condition \(BC\) and initial condition \(IC\) terms \(third and fourth terms\) enforce the boundary and initial conditions\.

By minimizingℒ⁡\(θ\)\\mathcal\{L\}\(\{\\mathbf\{\\theta\}\}\), the networkuθ​\(𝐱,t\)u\_\{\\mathbf\{\\theta\}\}\(\\mathbf\{x\},t\)learns to approximateu⁡\(𝐱,t\)u\(\\mathbf\{x\},t\)in a way that respects both the data and the physical constraints\. The collocation points for evaluating these residuals are typically composed as:

\{\(𝐱j,tj\)\}j=1Nphysicsin the domainΩ×\(0,T\]\\\{\(\\mathbf\{x\}\_\{j\},t\_\{j\}\)\\\}\_\{j=1\}^\{N\_\{\\text\{physics\}\}\}\\quad\\text\{in the domain\}\\quad\\Omega\\times\(0,T\]\(6a\)\{\(𝐱m,tm\)\}m=1NBCon the boundary∂Ω×\(0,T\]\\\{\(\\mathbf\{x\}\_\{m\},t\_\{m\}\)\\\}\_\{m=1\}^\{N\_\{\\text\{BC\}\}\}\\quad\\text\{on the boundary\}\\quad\\partial\\Omega\\times\(0,T\]\(6b\)\{\(𝐱n,tn\)\}n=1NICfor the initial condition\.\\\{\(\\mathbf\{x\}\_\{n\},t\_\{n\}\)\\\}\_\{n=1\}^\{N\_\{\\text\{IC\}\}\}\\quad\\text\{for the initial condition\}\.\(6c\)
In practice, PINNs impose the physics pointwise, i\.e\. only at the chosen collocation points\. Consequently, residuals are typically small near these points but may increase in regions without collocation point coverage\. Moreover, PINNs are sensitive to the values ofλphysics\\lambda\_\{\\text\{physics\}\},λBC\\lambda\_\{\\text\{BC\}\}, andλIC\\lambda\_\{\\text\{IC\}\}\. Careful selection of these weights is crucial for balanced training, and several adaptive strategies have been proposed\[[62](https://arxiv.org/html/2611.21933#bib.bib15),[60](https://arxiv.org/html/2611.21933#bib.bib17),[61](https://arxiv.org/html/2611.21933#bib.bib19),[6](https://arxiv.org/html/2611.21933#bib.bib20)\]\. Nevertheless, this remains an open research challenge and poses significant challenges to the application of PINNs for trajectory optimization\.

In the next section, Sec\.[3\.2](https://arxiv.org/html/2611.21933#S3.SS2), we will show the naive PINN\-formulation specifically for trajectory optimization problems\. We will show the numerical challenges with this plain setup in the experimental section Sec\.[5](https://arxiv.org/html/2611.21933#S5), and demonstrate how to fix the boundary conditions in Sec\.[4](https://arxiv.org/html/2611.21933#S4)\. Importantly, our approach is conceptually much simpler and computationally cheaper than adaptive tuning strategies, while simultaneously imposing the boundary conditions*exactly*instead of approximately in \([5](https://arxiv.org/html/2611.21933#S3.Ex1)\), by specifically exploiting the structure of trajectory optimization problems\.

### 3\.2Trajectory optimization with PINNs

We consider trajectory optimization governed by Newton’s equations, where the solution is the position𝐫⁡\(t\)\\mathbf\{r\}\(t\)\. For a system withNMN\_\{M\}massesMqM\_\{q\}at positions𝐏q=\(xq,yq,zq\)\\mathbf\{P\}\_\{q\}=\(x\_\{q\},y\_\{q\},z\_\{q\}\), the gravitational potentialΦ\\Phiat a position𝐫=\(x,y,z\)\\mathbf\{r\}=\(x,y,z\)is given by

Φ⁡\(𝐫\)=∑q=1NM−G​Mq‖𝐫−𝐏q‖\.\\displaystyle\\Phi\(\\mathbf\{r\}\)=\\sum\_\{q=1\}^\{N\_\{M\}\}\\frac\{\-GM\_\{q\}\}\{\\left\\lVert\\mathbf\{r\}\-\\mathbf\{P\}\_\{q\}\\right\\rVert\}\.\(7\)For a spacecraft of massmsm\_\{s\}exerting thrust𝐅\\mathbf\{F\}, Newton’s equation of motion becomes

msd2​𝐫d​t2=−∑q=1NMGmsMq𝐫−𝐏q‖𝐫−𝐏q‖3\+𝐅\.\\displaystyle m\_\{s\}\\frac\{d^\{2\}\\mathbf\{r\}\}\{dt^\{2\}\}=\-\\sum\_\{q=1\}^\{N\_\{M\}\}Gm\_\{s\}M\_\{q\}\\frac\{\\mathbf\{r\}\-\\mathbf\{P\}\_\{q\}\}\{\\left\\lVert\\mathbf\{r\}\-\\mathbf\{P\}\_\{q\}\\right\\rVert^\{3\}\}\+\\mathbf\{F\}\.\(8\)In order to use PINNs to optimize spacecraft trajectories, the physical law must be written in residual form\. Introducing the time\-of\-flight \(TOF\)TT, we use normalized timeτ:=t/T\\tau:=t/Tand note that the TOF equals the terminal time ift0=0t\_\{0\}=0\. Setting a normalized spacecraft massmS=1m\_\{S\}=1, the thrust becomes

𝐅⁡\(τ\)=1T2​d2​𝐫d​τ2\+∑q=1NMG​Mq​𝐫⁡\(τ\)−𝐏q‖𝐫⁡\(τ\)−𝐏q‖3\.\\displaystyle\\mathbf\{F\}\(\\tau\)=\\frac\{1\}\{T^\{2\}\}\\frac\{d^\{2\}\\mathbf\{r\}\}\{d\\tau^\{2\}\}\+\\sum\_\{q=1\}^\{N\_\{M\}\}GM\_\{q\}\\frac\{\\mathbf\{r\}\(\\tau\)\-\\mathbf\{P\}\_\{q\}\}\{\\left\\lVert\\mathbf\{r\}\(\\tau\)\-\\mathbf\{P\}\_\{q\}\\right\\rVert^\{3\}\}\.\(9\)In these units, we haveτ∈\[0,1\]\\tau\\in\\left\[0,1\\right\]and the spacecraft takes one normalized time unit to complete any trajectory given by a boundary value problem\. We parameterize the trajectory𝐫⁡\(τ\)\\mathbf\{r\}\(\\tau\)via a neural network𝒩θ​\(τ\)≡𝐫θ​\(τ\)\\mathcal\{N\}\_\{\\mathbf\{\\theta\}\}\(\\tau\)\\equiv\\mathbf\{r\}\_\{\\mathbf\{\\theta\}\}\(\\tau\)with weights and biasesθ\{\\mathbf\{\\theta\}\}, more precisely𝒩θ:ℝ→ℝ3\\mathcal\{N\}\_\{\\mathbf\{\\theta\}\}:\\mathbb\{R\}\\to\\mathbb\{R\}^\{3\},τ↦\(x,y,z\)⊤\\tau\\mapsto\(\{x\},\{y\},\{z\}\)^\{\\top\}, and substitute

𝐫⁡\(τ\)≈𝐫θ​\(τ\)≡𝒩θ​\(τ\)\.\\displaystyle\\mathbf\{r\}\(\\tau\)\\approx\\mathbf\{r\}\_\{\\mathbf\{\\theta\}\}\(\\tau\)\\equiv\\mathcal\{N\}\_\{\\mathbf\{\\theta\}\}\(\\tau\)\.\(10\)Following from this definition, the thrust in \([9](https://arxiv.org/html/2611.21933#S3.E9)\) becomes a function of the neural network and hence a function ofθ\{\\mathbf\{\\theta\}\}\. Such a neural network constitutes a parametrization of the ansatz function like any other ansatz function, with a major distinction being that closed forms are rarely available, but fast access to the gradient is provided by automatic differentiation\. The optimization problem, for a given TOF, then is to minimize the integrated thrust magnitude subject to constraints,

θopt\\displaystyle\\mathbf\{\\theta\}\_\{\\text\{opt\}\}=arg⁡minθ⁡ℒ⁡\(θ\)\\displaystyle=\\arg\\min\_\{\\mathbf\{\\theta\}\}\\mathcal\{L\}\(\\mathbf\{\\theta\}\)=arg⁡minθ⁡\(∑n=1N‖𝐅⁡\(τn\)‖2\+boundary condition terms\),\\displaystyle=\\arg\\min\_\{\\mathbf\{\\theta\}\}\\Bigg\(\\sum\_\{n=1\}^\{N\}\\left\\lVert\\mathbf\{F\}\(\\tau\_\{n\}\)\\right\\rVert^\{2\}\+\\text\{boundary condition terms\}\\,\\Bigg\),\(11\)atNNcollocation pointsτn∈\[0,1\]\\tau\_\{n\}\\in\\left\[0,1\\right\],n=1,⋯,Nn=1,\\cdots,N\. In other words, the trajectory𝐫θ​\(τ\)\\mathbf\{r\}\_\{\\mathbf\{\\theta\}\}\(\\tau\)corresponds to the solution ansatzuθu\_\{\\mathbf\{\\theta\}\}of the PINN in Sec\.[3\.1](https://arxiv.org/html/2611.21933#S3.SS1), while the thrust \([9](https://arxiv.org/html/2611.21933#S3.E9)\) corresponds to the physical law \([1](https://arxiv.org/html/2611.21933#S3.E1)\), which in turn results in the optimization objective \([11](https://arxiv.org/html/2611.21933#S3.Ex1a)\) corresponding to \([4](https://arxiv.org/html/2611.21933#S3.E4)\) and collocation pointsτn\\tau\_\{n\}in place of \([6a](https://arxiv.org/html/2611.21933#S3.E6.1)\) in Sec\.[3\.1](https://arxiv.org/html/2611.21933#S3.SS1)\.

The initial and terminal conditions of the trajectory are realized by a starting position vector𝐫⁡\(t0\)\\mathbf\{r\}\(t\_\{0\}\)at time\-zerot0t\_\{0\}and a terminal position vector𝐫⁡\(T\)\\mathbf\{r\}\(T\)at terminal timeTT, as well as possible constraints on the time\-derivatives of the position vectors at trajectory start and trajectory end, i\.e\.𝐫˙​\(t0\)\\dot\{\\mathbf\{r\}\}\(t\_\{0\}\),𝐫˙​\(tT\)\\dot\{\\mathbf\{r\}\}\(t\_\{T\}\),𝐫¨​\(t0\)\\ddot\{\\mathbf\{r\}\}\(t\_\{0\}\),𝐫¨​\(T\)\\ddot\{\\mathbf\{r\}\}\(T\), etc\.\.̇ In the naive PINN\-formulation, the boundary conditions of a trajectory, i\.e\. here the initial and terminal conditions, are enforced through additional loss terms at collocation points corresponding to the initial and boundary condition in \([6b](https://arxiv.org/html/2611.21933#S3.E6.2)\)\-\([6c](https://arxiv.org/html/2611.21933#S3.E6.3)\)\. Therefore, the full loss function in \([11](https://arxiv.org/html/2611.21933#S3.Ex1a)\) writes as

ℒ⁡\(θ\)=∑n=1N‖𝐅⁡\(τn\)‖2\+λ1​‖𝐫⁡\(t0\)−𝐫θ​\(t0\)‖2\+λ2​‖𝐫⁡\(T\)−𝐫θ​\(T\)‖2\+λ3​‖𝐫˙​\(t0\)−𝐫˙θ​\(t0\)‖2\+λ4​‖𝐫˙​\(T\)−𝐫˙θ​\(T\)‖2\+⋯\.\\mathcal\{L\}\(\\mathbf\{\\theta\)\}=\\sum\_\{n=1\}^\{N\}\\left\\lVert\\mathbf\{F\}\(\\tau\_\{n\}\)\\right\\rVert^\{2\}\\ \+\\lambda\_\{1\}\\left\\lVert\\mathbf\{r\}\(t\_\{0\}\)\-\\mathbf\{r\}\_\{\\mathbf\{\\theta\}\}\(t\_\{0\}\)\\right\\rVert^\{2\}\+\\lambda\_\{2\}\\left\\lVert\\mathbf\{r\}\(T\)\-\\mathbf\{r\}\_\{\\mathbf\{\\theta\}\}\(T\)\\right\\rVert^\{2\}\+\\lambda\_\{3\}\\left\\lVert\\dot\{\\mathbf\{r\}\}\(t\_\{0\}\)\-\\dot\{\\mathbf\{r\}\}\_\{\\mathbf\{\\theta\}\}\(t\_\{0\}\)\\right\\rVert^\{2\}\+\\lambda\_\{4\}\\left\\lVert\\dot\{\\mathbf\{r\}\}\(T\)\-\\dot\{\\mathbf\{r\}\}\_\{\\mathbf\{\\theta\}\}\(T\)\\right\\rVert^\{2\}\+\\cdots\.\(12\)Then the optimized trajectory is given by the solution to \([11](https://arxiv.org/html/2611.21933#S3.Ex1a)\) via the PINN prediction at optimal weights and biases,𝐫opt​\(τ\)≡𝐫θopt​\(τ\)\\mathbf\{r\}\_\{\\text\{opt\}\}\(\\tau\)\\equiv\\mathbf\{r\}\_\{\\mathbf\{\\theta\}\_\{\\text\{opt\}\}\}\(\\tau\), i\.e\. settingθ=θopt\{\\mathbf\{\\theta\}\}=\{\\mathbf\{\\theta\}\}\_\{\\text\{opt\}\}in the ansatz function\. This procedure corresponds to training an unsupervised PINN as introduced before in Sec\.[3\.1](https://arxiv.org/html/2611.21933#S3.SS1)where the physics\-lossℒphysics\\mathcal\{L\}\_\{\\text\{physics\}\}is defined as the integrated thrust magnitude\.

Conveniently, the formulation with normalized time allows us to treat the TOFTTexplicitly as an optimization parameter that can be backpropagated and learned in the same way as the weights and biasesθ\\mathbf\{\\theta\}, formally

\(θopt,Topt\)=arg⁡minθ,T⁡ℒ⁡\(θ\)\.\\displaystyle\(\\mathbf\{\\theta\}\_\{\\text\{opt\}\},T\_\{\\text\{opt\}\}\)=\\arg\\min\_\{\\mathbf\{\\theta\},T\}\\mathcal\{L\}\(\\mathbf\{\\theta\}\)\.\(13\)In the implementation, this is as simple as addingTTto the list of optimizable parameters in addition toθ\\mathbf\{\\theta\}\. In addition, Eq\. \([13](https://arxiv.org/html/2611.21933#S3.E13)\) reveals a major advantage of PINNs over classical solvers: the quantity of interest becomes auto\-differentiable with respect to the physical parameters, be it TOF or other optimizable parameters\. This in turn enablesefficientgradient\-based approaches to inverse problems\. Further, we do not need a fixed discretization of the domainΩ×\(0,T\]\\Omega\\times\(0,T\]at inference time\.

However, it is clear that even adaptive strategies face an increasingly difficult problem of tuning theλ\\lambda’s in the formulation of \([12](https://arxiv.org/html/2611.21933#S3.E12)\)\. Our experiments will demonstrate in Sec\.[5](https://arxiv.org/html/2611.21933#S5)that this naive PINN formulation leads to low accuracy in satisfying the boundary conditions, slow convergence and sometimes even prevents us from convergence to a satisfying solution at all\. In the next section, we introduce a simple transformation of the PINN to eliminate the boundary conditions in the loss function, thereby guaranteeing exact fulfillment of boundary conditions and eliminating convergence issues, and tuning effort, associated with the parametersλ\\lambdain \([12](https://arxiv.org/html/2611.21933#S3.E12)\)\.

## 4Exact boundary conditions by PINN\-transformation

We now introduce a transformed trajectory representation that satisfies the boundary conditions \(BCs\) exactly\. While for the naive PINN formulation we had𝐫θ​\(τ\)≡𝒩θ​\(τ\)\\mathbf\{r\}\_\{\\mathbf\{\\theta\}\}\(\\tau\)\\equiv\\mathcal\{N\}\_\{\\mathbf\{\\theta\}\}\(\\tau\), we define the transformed PINNρθ\\rho\_\{\\mathbf\{\\theta\}\}with exact boundary conditions as a function of𝐫θ\\mathbf\{r\}\_\{\\mathbf\{\\theta\}\}as

ρθ​\(τ\):=𝐑⁡\(τ\)\+ψ⁡\(τ\)​𝐫θ​\(τ\)\+ϕ⁡\(τ\)​\(𝐕⁡\(τ\)−𝐑˙​\(τ\)\),\\displaystyle\\mathbf\{\\rho\}\_\{\\mathbf\{\\theta\}\}\(\\tau\):=\\mathbf\{R\}\(\\tau\)\+\\psi\(\\tau\)\\mathbf\{r\}\_\{\\mathbf\{\\theta\}\}\(\\tau\)\+\\phi\(\\tau\)\\Big\(\\mathbf\{V\}\(\\tau\)\-\\dot\{\\mathbf\{R\}\}\(\\tau\)\\Big\),\(14\)with its time\-derivative

ρ˙θ​\(τ\)\\displaystyle\\dot\{\\mathbf\{\\rho\}\}\_\{\\mathbf\{\\theta\}\}\(\\tau\)=𝐑˙​\(τ\)\+ψ˙​\(τ\)​𝐫θ​\(τ\)\+ψ⁡\(τ\)​𝐫˙θ​\(τ\)\+ϕ⁡\(τ\)​\(𝐕˙​\(τ\)−𝐑¨​\(τ\)\)\+ϕ˙​\(τ\)​\(𝐕⁡\(τ\)−𝐑˙​\(τ\)\),\\displaystyle=\\dot\{\\mathbf\{R\}\}\(\\tau\)\+\\dot\{\\psi\}\(\\tau\)\\mathbf\{r\}\_\{\\mathbf\{\\theta\}\}\(\\tau\)\+\\psi\(\\tau\)\\dot\{\\mathbf\{r\}\}\_\{\\mathbf\{\\theta\}\}\(\\tau\)\+\\phi\(\\tau\)\\Big\(\\dot\{\\mathbf\{V\}\}\(\\tau\)\-\\ddot\{\\mathbf\{R\}\}\(\\tau\)\\Big\)\+\\dot\{\\phi\}\(\\tau\)\\Big\(\\mathbf\{V\}\(\\tau\)\-\\dot\{\\mathbf\{R\}\}\(\\tau\)\\Big\),\(15\)whereϕ\\phiandψ\\psiare so\-called distance functions\[[56](https://arxiv.org/html/2611.21933#bib.bib6)\]constructed such that they guarantee the boundary conditions to be fulfilled byρθ\\rho\_\{\\mathbf\{\\theta\}\}andρ˙θ\\dot\{\\rho\}\_\{\\mathbf\{\\theta\}\}at the boundary for all parameter values ofθ\\mathbf\{\\theta\}in conjunction with following choice:

𝐑⁡\(τ\)\\displaystyle\\mathbf\{R\}\(\\tau\)=τ⁡\(𝐫T−𝐫0\)\+𝐫0,\\displaystyle=\\tau\(\\mathbf\{r\}\_\{T\}\-\\mathbf\{r\}\_\{0\}\)\+\\mathbf\{r\}\_\{0\},\(16\)𝐕⁡\(τ\)\\displaystyle\\mathbf\{V\}\(\\tau\)=τ⁡\(𝐯T−𝐯0\)\+𝐯0,\\displaystyle=\\tau\(\\mathbf\{v\}\_\{T\}\-\\mathbf\{v\}\_\{0\}\)\+\\mathbf\{v\}\_\{0\},\(17\)where𝐫T=𝐫⁡\(T\)\\mathbf\{r\}\_\{T\}=\\mathbf\{r\}\(T\),𝐫0=𝐫⁡\(t0\)\\mathbf\{r\}\_\{0\}=\\mathbf\{r\}\(t\_\{0\}\),𝐯T=𝐫˙​\(T\)\\mathbf\{v\}\_\{T\}=\\dot\{\\mathbf\{r\}\}\(T\),𝐯0=𝐫˙​\(t0\)\\mathbf\{v\}\_\{0\}=\\dot\{\\mathbf\{r\}\}\(t\_\{0\}\)is defined by the boundary conditions222Note the slight abuse of notation in that we write𝐫⁡\(t\)\\mathbf\{r\}\(t\)and𝐫⁡\(τ\)\\mathbf\{r\}\(\\tau\)both with the same symbol𝐫\\mathbf\{r\}albeit the expression slightly differs depending on whether the argument is time or normalized time\. We here use un\-normalized time for clarity of argumentation regarding the requirements forϕ\\phiandψ\\psi, and avoid overloading notation to distinguish𝐫⁡\(t\)\\mathbf\{r\}\(t\),ψ⁡\(t\)\\psi\(t\)from their normalized time\-counterparts e\.g\. by𝐫~​\(τ\)\\tilde\{\\mathbf\{r\}\}\(\\tau\)andψ~​\(τ\)\\tilde\{\\psi\}\(\\tau\)\. Note that this normalization must be anticipated in the boundary conditions\.\. Then, for the transformed PINN to satisfy the boundary conditions∀θ\\forall\\mathbf\{\\theta\}, we need to construct \(i\)ψ\\psisuch thatψ⁡\(t0\)=0\\psi\(t\_\{0\}\)=0,ψ⁡\(T\)=0\\psi\(T\)=0,ψ˙​\(t0\)=0\\dot\{\\psi\}\(t\_\{0\}\)=0andψ˙​\(T\)=0\\dot\{\\psi\}\(T\)=0, as well as \(ii\)ϕ\\phisuch thatϕ⁡\(t0\)=0\\phi\(t\_\{0\}\)=0,ϕ⁡\(T\)=0\\phi\(T\)=0,ϕ˙​\(t0\)=1\\dot\{\\phi\}\(t\_\{0\}\)=1andϕ˙​\(T\)=1\\dot\{\\phi\}\(T\)=1\. Settingt0=0t\_\{0\}=0and recalling normalized timeτ:=t/T,τ∈\[0,1\]\\tau:=t/T,~\\tau\\in\[0,1\], one possible solution is

ψ⁡\(τ\)\\displaystyle\\psi\(\\tau\)=τ2​\(1−τ\)2,\\displaystyle=\\tau^\{2\}\(1\-\\tau\)^\{2\},\(18\)ϕ⁡\(τ\)\\displaystyle\\phi\(\\tau\)=2​τ3−3​τ2\+τ\.\\displaystyle=2\\tau^\{3\}\-3\\tau^\{2\}\+\\tau\.\(19\)We may then solve the trajectory optimization problem with exact boundary conditions by substitutingρθ\\mathbf\{\\rho\}\_\{\\mathbf\{\\theta\}\}for𝐫θ\\mathbf\{r\}\_\{\\mathbf\{\\theta\}\}in Eq\.[9](https://arxiv.org/html/2611.21933#S3.E9)and Eq\. \([11](https://arxiv.org/html/2611.21933#S3.Ex1a)\), respectively\. In this construction, the transformed PINN is always and a priori consistent with the boundary conditions and the velocity𝐯​\(t\)≡ρ˙​\(t\)\\mathbf\{v\}\(t\)\\equiv\\dot\{\\mathbf\{\\rho\}\}\(t\), which is easily verified by the substitution of \([16](https://arxiv.org/html/2611.21933#S4.E16)\)\-\([19](https://arxiv.org/html/2611.21933#S4.E19)\) into \([14](https://arxiv.org/html/2611.21933#S4.E14)\) and \([15](https://arxiv.org/html/2611.21933#S4.E15)\)\. This consistency is not fulfilled in the ordinary PINN, nor in other constructions such as two neural networks learning position and velocity independently, respectively, nor by a single neural network with multiple outputs learning position and velocity jointly\.

### 4\.1Implementation remarks

We construct a neural network𝒩θ:ℝ→ℝ3\\mathcal\{N\}\_\{\\mathbf\{\\theta\}\}:\\mathbb\{R\}\\to\\mathbb\{R\}^\{3\},τ↦\(x,y,z\)⊤\\tau\\mapsto\(\{x\},\{y\},\{z\}\)^\{\\top\}, and define the transformed trajectoryρθ​\(τ\)\\mathbf\{\\rho\}\_\{\\mathbf\{\\theta\}\}\(\\tau\)via the transformation above\. I\.e\. both𝐫θ\\mathbf\{r\}\_\{\\mathbf\{\\theta\}\}andρθ\\mathbf\{\\rho\}\_\{\\mathbf\{\\theta\}\}take time as input and yield a three\-dimensional vector as output\. By construction,ρθ​\(τ\)\\mathbf\{\\rho\}\_\{\\mathbf\{\\theta\}\}\(\\tau\)satisfies the boundary conditions for allθ\\mathbf\{\\theta\}and remains fully differentiable with respect to bothτ\\tauandθ\\mathbf\{\\theta\}\. Training is then achieved by solving the optimization problem with either fixed TOF Eq\. \([11](https://arxiv.org/html/2611.21933#S3.Ex1a)\) or optimizable TOF Eq\. \([13](https://arxiv.org/html/2611.21933#S3.E13)\), and the argument of free terminal time from the PINN in Sec\.[3\.1](https://arxiv.org/html/2611.21933#S3.SS1)carries over to the transformed PINN\. The explicit expression forρ˙θ​\(τ\)\\dot\{\\mathbf\{\\rho\}\}\_\{\\mathbf\{\\theta\}\}\(\\tau\)is not required in implementation, since automatic differentiation provides all necessary derivatives\. In other words, the derivative expression \([15](https://arxiv.org/html/2611.21933#S4.E15)\) is only used to derive suitable conditions forϕ⁡\(τ\)\\phi\(\\tau\)andψ⁡\(τ\)\\psi\(\\tau\)to enforce the boundary conditions\.

### 4\.2Generalizations

Regarding construction of the transformation, it may be advantageous to introduce an alternative distance functionχ⁡\(τ\)\\chi\(\\tau\)and consider a modified construction of the formχ⁡\(τ\)​𝐑​\(τ\)\\chi\(\\tau\)\\,\\mathbf\{R\}\(\\tau\)\. This can be used to eliminate𝐑˙​\(τ\)\\dot\{\\mathbf\{R\}\}\(\\tau\)at the boundaryτ∈\{0,1\}\\tau\\in\\\{0,1\\\}and replace the term\(𝐕​\(τ\)−𝐑˙​\(τ\)\)\\big\(\\mathbf\{V\}\(\\tau\)\-\\dot\{\\mathbf\{R\}\}\(\\tau\)\\big\)by𝐑⁡\(τ\)\\mathbf\{R\}\(\\tau\)\. Moreover, the distance functionsϕ⁡\(τ\)\\phi\(\\tau\),ψ⁡\(τ\)\\psi\(\\tau\), orχ⁡\(τ\)\\chi\(\\tau\)may themselves contain learnable parameters\. The proposed construction also naturally extends beyond kinematic constraints, for instance to \(i\) alternative coordinate systems such as cylindrical or spherical coordinates, \(ii\) higher\-order constraints by extending the construction to𝐫¨\\ddot\{\\mathbf\{r\}\}, or \(iii\) constraints in phase space\. In particular, parametrizing higher\-order distance functions may enable the learning of maneuvers with short impulses or acceleration limits and shall be the subject of future work\.

### 4\.3Interpretation as constraint embedding of a Rayleigh\-type variational problem

The proposed transformation can be interpreted as a reparameterization of the solution space that embeds the boundary conditions directly into the function class\. While for the ordinary PINN without exact BCs \(i\.e\. only soft BCs\) we have𝐫θ∈𝒱θ\\mathbf\{r\}\_\{\\mathbf\{\\theta\}\}\\in\\mathcal\{V\}\_\{\\mathbf\{\\theta\}\}with𝒱θ⊂𝒱\\mathcal\{V\}\_\{\\mathbf\{\\theta\}\}\\subset\\mathcal\{V\}\(cf\. Sec\.[3\.1](https://arxiv.org/html/2611.21933#S3.SS1)and recall our definitionuθ≡𝐫θu\_\{\\mathbf\{\\theta\}\}\\equiv\\mathbf\{r\}\_\{\\mathbf\{\\theta\}\}\), the transformed PINN with exact BCs isρθ∈𝒱BC\\mathbf\{\\rho\}\_\{\\mathbf\{\\theta\}\}\\in\\mathcal\{V\}\_\{\\mathrm\{BC\}\}with𝒱BC⊂𝒱θ\\mathcal\{V\}\_\{\\mathrm\{BC\}\}\\subset\\mathcal\{V\_\{\\mathbf\{\\theta\}\}\}\. Instead of enforcing constraints through penalty terms, which leads to a multi\-objective optimization problem, the transformation restricts the function space generated by the neural network to functions that satisfy the boundary conditions by construction\. As a result, the original constrained optimization problem is converted into an unconstrained optimization over admissible trajectories\. This perspective explains both the improved convergence behavior and the reduced sensitivity to hyperparameters observed in the numerical experiments\. Lastly, the resulting optimization problem corresponds to a continuous\-thrust generalization of a Lambert boundary value problem and to a Rayleigh\-type variational problem in which the integrated control effort is minimized over the space of admissible trajectories\.

## 5Numerical Experiments

We demonstrate five numerical experiments of increasing complexity: \(A\) a swing by trajectory for a spacecraft passing three celestial bodies in 2D, \(B\) a swing by trajectory for a spacecraft passing three celestial bodies in 3D, \(C\) a circular orbital transfer in polar coordinates with fixed terminal angle, \(D\) a circular orbital transfer in polar coordinates with free terminal angle, and \(E\) hold\-point rendezvous maneuver in low earth orbit\. In all experiments, we compare our transformed PINN with exact boundary conditions \(BCs\) to a baseline obtained with the classical numerical solver OpenGoddard\[[32](https://arxiv.org/html/2611.21933#bib.bib60)\]\. Where possible, we also compare our transformed PINN with exact boundary conditions to the naive PINN without exact boundary conditions \(i\.e\. soft BCs\)\. However, in examples \(C\)\-\(E\) the naive PINNs failed to converge to a satisfactory solution, suggesting that our method from Sec\.[4](https://arxiv.org/html/2611.21933#S4)enables more difficult problems that are beyond reach with the naive PINN\. In all cases, we compare to the “best” naive PINN, meaning that the PINN with optimized weighting parameters, i\.e\. theλ\\lambda’s in \([5](https://arxiv.org/html/2611.21933#S3.Ex1)\), is selected\. Supplementary material on the grid search and an illustration of the computational challenges to select the bestλ\\lambda’s is provided in Appendix[A](https://arxiv.org/html/2611.21933#A1)\. In all experiments, we report the velocity increment defined via the integrated thrust magnitude as

Δ​V=∫0T\|\|𝐅⁡\(t\)\|\|m⁡\(t\)​𝑑t\\displaystyle\\Delta V=\\int\_\{0\}^\{T\}\\frac\{\\lvert\\lvert\\mathbf\{F\}\(t\)\\rvert\\rvert\}\{m\(t\)\}\\,dt\(20\)where for the sake of simplicity we setm⁡\(t\)≡1m\(t\)\\equiv 1\. The method can in principle be extended to non\-constant masses\. The integral was computed using 1000 equi\-spaced collocation points, i\.e\. ten times as many as during training\. Note that the thrust\-term in \([11](https://arxiv.org/html/2611.21933#S3.Ex1a)\) is a discrete approximation closely related toΔ​V\\Delta Vin \([20](https://arxiv.org/html/2611.21933#S5.E20)\)\.

For all neural networks, we used 3 layers with 50 densely connected neurons each\. The input was the normalized time, and the output was a position in 2D\- or 3D\-space, respectively\. For the examples of Sec\.[5\.1](https://arxiv.org/html/2611.21933#S5.SS1)\-[5\.2](https://arxiv.org/html/2611.21933#S5.SS2), this output was in cartesian coordinates, whilst examples of Sec\.[5\.3](https://arxiv.org/html/2611.21933#S5.SS3)\-[5\.5](https://arxiv.org/html/2611.21933#S5.SS5)were in polar coordinates\. For training, ADAM and LBFGS optimizers were used consecutively\. The trajectory was considered converged after the relative change between two training iterations is smaller thanε\\varepsilon, withε=10−6\\varepsilon=10^\{\-6\}in the experiments of Sec\.[5\.1](https://arxiv.org/html/2611.21933#S5.SS1)\-[5\.2](https://arxiv.org/html/2611.21933#S5.SS2)andε=10−5\\varepsilon=10^\{\-5\}in the experiments of Sec\.[5\.3](https://arxiv.org/html/2611.21933#S5.SS3)\-[5\.5](https://arxiv.org/html/2611.21933#S5.SS5)\. The number of ADAM iterations was set individually for all experiments while the maximum number of iterations for the LBFGS optimizer was set to10510^\{5\}in the experiments of Sec\.[5\.1](https://arxiv.org/html/2611.21933#S5.SS1)\-[5\.2](https://arxiv.org/html/2611.21933#S5.SS2)\. No LBFGS was necessary in the experiments of Sec\.[5\.3](https://arxiv.org/html/2611.21933#S5.SS3)\-[5\.5](https://arxiv.org/html/2611.21933#S5.SS5)\. The loss term was computed on 100 linearly equi\-spaced collocation points for normalized time\.

Unless stated otherwise, examples reportingnormalized units\(Sec\.[5\.1](https://arxiv.org/html/2611.21933#S5.SS1), Sec,[5\.2](https://arxiv.org/html/2611.21933#S5.SS2)\) employ dimensionless variables, where time is normalized by the TOF, the spacecraft mass is set to unity, and spatial coordinates are scaled such that the computational domain spans a unit square\. The resultingΔ​V\\Delta Vfor all examples across all methods is summarized in Sec\.[5\.6](https://arxiv.org/html/2611.21933#S5.SS6)\. All examples report figures of the trajectories, training curves, required force magnitude \(RFM\) and gravity experienced by the spacecraft and thrust magnitude along the trajectory\. The computational time for PINN training and the classical solver is comparable and on the order of minutes for the demonstrated examples\.

All codes are implemented in PyTorch and available at[https://github\.com/s\-schoenherr/SpacePINN](https://github.com/s-schoenherr/SpacePINN)\. We defer a parameter study with respect to fixed and free TOF to Appendix[B](https://arxiv.org/html/2611.21933#A2)\.

### 5\.1Swing\-by with three celestial bodies in 2D

We seek to optimize a swing\-by trajectory for a spacecraft in a given gravitational potential and given boundary conditions\. This problem was originally posed by Seo\[[53](https://arxiv.org/html/2611.21933#bib.bib5)\]\. We will start with two spatial dimensions and then extend to three dimensions\. The gravitational potential is defined by three celestial body coordinates and masses in the\(x,y\)\(x,y\)plane in Tab\.[1](https://arxiv.org/html/2611.21933#S5.T1)\.

Table 1:Positions and gravitational massesG​MiGM\_\{i\}of celestial bodiesIn this first example, our neural network will map𝒩θ:ℝ→ℝ2\\mathcal\{N\}\_\{\\mathbf\{\\theta\}\}:\\mathbb\{R\}\\to\\mathbb\{R\}^\{2\},τ↦\(x,y\)⊤\\tau\\mapsto\(\{x\},\{y\}\)^\{\\top\}, omitting thezz\-coordinate\. We will later extend this to three dimensions\. We do this to rule out solutions that deviate from the\(x,y\)\(x,y\)\-plane and isolate a direct comparison in the first experiment, whilst such solutions are permitted in the second experiment\.

The boundary conditions arer​\(t0\)=\(−1,−1\)\\textbf\{r\}\(t\_\{0\}\)=\(\-1,\-1\)andr​\(T\)=\(1,1\)\\textbf\{r\}\(T\)=\(1,1\), respectively\. The initial and terminal state velocities remain free, i\.e\. we setϕ⁡\(t\)≡0\\phi\(t\)\\equiv 0in \([14](https://arxiv.org/html/2611.21933#S4.E14)\)\. The TOF is optimizable with an initial guess ofT=1T=1\. We found local optima for the TOF close to the initial value, however hyperparameter tuning is more sensitive with optimizable TOF\. The results for both the best naive PINN \(λB​C=0\.324\\lambda\_\{BC\}=0\.324\) and the transformed PINN with exact boundary conditions are shown in Fig\.[1\(a\)](https://arxiv.org/html/2611.21933#S5.F1.sf1)\. Whilst both trajectories show reasonable qualitative agreement with the classical solver’s baseline, the PINNs cannot compete in terms of integrated thrust magnitudeΔ​V\\Delta V\(cf\. Tab\.[2](https://arxiv.org/html/2611.21933#S5.T2)\)\. While the transformed PINN substantially improves convergence behavior and exact satisfaction of boundary conditions compared to the naive PINN, the classical solver remains superior in terms of integrated thrust optimality\. We attribute this to the fact that the PINN solves the variational problem only approximately through residual minimization at finitely many collocation points and through a highly nonconvex neural\-network parametrization\. The solution trajectory \(top left panel\) of the naive PINN is qualitatively similar to the transformed PINN with exact boundary conditions\. However, the naive PINN deviates substantially from the boundary conditions and converges significantly slower to an optimum with larger physics loss \(top right panel\)\. The convergence curve also shows that the naive PINN stalls at a residual error for the boundary conditions on the order of10−410^\{\-4\}\. Lastly, the acting gravity and thrust magnitude \(bottom panels\) differ substantially quantitatively between the two models\. In particular, transformed PINN with exact boundary conditions achieves consistently lower required thrust compared to the naive PINN, see also summary metrics and statistics in Sec\.[5\.6](https://arxiv.org/html/2611.21933#S5.SS6)\.

We found that such differences in the convergence curves persist throughout the experiments, and we offer a simple explanation for the phenomenon\. While the naive PINN needs to balance the satisfaction of the boundary conditions against minimizing the thrust, the transformed PINN need not make any such balance and solely needs to minimize the thrust\.

\(a\)Trajectories\. Green disks are celestial bodies\.\(b\)Training curve\.\(c\)Gravity and required force magnitude \(RFM\)\.\(d\)Thrust magnitude\.
Figure 1:Swing\-by trajectory in 2D\. Comparison of the naive PINN, our method \(exact BC\), and the baseline\. All units normalized\. Arrows scaled for visualization\.
### 5\.2Swing\-by with three celestial bodies in 3D

We extend the previous 2D experiment in Sec\.[5\.1](https://arxiv.org/html/2611.21933#S5.SS1)to a plane in three dimensions\. The celestial body coordinates as well as the boundary conditions remain the same as in Sec\.[5\.1](https://arxiv.org/html/2611.21933#S5.SS1)within the\(x,y\)\(x,y\)\-plane, i\.e\. thezz\-coordinates of the boundary conditions and celestial body coordinates set to zero\. The output of the PINNs is now a position in 3D, i\.e\.𝒩θ:ℝ→ℝ3\\mathcal\{N\}\_\{\\mathbf\{\\theta\}\}:\\mathbb\{R\}\\to\\mathbb\{R\}^\{3\},τ↦\(x,y,z\)⊤\\tau\\mapsto\(\{x\},\{y\},\{z\}\)^\{\\top\}\.

The results for both the optimal naive PINN \(withλB​C=0\.421\\lambda\_\{BC\}=0\.421\) and the transformed PINN with exact boundary conditions are shown in Fig\.[2](https://arxiv.org/html/2611.21933#S5.F2)\. The naive PINN converges to a solution trajectory \(top left panel set\) that leaves the\(x,y\)\(x,y\)\-plane and circumvents the gravitational centers altogether333We confirmed that this is a physically meaningful solution as follows: The naive PINNs solution was used as an initial guess for a separate run with the numerical solver OpenGoddard, which then in turn would converge to a qualitatively similar solution\. In contrast, OpenGoddard initialized with a straight line between the initial and terminal position converged to the baseline trajectory shown in Fig\.[2](https://arxiv.org/html/2611.21933#S5.F2)\.\.

In contrast, the solution of the transformed PINN with exact boundary conditions still finds an in\-plane trajectory qualitatively similar to the baseline\. This however demanded a more sophisticated initialization or pre\-conditioning scheme as follows: First, we start by initializing a transformed PINN with auxiliary boundary conditions for initial and final state velocities𝐯0=𝐯T=\[1,1,0\]\\mathbf\{v\}\_\{0\}=\\mathbf\{v\}\_\{T\}=\[1,1,0\], in addition to the initial and terminal positions, in order to prime the PINN for an in\-plane trajectory\. Second, after pre\-training for 2000 iterations with ADAM, we find a sub\-optimal trajectory in the\(x,y\)\(x,y\)\-plane\. Third, the auxiliary boundary conditions are relaxed again, i\.e\. training resumes with the pre\-conditioned PINN, pre\-optimized TOF as well as now free \(optimizable\) initial and terminal velocities\. We found that the transformed PINN with exact boundary conditions converges to a qualitatively similar solution as the optimal naive PINN unless this pre\-conditioning step is applied\. We interpret the pre\-conditioning step as an additional, potentially symmetry breaking implicit biases\. Importantly, this pre\-conditioning step is not available in the naive PINN formulation without escalation of hyperparameter tuning for lack of a direct mechanism to impose boundary conditions exactly\.

Regarding convergence and thrust as a function of time, we have similar findings as in the 2D example, with the notable exception that it is indeed more favorable to escape the\(x,y\)\(x,y\)\-plane for a minimal thrust swing\-by, as is confirmed by the convergence curve \(top right panel\)\.

\(a\)Trajectories\. Green disks are celestial bodies\.\(b\)Training curve\.\(c\)Gravity and required force magnitude \(RFM\)\.\(d\)Thrust magnitude\.
Figure 2:Swing\-by trajectory in 3D\. Comparison of the naive PINN, our method \(exact BC\) with and without preconditioning, and the baseline\. All units normalized\. Arrows scaled for visualization\.
### 5\.3Circular Orbit Transfer with Fixed Terminal Angle

Next we consider an orbit transfer problem from a circular Low Earth Orbit \(LEO\) to a higher circular Medium Earth Orbit \(MEO\)\. For this, our method from Sec\.[4](https://arxiv.org/html/2611.21933#S4)can be extended straightforward to polar coordinates with polar unit vectors𝐞r\\mathbf\{e\}\_\{r\}and𝐞α\\mathbf\{e\}\_\{\\alpha\}as follows\.

𝐫\\displaystyle\\mathbf\{r\}=r​𝐞r\\displaystyle=r\\mathbf\{e\}\_\{r\}𝐫˙\\displaystyle\\dot\{\\mathbf\{r\}\}=r˙​𝐞r\+r​α˙​𝐞α\\displaystyle=\\dot\{r\}\\mathbf\{e\}\_\{r\}\+r\\dot\{\\alpha\}\\mathbf\{e\}\_\{\\alpha\}𝐫¨\\displaystyle\\ddot\{\\mathbf\{r\}\}=\(r¨−r​α˙2\)​𝐞r\+\(2​r˙​α˙\+r​α¨\)​𝐞α\\displaystyle=\(\\ddot\{r\}\-r\\dot\{\\alpha\}^\{2\}\)\\mathbf\{e\}\_\{r\}\+\(2\\dot\{r\}\\dot\{\\alpha\}\+r\\ddot\{\\alpha\}\)\\mathbf\{e\}\_\{\\alpha\}\(21\)withα\\alphabeing the azimuthal angle coordinate in radians andrrthe norm of𝐫\\mathbf\{r\}\. We may substitute𝐫⁡\(τ\)\\mathbf\{r\}\(\\tau\)for a neural network𝐫θ:𝒩θ​\(τ\)→\(r⁡\(τ\),α⁡\(τ\)\)\\mathbf\{r\}\_\{\\mathbf\{\\theta\}\}:\\mathcal\{N\}\_\{\\mathbf\{\\theta\}\}\(\\tau\)\\to\\big\(r\(\\tau\),\\alpha\(\\tau\)\\big\)into \([11](https://arxiv.org/html/2611.21933#S3.Ex1a)\) to obtain a PINN that outputs polar coordinates\(r,α\)\(r,\\alpha\)as a function of normalized time\. Note that one must use the acceleration in polar coordinates \([21](https://arxiv.org/html/2611.21933#S5.E21)\) for the thrust vector \([9](https://arxiv.org/html/2611.21933#S3.E9)\)\. The transformation of the PINN to satisfy boundary conditions is then analogous to Sec\.[4](https://arxiv.org/html/2611.21933#S4), however for this type of problem the boundary conditions typically are slightly different\. We may now consider boundary conditions

𝐫0\\displaystyle\\mathbf\{r\}\_\{0\}=𝐫⁡\(t0\)=\(r0,α0\)T\\displaystyle=\\mathbf\{r\}\(t\_\{0\}\)=\\big\(r\_\{0\},\\alpha\_\{0\}\\big\)^\{T\}𝐫T\\displaystyle\\mathbf\{r\}\_\{T\}=𝐫⁡\(T\)=\(rT,αT\)T\.\\displaystyle=\\mathbf\{r\}\(T\)=\\big\(r\_\{T\},\\alpha\_\{T\}\\big\)^\{T\}\.\(22\)withr0=RLEO=6871​kmr\_\{0\}=R\_\{\\text\{LEO\}\}=6871~\\text\{km\}, corresponding to 500 km above ground, andrT=RMEO=8371​kmr\_\{T\}=R\_\{\\text\{MEO\}\}=8371~\\text\{km\}, corresponding to 2000 km above ground, and a fixed radian phases ofα0=0\\alpha\_\{0\}=0andαT=π\\alpha\_\{T\}=\\pi\. In orbital transfer problems, we may typically desire to constrain the orbits throughr⁡\(t0\)r\(t\_\{0\}\)andr⁡\(T\)r\(T\)as well as radial velocityr˙​\(T\)=0\\dot\{r\}\(T\)=0but refrain from constraining the angular phasesα⁡\(t0\)\\alpha\(t\_\{0\}\),α⁡\(T\)\\alpha\(T\)or tangential velocityα˙​\(τ\)\\dot\{\\alpha\}\(\\tau\)\. For this experiment, we do constrain the phases as defined above in order to obtain a solution that is comparable to an analytic solution of a Hohmann\-transfer\. We will relax this constraint in another experiment later in Sec\.[5\.4](https://arxiv.org/html/2611.21933#S5.SS4)\. The initial and terminal orbital velocities of the LEO and MEO were constrained toα˙​\(t0\)=VLEORLEO\\dot\{\\alpha\}\(t\_\{0\}\)=\\frac\{V\_\{\\mathrm\{LEO\}\}\}\{R\_\{\\mathrm\{LEO\}\}\}withVLEO=G​MEarthRLEO=7\.61​km/sV\_\{\\mathrm\{LEO\}\}=\\sqrt\{\\frac\{GM\_\{\\mathrm\{Earth\}\}\}\{R\_\{\\mathrm\{LEO\}\}\}\}=7\.61~\\mathrm\{km/s\}andα˙​\(T\)=VMEORMEO\\dot\{\\alpha\}\(T\)=\\frac\{V\_\{\\mathrm\{MEO\}\}\}\{R\_\{\\mathrm\{MEO\}\}\}withVMEO=G​MEarthRMEO=6\.90​km/sV\_\{\\mathrm\{MEO\}\}=\\sqrt\{\\frac\{GM\_\{\\mathrm\{Earth\}\}\}\{R\_\{\\mathrm\{MEO\}\}\}\}=6\.90~\\mathrm\{km/s\}, and the initial and terminal radial velocity was constrained to zero,r˙​\(0\)=r˙​\(T\)=0\\dot\{r\}\(0\)=\\dot\{r\}\(T\)=0\. Lastly, the TOF is treated as optimizable and was initialized atTinit=π​\(RLEO\+RMEO\)38​G​MEarth≈3315​sT\_\{\\mathrm\{init\}\}=\\pi\\sqrt\{\\frac\{\(R\_\{\\mathrm\{LEO\}\}\+R\_\{\\mathrm\{MEO\}\}\)^\{3\}\}\{8GM\_\{\\mathrm\{Earth\}\}\}\}\\approx 3315~\\mathrm\{s\}\.

The model was trained with ADAM until a relative loss change of 1e\-5\. The results are shown in Fig\.[3](https://arxiv.org/html/2611.21933#S5.F3)\. For reference, we compare the obtained trajectory to the classical Hohmann transfer, which provides the optimal impulsiveΔ​V\\Delta Vfor transfers between circular coplanar orbits\. While the Hohmann solution assumes instantaneous velocity changes, the proposed method yields a continuous low\-thrust trajectory\. The trajectory \(top left panel\) of the transformed PINN with exact boundary conditions is qualitatively again similar to OpenGoddard’s numerical solver baseline\. However, in this example the transformed PINN achieves a lowerΔ​V\\Delta Vthan the classical solver, which yields relatively large impulses close to start and end of the trajectory \(bottom panels\)\. We observe that the transformed PINN solution produces a smooth thrust profile that approximates the impulsive solution while maintaining comparable total control effort\.

We observed that this holds even when we run the classical solver again with a warm re\-start where the initialization is based on the transformed PINN’s resulting superior trajectory\. The training curve \(top right\) shows relatively fast convergence of the transformed PINN with exact boundary conditions\. Note the naive PINN without exact boundary conditions \(i\.e\. soft BCs\) failed to converge to a satisfactory solution and is hence omitted here\.

\(a\)Trajectories\.\(b\)Training curve\.\(c\)Gravity and required force magnitude \(RFM\)\.\(d\)Thrust magnitude\.
Figure 3:Circular orbit transfer with fixed terminal angle\. Comparison of our transformed PINN with exact boundary conditions to the OpenGoddard baseline\.
### 5\.4Circular Orbit Transfer with Free Terminal Angle

We extend the previous example by relaxing the constraints on the phase angle, in particular with free terminal phase, and we consider a transfer from Low Earth Orbit \(LEO\) to Geostationary Earth Orbit \(GEO\)\. With an initial phase ofα⁡\(t0\)=0\\alpha\(t\_\{0\}\)=0, the initial boundary conditions are given by\(r⁡\(t0\),α⁡\(t0\)\)T=\(RLEO,0\)T\(r\(t\_\{0\}\),\\alpha\(t\_\{0\}\)\)^\{T\}=\(R\_\{\\mathrm\{LEO\}\},0\)^\{T\}and\(r˙​\(t0\),α˙​\(t0\)\)T=\(0,VLEO/RLEO\)T\(\\dot\{r\}\(t\_\{0\}\),\\dot\{\\alpha\}\(t\_\{0\}\)\)^\{T\}=\(0,V\_\{\\mathrm\{LEO\}\}/R\_\{\\mathrm\{LEO\}\}\)^\{T\}\. At the terminal point, the phase angle is free\. Thus, only the orbital radius and tangential velocity are prescribed, i\.e\.r⁡\(T\)=RGEOr\(T\)=R\_\{\\mathrm\{GEO\}\}and\(r˙​\(T\),α˙​\(T\)\)T=\(0,VGEO/RGEO\)T\(\\dot\{r\}\(T\),\\dot\{\\alpha\}\(T\)\)^\{T\}=\(0,V\_\{\\mathrm\{GEO\}\}/R\_\{\\mathrm\{GEO\}\}\)^\{T\}whileα⁡\(T\)\\alpha\(T\)is optimized\.

The optimization was initialized with a terminal\-angle guess of2\.10​π2\.10\\,\\piand an initial guess for the TOF that equals the Hohmann transfer time multiplied by a factor of1\.091\.09\. This problem proved to be sensitive to the initial guess of the terminal angle as well as the TOF\. Hence, a grid search over both quantities was carried out in order to identify a robust initialization region with favorable convergence behavior and low control cost\. The model was trained with ADAM until a relative loss change of 1e\-5\.

The results are shown in Fig\.[4](https://arxiv.org/html/2611.21933#S5.F4)\. The transformed PINN with exact boundary conditions converged reliably to a smooth low\-thrust transfer and produced a trajectory qualitatively similar to the OpenGoddard reference solution\. As in the previous example, the naive PINN without exact boundary conditions \(i\.e\. soft BCs\) failed to converge to a satisfactory solution\. Both trajectories become tangential to the orbit as they close in on the orbit444This is not obviously visible for the baseline on this figure’s scale due to the impulse\-like correction happening only shortly before entering orbit\. We omit an additional plot on smaller axes scales for brevity\.\.

\(a\)Trajectories\.\(b\)Training curve\.\(c\)Gravity and required force magnitude \(RFM\)\.\(d\)Thrust magnitude\.
Figure 4:Circular orbit transfer with free terminal angle\. Comparison of our transformed PINN with exact boundary conditions to the OpenGoddard baseline\.
### 5\.5Hold\-Point Rendezvous in Low Earth Orbit

To complement the orbital\-transfer examples by a local proximity\-operations scenario, we consider a rendezvous hold\-point problem on a circular orbit in the earth\-centered inertial frame\. The target follows the same circular LEO reference orbit as in the previous orbit\-transfer example[5\.3](https://arxiv.org/html/2611.21933#S5.SS3)\. Introducing subscriptsaaandbbfor chaser and target, respectively, in polar coordinates, its initial state is\(rb​\(t0\),αb​\(t0\)\)T=\(RLEO,0\)T\(r\_\{b\}\(t\_\{0\}\),\\alpha\_\{b\}\(t\_\{0\}\)\)^\{T\}=\(R\_\{\\mathrm\{LEO\}\},0\)^\{T\}and\(r˙b​\(t0\),α˙b​\(t0\)\)T=\(0,VLEO/RLEO\)T\.\(\\dot\{r\}\_\{b\}\(t\_\{0\}\),\\dot\{\\alpha\}\_\{b\}\(t\_\{0\}\)\)^\{T\}=\\left\(0,V\_\{\\mathrm\{LEO\}\}/\{R\_\{\\mathrm\{LEO\}\}\}\\right\)^\{T\}\.Although the PINN is implemented in polar coordinates, the rendezvous scenario is most naturally specified in the target\-centered Cartesian frame\. Let𝐱b​\(t\)=RLEO​𝐞r​\(t\)\\mathbf\{x\}\_\{b\}\(t\)=R\_\{\\mathrm\{LEO\}\}~\\mathbf\{e\}\_\{r\}\(t\)and𝐱˙b​\(t\)=VLEO​𝐞α​\(t\)\\dot\{\\mathbf\{x\}\}\_\{b\}\(t\)=V\_\{\\mathrm\{LEO\}\}~\\mathbf\{e\}\_\{\\alpha\}\(t\)denote the target position and velocity, where𝐞r​\(t\)\\mathbf\{e\}\_\{r\}\(t\)and𝐞α​\(t\)\\mathbf\{e\}\_\{\\alpha\}\(t\)are the radial and along\-track unit vectors of the target orbit\. The chaser position is represented relative to the target as𝐱a​\(t\)=𝐱b​\(t\)\+δ​ra​\(t\)​𝐞r​\(t\)\+δ​sa​\(t\)​𝐞α​\(t\)\\mathbf\{x\}\_\{a\}\(t\)=\\mathbf\{x\}\_\{b\}\(t\)\+\\delta r\_\{a\}\(t\)\\mathbf\{e\}\_\{r\}\(t\)\+\\delta s\_\{a\}\(t\)\\mathbf\{e\}\_\{\\alpha\}\(t\)whereδ​ra​\(t\)\\delta r\_\{a\}\(t\)andδ​sa​\(t\)\\delta s\_\{a\}\(t\)are the radial and along\-track relative offsets, respectively\. The initial relative offset is chosen asδ​ra​\(t0\)=−1\.0​km\\delta r\_\{a\}\(t\_\{0\}\)=\-1\.0\\,\\mathrm\{km\}andδ​sa​\(t0\)=−0\.3​km\\delta s\_\{a\}\(t\_\{0\}\)=\-0\.3\\,\\mathrm\{km\}with𝐱˙a​\(t0\)=𝐱˙b​\(t0\)\.\\dot\{\\mathbf\{x\}\}\_\{a\}\(t\_\{0\}\)=\\dot\{\\mathbf\{x\}\}\_\{b\}\(t\_\{0\}\)\.The terminal point is defined as a hold point with a prescribed radial offseth=30h=30~m outside the target orbit𝐱a​\(T\)=𝐱b​\(T\)\+h​𝐞r​\(T\)\\mathbf\{x\}\_\{a\}\(T\)=\\mathbf\{x\}\_\{b\}\(T\)\+h\\mathbf\{e\}\_\{r\}\(T\)\. This corresponds to matching the target’s angular rate at the terminal point𝐱˙a​\(T\)=\(RLEO\+h\)​VLEORLEO​𝐞α​\(T\)\.\\dot\{\\mathbf\{x\}\}\_\{a\}\(T\)=\(R\_\{\\mathrm\{LEO\}\}\+h\)\\frac\{V\_\{\\mathrm\{LEO\}\}\}\{R\_\{\\mathrm\{LEO\}\}\}\\mathbf\{e\}\_\{\\alpha\}\(T\)\.For training, these Cartesian boundary states are converted to polar coordinates\(r,α\)\(r,\\alpha\)and polar velocities\(r˙,α˙\)\(\\dot\{r\},\\dot\{\\alpha\}\)\. Altogether, the problem is solved as a free\-TOF low\-thrust rendezvous\-to\-hold\-point transfer with exact boundary conditions\. The model is trained with ADAM until a relative loss change of 10e\-7\.

The results are shown in Fig\.[5](https://arxiv.org/html/2611.21933#S5.F5), including an orbit\-frame overview of the global motion, a local vertical local horizontal \(LVLH\) relative\-motion plot, a thrust\-magnitude history in minutes and the total loss convergence plot\. The transformed PINN successfully computes a low\-thrust rendezvous trajectory to the prescribed hold point, satisfying all boundary conditions exactly and exhibiting stable convergence without the need for boundary\-condition loss tuning\.

\(a\)Orbit frame overview and trajectories\.\(b\)Training curve\.\(c\)LVLH Relative Motion, target centered at the origin\.\(d\)Thrust magnitude\.
Figure 5:Hold\-Point Rendezvous in LEO with initial offset between target and chase\.
### 5\.6Summary metrics and statistical analyses

TheΔ​V\\Delta Vof all above examples is summarized for all methods in Tab\.[2](https://arxiv.org/html/2611.21933#S5.T2)\. Furthermore, we show box plots ofΔ​V\\Delta Vfor 100 randomly initialized neural networks in Fig\.[6](https://arxiv.org/html/2611.21933#S5.F6), to demonstrate robustness and consistency of performance\. We limit the statistical comparison to the examples in Sec\.[5\.1](https://arxiv.org/html/2611.21933#S5.SS1)and Sec\.[5\.2](https://arxiv.org/html/2611.21933#S5.SS2)since for the advanced examples \(Sec\.[5\.3](https://arxiv.org/html/2611.21933#S5.SS3), Sec\.[5\.4](https://arxiv.org/html/2611.21933#S5.SS4)and Sec\.[5\.5](https://arxiv.org/html/2611.21933#S5.SS5)\), the naive PINN with soft BCs failed to converge\.

In 2D, the transformed PINN with exact BC consistently outperforms PINN with soft BC\. In 3D, the transformed PINN with exact BC again improves over PINN with soft BC, while the pre\-trained variant attains competitive median performance but exhibits a small number of outliers that failed to converge\. The exact\-BC formulation achieves lowerΔ​V\\Delta V, shorter TOF, and faster convergence across most random seed initializations\. The pre\-trained variant improves median performance but introduces heavy tails, visible in theΔ​V\\Delta Vand convergence distributions\. We find consistently that the naive PINN underperforms or fails because it must trade off objectives, whilst the transformed PINN removes that trade\-off entirely\. By restricting the function space to feasible trajectories, the transformation modifies the optimization landscape and effectively removes infeasible local minima that arise from boundary condition violations\. Mind that different methods can converge to other distinct local optima – not related to the boundary conditions, nevertheless – resulting in additional outliers\.

\(a\)Swing\-by in 2D \(Sec\.[5\.1](https://arxiv.org/html/2611.21933#S5.SS1)\)\. From left to right:Δ​V\\Delta V, TOF, training iterations until convergence\.\(b\)Swing\-by in 3D \(Sec\.[5\.2](https://arxiv.org/html/2611.21933#S5.SS2)\)\. From left to right:Δ​V\\Delta V, TOF, training iterations until convergence\.
Figure 6:Box plots for 100 randomly initialized PINNs, comparing naive PINNs with soft BCs to our transformed PINNs with exact BCs, and a pre\-conditioned variant of the latter\.Table 2:Summary metrics and comparisons for the five numerical examples Sec\.[5\.1](https://arxiv.org/html/2611.21933#S5.SS1)to Sec\.[5\.5](https://arxiv.org/html/2611.21933#S5.SS5)\. We used 100 samples each\. PINNs with soft BCs did not converge in some examples \(denoted as n\.c\.\)\.∗One outlier run was excluded from the reported statistics\.

∗∗Seven outlier runs were excluded from the reported statistics\.

∗∗∗Specifically, the solution by PINN with exact BCs was input as initial guess for a new solve with OpenGoddard\.

†Baseline optimization stopped after reaching iteration budget\.

## 6Discussion

The numerical results presented in Sec\.[5](https://arxiv.org/html/2611.21933#S5)demonstrate that the proposed transformation\-based PINN formulation enables exact enforcement of boundary and kinematic constraints while substantially improving convergence behavior compared to ordinary PINNs\. Across all test cases, transformed PINN with exact boundary conditions \(BCs\) either \(i\) consistently achieved lower residual losses, faster convergence, and smoother thrust profiles without requiring adjustment of boundary\-loss weights, or \(ii\) enabled converged solutions in the first place\.

The results imply a fundamental limitation of standard PINN formulations for trajectory optimization: the need to balance competing objectives in a multi\-objective loss function\. The introduction of boundary\-condition penalty terms leads to a sensitivity with respect to loss weights, which can significantly affect both convergence and solution quality\. By eliminating these penalty terms entirely, the proposed method removes a key source of instability and, in the absence of additional objectives or constraints, reduces the training problem to a single\-objective optimization\.

From a methodological perspective, the proposed transformation can be interpreted as a reparameterization of the solution space that embeds the boundary conditions directly into the function class\. This converts the original constrained optimization problem into an unconstrained optimization over admissible trajectories\. This viewpoint explains the improved robustness and reduced hyperparameter sensitivity observed in the experiments and connects the approach to classical ideas in constrained optimization\.

In comparison to classical trajectory optimization methods, such as direct collocation, the proposed approach offers a fundamentally different trade\-off\. While classical methods operate on discretized trajectories and require explicit enforcement of constraints at collocation points, the present formulation yields a continuous representation of the trajectory that is \(auto\-\)differentiable with respect to both time and problem parameters\. This enables gradient\-based optimization in function space and provides a natural pathway toward solving inverse problems and sensitivity analysis within the same framework\.

The comparison with the Hohmann\-type transfer illustrates that the proposed method can recover physically meaningful solutions that are consistent with classical astrodynamics\. While the Hohmann transfer represents an optimal impulsive solution, the PINN\-based approach produces a smooth, continuous low\-thrust trajectory that approximates this behavior without requiring explicit discretization of control inputs\. This suggests that the PINN solution can be interpreted as a continuous generalization of classicalΔ​V\\Delta V\-optimal transfers\.

The observed performance gains were more pronounced for increasingly complex dynamical systems, where the ordinary PINN often failed to converge, converged to the “wrong” solution or required extensive tuning ofλBC\\lambda\_\{\\mathrm\{BC\}\}\. In contrast, the transformed PINN with exact boundary conditions maintained reliable convergence even in free TOF optimization\. The transformed formulation also integrates seamlessly with existing PINN toolchains, preserving differentiability and using standard optimizers such as ADAM and L\-BFGS\.

Overall, the results indicate that embedding constraints directly into the neural network parameterization is a promising direction in trajectory optimization\. By combining exact constraint satisfaction with continuous representations and automatic differentiation, the proposed framework provides a flexible alternative to traditional methods\.

### 6\.1Limitations

At the same time, the experiments reveal several limitations\. In higher\-dimensional settings, such as the 3D swing\-by problem, the optimization landscape may contain multiple, physically meaningful local minima\. As a result, the method can benefit from suitable initialization or pre\-conditioning strategies, but exact boundary\-condition enforcement alone does not guarantee convergence to globally optimal trajectories\. This behavior is consistent with the non\-convex nature of neural network training and is not unique to the proposed approach; similar challenges arise in classical methods such as direct collocation, albeit with more explicit control through initialization\.

Beyond this, the present study is limited to relatively simple problems with idealized gravitational models\. Extending the approach to realistic astrodynamics scenarios will require incorporating perturbations, eclipses, and higher\-order dynamics\. Moreover, scalability to long\-duration missions may benefit from domain decomposition or adaptive collocation strategies, as the current formulation relies on a single monolithic neural network\. While the proposed transformation removes the need for boundary\-condition tuning, performance still depends on neural\-network architecture and optimizer settings, which warrant further systematic investigation\. Although convergence behavior is improved empirically, a formal analysis of convergence rates and numerical conditioning remains an open challenge\. Finally, the presence of multiple local optima can lead to qualitatively distinct solution families, not all of which may be desirable from a mission\-design perspective\. While such local minima often correspond to physically meaningful trajectories, they complicate optimization and require careful initialization or guidance\. This challenge is not unique to the proposed approach and also arises in classical methods such as direct collocation, where more explicit convergence control can be achieved through tailored initialization strategies\.

Importantly, the proposed approach directly applies to equality\-type kinematic constraints; extensions to inequality\-, path\- or control constraints such as bounded thrust remains an open direction\. Overall, the proposed transformation introduces a principled way to embed boundary information directly into the neural representation for trajectory optimization with practically relevant advantages\.

### 6\.2Outlook

Our PINN\-based approach, incl\. the exact imposition of boundary conditions, could also be combined with the popular reinforcement learning approaches\[[57](https://arxiv.org/html/2611.21933#bib.bib64),[8](https://arxiv.org/html/2611.21933#bib.bib65),[51](https://arxiv.org/html/2611.21933#bib.bib62),[23](https://arxiv.org/html/2611.21933#bib.bib66),[24](https://arxiv.org/html/2611.21933#bib.bib67)\]to a “physics\-informed reinforcement learning" approach\. Another promising direction would be neural operators\[[41](https://arxiv.org/html/2611.21933#bib.bib14)\], incl\. physics\-informed variants of neural operators\[[39](https://arxiv.org/html/2611.21933#bib.bib69)\]\.

Operator learning offers a natural next step for extending the present framework beyond trajectory optimization toward learning entire solution maps between function spaces\. Instead of parameterizing a single trajectory, operator\-based models can learn mappings from boundary conditions, system parameters, or force fields directly to optimal trajectories\. This could enable rapid amortized inference across families of problems and reduce the need to retrain models for each new instance\. In combination with exact constraint enforcement as proposed in this work, operator learning could provide a scalable pathway toward real\-time, physics\-consistent trajectory design\.

## 7Conclusion

In this work, we introduced a simple yet effective transformation of physics\-informed neural networks \(PINNs\) for spacecraft trajectory optimization that enforces boundary conditions exactly by construction\. By embedding boundary conditions directly into the function class via distance functions, the proposed approach reduces the need for penalty\-based enforcement and reduces the problem to a single\-objective formulation\. This alleviates a central challenge in standard PINNs, namely sensitivity to loss\-weight tuning and training instability\.

Across trajectory optimization problems of increasing complexity, including multi\-body swing\-by maneuvers and low\-thrust orbital transfers, we showed improved convergence, robustness, and solution quality compared to naive PINN formulations\. The transformed PINN consistently achieves exact boundary satisfaction, avoids residual constraint violations, and remains stable even where standard PINNs fail to converge\. Moreover, the method can yield competitive performance relative to classical solvers while offering additional flexibility through auto\-differentiability with respect to physical parameters such as the TOF and without requiring fixed discretizations\.

At the same time, several challenges remain\. Performance still depends on architectural and optimization choices, and complex problems may exhibit multiple physically meaningful local optima, necessitating suitable initialization or pre\-conditioning strategies\. Extending the framework to more realistic astrodynamics scenarios with perturbations and operational constraints is an important next step\. Finally, a rigorous theoretical analysis of convergence and numerical conditioning remains an open question\.

More broadly, the proposed approach can be viewed as a constraint\-embedding strategy for scientific machine learning\. By reparameterizing the solution space to satisfy constraints a priori, it provides a principled alternative to penalty\-based formulations that removes a fundamental failure mode of PINNs\.

We anticipate that this perspective can contribute to improving robustness and reliability of physics\-informed learning methods in engineering and the physical sciences\.

## Appendix

## Appendix AHyperparameter tuning

For fair comparison between an ordinary PINN and a transformed PINN with exact boundary conditions, the only tunable hyperparameter is the loss\-weightλB​C\\lambda\_\{BC\}pertaining to the boundary condition\-term in the loss function and settingλB​C=λi,∀i∈\{1,2,3,4\}\\lambda\_\{BC\}=\\lambda\_\{i\},~~\\forall i\\in\\\{1,2,3,4\\\}\. This was achieved via a simple grid search; training 1000 models in the range of\[10−3,103\]\[10^\{\-3\},10^\{3\}\]\. The results are shown in Fig\.[7](https://arxiv.org/html/2611.21933#A1.F7)for both swing\-by examples in 2D and 3D, Sec\.[5\.1](https://arxiv.org/html/2611.21933#S5.SS1)and[5\.2](https://arxiv.org/html/2611.21933#S5.SS2)\.

The PyTorch LBFGS optimizer was used with parametermax\_iter=10, corresponding to at most 10 internal quasi\-Newton iterations per optimizer step\. The settings for the optimizers and resulting hyperparameters are shown in Table[3](https://arxiv.org/html/2611.21933#A1.T3)\. We omit the corresponding results for the orbital transfer examples \(Sec\.[5\.3](https://arxiv.org/html/2611.21933#S5.SS3)and[5\.4](https://arxiv.org/html/2611.21933#S5.SS4)\) due to failure to find a solution that converges to a satisfying solution at all\.

The resulting hyperparameter search is visualized in Fig\.[7](https://arxiv.org/html/2611.21933#A1.F7)\. The blue curve shows the smallest total loss value of a training process withNA​D​A​MN\_\{ADAM\}= 2000 and at maximumNL​B​F​G​SN\_\{LBFGS\}=10510^\{5\}iterations, for a givenλB​C\\lambda\_\{BC\}and trainable TOF, initialized withT=1T=1\. The networks were trained 1000 times, respectively\. We observe in Fig\.[7](https://arxiv.org/html/2611.21933#A1.F7)that the ordinary PINN without exact boundary conditions \(i\.e\. soft BCs\) is sensitive to the choice ofλB​C\\lambda\_\{BC\}\. We also observe that the minimum of the total loss does not necessarily need to coincide with the minimum of the boundary condition\-loss term alone\.

Table 3:Optimizer parameters and results of boundary loss\-weight \(λB​C\\lambda\_\{BC\}\) grid search\. From left to right: problem dimension, total TOFTTset to fixed or optimizable, number of training iterations with ADAM, maximum number of training iterations with LBFGS optimizer, respectively, and the best resulting boundary weightλB​C\\lambda\_\{BC\}\.Dim\.TTNA​D​A​MN\_\{ADAM\}max⁡\(NL​B​F​G​S\)\\max\(N\_\{LBFGS\}\)BestλB​C\\lambda\_\{BC\}2Optimizable20001000000\.3243Optimizable20001000000\.421\(a\)2D problem \(Sec\.[5\.1](https://arxiv.org/html/2611.21933#S5.SS1)\) with optimizable TOF\(b\)3D problem \(Sec\.[5\.2](https://arxiv.org/html/2611.21933#S5.SS2)\) with optimizable TOF
Figure 7:Boundary loss weightλB​C\\lambda\_\{BC\}grid search for the naive PINNs\. Best total loss values are marked with crosses\.
## Appendix BInfluence of TOF

Before, the experiment was conducted with a free terminal time or time\-of\-flight \(TOF\)TTaccording to \([13](https://arxiv.org/html/2611.21933#S3.E13)\), where the initial guess forTTat the start of the optimization wasT=1T=1\. Next we want to investigate the influence of the TOF parameter on the optimization and solution\. For this, we consider a variant of the experiment in Sec\.[5\.2](https://arxiv.org/html/2611.21933#S5.SS2)with fixed TOFTTand conduct a parameter sweep overTTin the rangeT∈\[1,1\.1\]T\\in\\left\[1,1\.1\\right\]in increments of0\.020\.02units as well as for discrete valuesT∈\{1\.0,1\.25,1\.5,1\.75,2\.0,3\.0\}T\\in\\\{1\.0,1\.25,1\.5,1\.75,2\.0,3\.0\\\}\. The purpose of this particular set of parameter values is to better visualize a bifurcation with respect toTT\.

The results are shown in Fig\.[8](https://arxiv.org/html/2611.21933#A2.F8)\. We observe from Fig\.[8\(a\)](https://arxiv.org/html/2611.21933#A2.F8.sf1)and[8\(b\)](https://arxiv.org/html/2611.21933#A2.F8.sf2)that a bifurcation is present roughly in the vicinity ofT∈\[1\.08,1\.1\]T\\in\[1\.08,1\.1\]\. As expected, for larger TOFs, the detour out\-of\-plane trajectory encountered before is favorable over a direct trajectory, and the arc length of the detour trajectory increases with fixed TOF\. This confirms physical consistency in that constraining the TOF to smaller values necessitates a shorter arc length whilst unconstrained TOF translates to virtually no restrictions on the trajectory’s arc length\. The thrust magnitude for the direct trajectory in Fig\.[8\(c\)](https://arxiv.org/html/2611.21933#A2.F8.sf3)and the detour trajectory in Fig\.[8\(d\)](https://arxiv.org/html/2611.21933#A2.F8.sf4)show that the thrust is indeed lower when allowing detour trajectories by allowing larger TOFs\. This is confirmed by the experienced gravity force as a function of time shown in Fig\.[8\(e\)](https://arxiv.org/html/2611.21933#A2.F8.sf5)and[8\(f\)](https://arxiv.org/html/2611.21933#A2.F8.sf6), respectively, where the three centers of gravity are increasingly experienced as one center of gravity at increasing distance\.

We conclude that \(i\) transformed PINNs can indeed find physically meaningful trajectories according to their constraints and that \(ii\) the TOF has a major influence on the experienced gravitational force and hence on the loss landscape and in turn on the convergence behavior\. The transition between the two regimes may indicate that the bifurcation causes a high\-loss barrier between the two local optima that result in the two qualitatively distinct family of trajectories\. In other words, if the PINN is initialized in either one regime, a high\-loss barrier prevents a gradient\-based optimizer from ever exploring the other regime\. Whilst our pre\-condition step from Sec\.[5\.2](https://arxiv.org/html/2611.21933#S5.SS2)can mitigate this issue, it remains future work to obtain full control of convergence behavior e\.g\. by finding appropriate annealing schedules for the TOF\. Nevertheless, the fact that the TOF can be treated as an auto\-differentiable optimization variable is a major advantage of the PINN framework over traditional numerical schemes, which is exacerbated in that the TOF in trajectory optimization usually appears in the boundary conditions\.

\(a\)\(b\)\(c\)\(d\)\(e\)\(f\)
Figure 8:Influence of TOF: Parameter sweep for fixed TOF values\. Left column:T∈\[1\.0,1\.1\]T\\in\\left\[1\.0,1\.1\\right\]in increments of 0\.02, resulting in direct trajectories on or close to the\(x,y\)\(x,y\)\-plane\. Right column:T∈\{1\.0,1\.25,1\.5,1\.75,2\.0,3\.0\}T\\in\\\{1\.0,1\.25,1\.5,1\.75,2\.0,3\.0\\\}resulting in detour trajectories leaving the\(x,y\)\(x,y\)\-plane\. Top row: Resulting trajectories\. Middle row: resulting thrust magnitude\. Bottom row: Gravitational pull experienced by the spacecraft\.
## Funding Sources

S\.R\. was financially supported by the Austrian Science Fund \(FWF\), Grant DOI: 10\.55776/J4774\.

## Acknowledgments

LLMs were used for stylistic improvements of the introduction\.

## References

- \[1\]\(2025\)Physics\-informed neural networks for vessel trajectory prediction: learning time\-discretized kinematic dynamics via finite differences\.InProceedings of the 19th International Symposium on Spatial and Temporal Data,pp\. 55–65\.External Links:[Document](https://dx.doi.org/10.1145/3748777.3748779)Cited by:[§2\.1](https://arxiv.org/html/2611.21933#S2.SS1.p1.1)\.
- \[2\]C\. Ampatzis and D\. Izzo\(2009\)Machine learning techniques for approximation of objective functions in trajectory optimisation\.InProceedings of the IJCAI 2009 Workshop on Artificial Intelligence in Space,ESA Special Publication, Vol\.ESA SP\-673,Pasadena, California, USA,pp\. 1–6\.Note:Workshop held 17–18 July 2009External Links:ISBN 978\-92\-9221\-237\-7Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p2.1)\.
- \[3\]G\. P\. Badura, E\. Arunkumar, M\. Velez\-Reyes, B\. Gunter, and K\. Ho\(2025\)Convergence and uncertainty of physics\-informed machine learning for cislunar orbit determination\.Journal of Spacecraft and Rockets62\(6\),pp\. 2212–2231\.External Links:[Document](https://dx.doi.org/10.2514/1.A36313)Cited by:[§2\.3](https://arxiv.org/html/2611.21933#S2.SS3.p2.1)\.
- \[4\]G\. P\. Badura, M\. Velez\-Reyes, B\. Gunter, C\. R\. Valenta, and K\. Ho\(2024\)Regularizing training of physics informed neural networks \(pinns\) for cislunar orbit determination via transfer learning\.InProceedings of the Advanced Maui Optical and Space Surveillance \(AMOS\) Technologies Conference,pp\. 17\.Cited by:[1st item](https://arxiv.org/html/2611.21933#S2.I1.i1.p1.1),[§2\.3](https://arxiv.org/html/2611.21933#S2.SS3.p2.1),[§2\.4](https://arxiv.org/html/2611.21933#S2.SS4.p1.1)\.
- \[5\]A\. L\. Barth, D\. Roach, O\. Ma, M\. Hernandez, A\. Fleming, and R\. Ashbaugh\(2025\)Machine learning\-based optimal trajectory planning for spacecraft passively flying around a satellite for proximity operations\.InAIAA SCITECH 2025 Forum,pp\. 1003\.External Links:[Document](https://dx.doi.org/10.2514/6.2025-1003)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p2.1)\.
- \[6\]S\. Basir and I\. Senocak\(2022\)Physics and equality constrained artificial neural networks: application to forward and inverse problems with multi\-fidelity data fusion\.Journal of Computational Physics463,pp\. 111301\.External Links:[Document](https://dx.doi.org/10.1016/j.jcp.2022.111301)Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p1.1),[§3\.1](https://arxiv.org/html/2611.21933#S3.SS1.p6.1)\.
- \[7\]B\. Benedikter, A\. D’Ambrosio, and R\. Furfaro\(2025\)Rocket ascent trajectory optimization via physics\-informed pontryagin neural networks\.InAIAA SCITECH 2025 Forum,pp\. 2532\.External Links:[Document](https://dx.doi.org/10.2514/6.2025-2532)Cited by:[§2\.3](https://arxiv.org/html/2611.21933#S2.SS3.p1.1)\.
- \[8\]S\. Bonasera, N\. Bosanac, C\. J\. Sullivan, I\. Elliott, N\. Ahmed, and J\. W\. McMahon\(2023\)Designing sun–earth l2 halo orbit stationkeeping maneuvers via reinforcement learning\.Journal of Guidance, Control, and Dynamics46\(2\),pp\. 301–311\.External Links:[Document](https://dx.doi.org/10.2514/1.G006783)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p2.1),[§6\.2](https://arxiv.org/html/2611.21933#S6.SS2.p1.1)\.
- \[9\]S\. L\. Brunton, J\. Nathan Kutz, K\. Manohar, A\. Y\. Aravkin, K\. Morgansen, J\. Klemisch, N\. Goebel, J\. Buttrick, J\. Poskin, A\. W\. Blom\-Schieber, T\. Hogan, and D\. McDonald\(2021\)Data\-driven aerospace engineering: Reframing the industry with machine learning\.AIAA Journal59\(8\),pp\. 2820–2847\.External Links:[Document](https://dx.doi.org/10.2514/1.J060131)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p1.1)\.
- \[10\]S\. Cai, Z\. Mao, Z\. Wang, M\. Yin, and G\. E\. Karniadakis\(2021\)Physics\-informed neural networks \(pinns\) for fluid mechanics: a review\.Acta Mechanica Sinica37\(12\),pp\. 1727–1738\.External Links:[Document](https://dx.doi.org/10.1007/s10409-021-01148-1)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p3.1)\.
- \[11\]X\. Cao, Y\. Cai, Y\. Li, S\. Xiaoqiang, L\. Chen, and H\. Wang\(2025\)Intelligent vehicle trajectory tracking control based on physics\-informed neural network dynamics model\.Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering239\(7\),pp\. 2315–2331\.External Links:[Document](https://dx.doi.org/10.1177/0954407024124)Cited by:[§2\.1](https://arxiv.org/html/2611.21933#S2.SS1.p1.1)\.
- \[12\]S\. Cuomo, V\. S\. Di Cola, F\. Giampaolo, G\. Rozza, M\. Raissi, and F\. Piccialli\(2022\)Scientific machine learning through physics–informed neural networks: where we are and what’s next\.Journal of Scientific Computing92\(3\),pp\. 88\.External Links:[Document](https://dx.doi.org/10.1007/s10915-022-01939-z)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p3.1)\.
- \[13\]D\. Dalton, A\. Lazarus, H\. Gao, and D\. Husmeier\(2024\)Boundary constrained gaussian processes for robust physics\-informed machine learning of linear partial differential equations\.Journal of Machine Learning Research25\(272\),pp\. 1–61\.External Links:[Link](http://jmlr.org/papers/v25/23-1508.html)Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p1.1),[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p3.1)\.
- \[14\]A\. Das and D\. K\. Giri\(2025\)Physics informed neural network for optimal rendezvous guidance of low\-thrust cubesats\.In2025 IEEE Space, Aerospace and Defence Conference \(SPACE\),pp\. 1–6\.External Links:[Document](https://dx.doi.org/10.1109/SPACE65882.2025.11171316)Cited by:[1st item](https://arxiv.org/html/2611.21933#S2.I1.i1.p1.1),[§2\.3](https://arxiv.org/html/2611.21933#S2.SS3.p1.1)\.
- \[15\]M\. G\. Dissanayake and N\. Phan\-Thien\(1994\)Neural\-network\-based approximations for solving partial differential equations\.communications in Numerical Methods in Engineering10\(3\),pp\. 195–201\.External Links:[Document](https://dx.doi.org/10.1002/cnm.1640100303)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p3.1)\.
- \[16\]K\. Drozd, R\. Furfaro, and D\. Mortari\(2024\)Rapidly exploring random trees with physics\-informed neural networks for constrained energy\-optimal rendezvous problems\.The Journal of the Astronautical Sciences71\(1\),pp\. 9\.External Links:[Document](https://dx.doi.org/10.1007/s40295-023-00426-3)Cited by:[§2\.3](https://arxiv.org/html/2611.21933#S2.SS3.p1.1)\.
- \[17\]A\. D’Ambrosio, B\. Benedikter, and R\. Furfaro\(2025\)Physics\-informed pontryagin neural networks for path\-constrained optimal control problems\.Journal of Guidance, Control, and Dynamics,pp\. 1–17\.External Links:[Document](https://dx.doi.org/10.2514/1.G008854)Cited by:[§2\.3](https://arxiv.org/html/2611.21933#S2.SS3.p1.1),[§2\.3](https://arxiv.org/html/2611.21933#S2.SS3.p2.1),[§2\.4](https://arxiv.org/html/2611.21933#S2.SS4.p1.1)\.
- \[18\]A\. D’Ambrosio and R\. Furfaro\(2024\)Learning fuel\-optimal trajectories for space applications via pontryagin neural networks\.Aerospace11\(3\),pp\. 228\.External Links:[Document](https://dx.doi.org/10.3390/aerospace11030228)Cited by:[1st item](https://arxiv.org/html/2611.21933#S2.I1.i1.p1.1),[§2\.3](https://arxiv.org/html/2611.21933#S2.SS3.p1.1),[§2\.4](https://arxiv.org/html/2611.21933#S2.SS4.p1.1)\.
- \[19\]A\. D’Ambrosio, E\. Schiassi, F\. Curti, and R\. Furfaro\(2021\)Physics\-informed neural networks for optimal proximity maneuvers with collision avoidance around asteroids\.InProceedings of the AAS/AIAA Astrodynamics Specialist Conference\. Big Sky: AAS/AIAA,Cited by:[§2\.3](https://arxiv.org/html/2611.21933#S2.SS3.p1.1)\.
- \[20\]S\. A\. Faroughi, N\. Pawar, C\. Fernandes, M\. Raissi, S\. Das, N\. K\. Kalantari, and S\. K\. Mahjour\(2022\)Physics\-guided, physics\-informed, and physics\-encoded neural networks in scientific computing\.arXiv preprint arXiv:2211\.07377\.Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p3.1)\.
- \[21\]L\. Federici, B\. Benedikter, and A\. Zavoli\(2021\)Deep learning techniques for autonomous spacecraft guidance during proximity operations\.Journal of Spacecraft and Rockets58\(6\),pp\. 1774–1785\.External Links:[Document](https://dx.doi.org/10.2514/1.A35076)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p2.1)\.
- \[22\]W\. N\. Fife, P\. Ghosh, and K\. J\. DeMars\(2024\)Probabilistic trajectory design via approximate gaussian mixture steering\.InProceedings of the AAS/AIAA Astrodynamics Specialist Conference,Broomfield, CO\.Note:AAS Paper 24\-153\. arXiv:2401\.02396Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p2.1)\.
- \[23\]B\. Gaudet, R\. Furfaro, and R\. Linares\(2020\)Reinforcement learning for angle\-only intercept guidance of maneuvering targets\.Aerospace Science and Technology99,pp\. 105746\.External Links:[Document](https://dx.doi.org/10.1016/j.ast.2020.105746)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p2.1),[§6\.2](https://arxiv.org/html/2611.21933#S6.SS2.p1.1)\.
- \[24\]B\. Gaudet, R\. Linares, and R\. Furfaro\(2020\)Deep reinforcement learning for six degree\-of\-freedom planetary landing\.Advances in Space Research65\(7\),pp\. 1723–1741\.External Links:[Document](https://dx.doi.org/10.1016/j.asr.2019.12.030)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p2.1),[§6\.2](https://arxiv.org/html/2611.21933#S6.SS2.p1.1)\.
- \[25\]M\. Ghanifar, A\. Nikkhah, M\. Kamzan, M\. Teshnehlab, and M\. Tayefi\(2025\)Online trajectory regeneration for multirotors via a proportional\-derivative physics\-informed neural network\.IEEE Access\.External Links:[Document](https://dx.doi.org/10.1109/ACCESS.2025.3598862)Cited by:[§2\.1](https://arxiv.org/html/2611.21933#S2.SS1.p1.1),[§2\.1](https://arxiv.org/html/2611.21933#S2.SS1.p2.1)\.
- \[26\]T\.C\. Goldman and K\.J\. Cowan\(2024\)An unsupervised physics\-informed neural network for finding optimal low\-thrust transfer trajectories with the direct method\.InAdvances in the Astronautical Sciences,\{\. A\. \}\. Christian, \{\. \}\. Mashiku, \{\. \}\. Valerino, and \{\. \}\. Mehta \(Eds\.\),Advances in the Astronautical Sciences\(English\)\.Note:2024 AAS/AIAA Astrodynamics Specialist ConferenceCited by:[§2\.3](https://arxiv.org/html/2611.21933#S2.SS3.p2.1)\.
- \[27\]J\. L\. Gonzalo and C\. Colombo\(2021\)On\-board collision avoidance applications based on machine learning and analytical methods\.In8th European Conference on Space Debris, ESA/ESOC, Darmstadt, Germany, Virtual Conference,pp\. 20–23\.Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p2.1)\.
- \[28\]T\. Guffanti, D\. Gammelli, S\. D’Amico, and M\. Pavone\(2024\)Transformers for trajectory optimization with application to spacecraft rendezvous\.In2024 IEEE Aerospace Conference,pp\. 1–13\.External Links:[Document](https://dx.doi.org/10.1109/AERO58975.2024.10521334)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p2.1)\.
- \[29\]N\. Gunantara\(2018\)A review of multi\-objective optimization: methods and its applications\.Cogent Engineering5\(1\),pp\. 1502242\.External Links:[Document](https://dx.doi.org/10.1080/23311916.2018.1502242)Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p3.1)\.
- \[30\]M\. Härkönen, M\. Lange\-Hegermann, and B\. Raiţă\(2023\)Gaussian process priors for systems of linear partial differential equations with constant coefficients\.InProceedings of the 40th International Conference on Machine Learning \(ICML\),pp\. 12587–12615\.External Links:2212\.14319Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p3.1)\.
- \[31\]S\. K\. Ibrahim, A\. Ahmed, M\. A\. E\. Zeidan, and I\. E\. Ziedan\(2020\)Machine learning techniques for satellite fault diagnosis\.Ain Shams Engineering Journal11\(1\),pp\. 45–56\.External Links:[Document](https://dx.doi.org/10.1016/j.asej.2019.08.006)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p1.1)\.
- \[32\]Interstellar Technologies Inc\.\(2017\)OpenGoddard: Trajectory Optimization for Python\.Note:[https://github\.com/istellartech/OpenGoddard](https://github.com/istellartech/OpenGoddard)Version 1\.x, accessed: 2026\-04\-10Cited by:[4th item](https://arxiv.org/html/2611.21933#S2.I1.i4.p1.1),[§5](https://arxiv.org/html/2611.21933#S5.p1.1)\.
- \[33\]M\.R\. Jones, T\.J\. Rogers, and E\.J\. Cross\(2023\)Constraining gaussian processes for physics\-informed acoustic emission mapping\.Mechanical Systems and Signal Processing188,pp\. 109984\.External Links:ISSN 0888\-3270,[Document](https://dx.doi.org/10.1016/j.ymssp.2022.109984)Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p1.1),[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p3.1)\.
- \[34\]G\. E\. Karniadakis, I\. G\. Kevrekidis, L\. Lu, P\. Perdikaris, S\. Wang, and L\. Yang\(2021\)Physics\-informed machine learning\.Nature Reviews Physics3\(6\),pp\. 422–440\.External Links:[Document](https://dx.doi.org/10.1038/s42254-021-00314-5)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p3.1),[§3\.1](https://arxiv.org/html/2611.21933#S3.SS1.p1.1)\.
- \[35\]I\. E\. Lagaris, A\. Likas, and D\. I\. Fotiadis\(1998\)Artificial neural networks for solving ordinary and partial differential equations\.IEEE transactions on neural networks9\(5\),pp\. 987–1000\.External Links:[Document](https://dx.doi.org/10.1109/72.712178)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p3.1)\.
- \[36\]R\. Lam, M\. Poloczek, P\. Frazier, and K\. E\. Willcox\(2018\)Advances in bayesian optimization with applications in aerospace engineering\.In2018 AIAA Non\-Deterministic Approaches Conference,pp\. 1656\.External Links:[Document](https://dx.doi.org/10.2514/6.2018-1656)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p1.1)\.
- \[37\]M\. Lange\-Hegermann\(2018\)Algorithmic linearly constrained gaussian processes\.InProceedings of the 32nd International Conference on Neural Information Processing Systems \(NeurIPS\),Vol\.31,pp\. 2141–2152\.Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p3.1)\.
- \[38\]C\. Li and R\. Zeng\(2026\)AW\-el\-pinns: a multi\-task learning physics\-informed neural network for euler\-lagrange systems in optimal control problems\.Neural Networks199,pp\. 108694\.External Links:ISSN 0893\-6080,[Document](https://dx.doi.org/10.1016/j.neunet.2026.108694)Cited by:[§2\.1](https://arxiv.org/html/2611.21933#S2.SS1.p2.1)\.
- \[39\]Z\. Li, H\. Zheng, N\. Kovachki, D\. Jin, H\. Chen, B\. Liu, K\. Azizzadenesheli, and A\. Anandkumar\(2024\)Physics\-informed neural operator for learning partial differential equations\.ACM/IMS Journal of Data Science1\(3\),pp\. 1–27\.External Links:[Document](https://dx.doi.org/10.1145/3648506)Cited by:[§6\.2](https://arxiv.org/html/2611.21933#S6.SS2.p1.1)\.
- \[40\]S\. Liu, H\. Zhongkai, C\. Ying, H\. Su, J\. Zhu, and Z\. Cheng\(2022\)A unified hard\-constraint framework for solving geometrically complex pdes\.Advances in Neural Information Processing Systems35,pp\. 20287–20299\.External Links:[Document](https://dx.doi.org/10.52202/068431-1475)Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p3.1)\.
- \[41\]L\. Lu, P\. Jin, G\. Pang, Z\. Zhang, and G\. E\. Karniadakis\(2021\)Learning nonlinear operators via deeponet based on the universal approximation theorem of operators\.Nature machine intelligence3\(3\),pp\. 218–229\.External Links:[Document](https://dx.doi.org/10.1038/s42256-021-00302-5)Cited by:[§6\.2](https://arxiv.org/html/2611.21933#S6.SS2.p1.1)\.
- \[42\]K\. Luong, M\. Wahab, and J\. Lee\(2025\)Simultaneous imposition of initial and boundary conditions via decoupled physics\-informed neural networks for solving initial\-boundary value problems\.Applied Mathematics and Mechanics46\(4\),pp\. 763–780\.External Links:[Document](https://dx.doi.org/10.1007/s10483-025-3240-7)Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p3.1)\.
- \[43\]F\. Masi, I\. Stefanou, P\. Vannucci, and V\. Maffi\-Berthier\(2021\)Thermodynamics\-based artificial neural networks for constitutive modeling\.Journal of the Mechanics and Physics of Solids147,pp\. 104277\.External Links:[Document](https://dx.doi.org/10.1016/j.jmps.2020.104277)Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p2.1)\.
- \[44\]N\. E\. Michek, P\. Mehta, and W\. W\. Huebsch\(2024\)Flight dynamic uncertainty quantification modeling using physics\-informed neural networks\.AIAA Journal62\(11\),pp\. 4234–4246\.External Links:[Document](https://dx.doi.org/10.2514/1.J063992)Cited by:[§2\.3](https://arxiv.org/html/2611.21933#S2.SS3.p2.1)\.
- \[45\]F\. Pioch, J\. H\. Harmening, A\. M\. Müller, F\. Peitzmann, D\. Schramm, and O\. e\. Moctar\(2023\)Turbulence modeling for physics\-informed neural networks: comparison of different rans models for the backward\-facing step flow\.Fluids8\(2\),pp\. 43\.External Links:[Document](https://dx.doi.org/10.3390/fluids8020043)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p1.1)\.
- \[46\]M\. Raissi, P\. Perdikaris, and G\. E\. Karniadakis\(2019\)Physics\-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations\.Journal of Computational Physics378,pp\. 686–707\.External Links:[Document](https://dx.doi.org/10.1016/j.jcp.2018.10.045)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p3.1),[§3\.1](https://arxiv.org/html/2611.21933#S3.SS1.p1.1)\.
- \[47\]S\. Ranftl and W\. von der Linden\(2021\)Bayesian surrogate analysis and uncertainty propagation\.InPhysical Sciences Forum,Vol\.3,pp\. 6\.External Links:[Document](https://dx.doi.org/10.3390/psf2021003006)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p1.1)\.
- \[48\]S\. Ranftl\(2022\)A connection between probability, physics and neural networks\.InPhysical Sciences Forum,Vol\.5,pp\. 11\.External Links:[Document](https://dx.doi.org/10.3390/psf2022005011)Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p2.1)\.
- \[49\]W\. Retagne, J\. Dauer, and G\. Waxenegger\-Wilfing\(2024\)Adaptive satellite attitude control for varying masses using deep reinforcement learning\.Frontiers in Robotics and AI11,pp\. 1402846\.External Links:[Document](https://dx.doi.org/10.3389/frobt.2024.1402846)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p1.1)\.
- \[50\]F\. M\. Rohrhofer, S\. Posch, C\. Gößnitzer, and B\. C\. Geiger\(2023\)Data vs\. physics: the apparent pareto front of physics\-informed neural networks\.IEEE Access11,pp\. 86252–86261\.External Links:[Document](https://dx.doi.org/10.1109/ACCESS.2023.3302892)Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p3.1)\.
- \[51\]F\. Salmaso, M\. Trisolini, and C\. Colombo\(2023\)A machine learning and feature engineering approach for the prediction of the uncontrolled re\-entry of space objects\.Aerospace10\(3\),pp\. 297\.External Links:[Document](https://dx.doi.org/10.3390/aerospace10030297)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p2.1),[§6\.2](https://arxiv.org/html/2611.21933#S6.SS2.p1.1)\.
- \[52\]E\. Schiassi, A\. D’Ambrosio, K\. Drozd, F\. Curti, and R\. Furfaro\(2022\)Physics\-informed neural networks for optimal planar orbit transfers\.Journal of Spacecraft and Rockets59\(3\),pp\. 834–849\.Cited by:[§2\.3](https://arxiv.org/html/2611.21933#S2.SS3.p1.1)\.
- \[53\]J\. Seo\(2024\)Solving real\-world optimization tasks using physics\-informed neural computing\.Scientific Reports14\(1\),pp\. 202\.External Links:[Document](https://dx.doi.org/10.1038/s41598-023-49977-3)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p3.1),[§2\.1](https://arxiv.org/html/2611.21933#S2.SS1.p1.1),[§5\.1](https://arxiv.org/html/2611.21933#S5.SS1.p1.1)\.
- \[54\]H\. Sheng and C\. Yang\(2021\)PFNN: a penalty\-free neural network method for solving a class of second\-order boundary\-value problems on complex geometries\.Journal of Computational Physics428,pp\. 110085\.External Links:ISSN 0021\-9991,[Document](https://dx.doi.org/10.1016/j.jcp.2020.110085)Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p3.1)\.
- \[55\]C\. Straub, P\. Brendel, V\. Medvedev, and A\. Rosskopf\(2025\)Hard\-constraining neumann boundary conditions in physics\-informed neural networks via fourier feature embeddings\.InICLR 2025 Workshop on Machine Learning Multiscale Processes,Note:arXiv: 2504\.01093External Links:[Link](https://openreview.net/forum?id=jKdZsWdRLZ)Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p3.1)\.
- \[56\]N\. Sukumar and A\. Srivastava\(2022\)Exact imposition of boundary conditions with distance functions in physics\-informed deep neural networks\.Computer Methods in Applied Mechanics and Engineering389,pp\. 114333\.External Links:ISSN 0045\-7825,[Document](https://dx.doi.org/10.1016/j.cma.2021.114333)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p5.1),[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p3.1),[§4](https://arxiv.org/html/2611.21933#S4.p1.3)\.
- \[57\]C\. J\. Sullivan, N\. Bosanac, and R\. L\. Anderson\(2023\)Designing low\-thrust transfers near earth–moon l2 via multi\-objective reinforcement learning\.Journal of Spacecraft and Rockets60\(2\),pp\. 634–647\.External Links:[Document](https://dx.doi.org/10.2514/1.A35463)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p2.1),[§6\.2](https://arxiv.org/html/2611.21933#S6.SS2.p1.1)\.
- \[58\]N\. Thuerey, K\. Weißenow, L\. Prantl, and X\. Hu\(2020\)Deep learning methods for reynolds\-averaged navier–stokes simulations of airfoil flows\.AIAA Journal58\(1\),pp\. 25–36\.External Links:[Document](https://dx.doi.org/10.2514/1.J058291)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p1.1)\.
- \[59\]B\. van der Heijden, X\. Li, G\. Lubineau, and E\. Florentin\(2025\)Enforcing physics onto pinns for more accurate inhomogeneous material identification\.Computer Methods in Applied Mechanics and Engineering441,pp\. 117993\.External Links:[Document](https://dx.doi.org/10.1016/j.cma.2025.117993)Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p2.1)\.
- \[60\]S\. Wang, S\. Sankaran, and P\. Perdikaris\(2022\)Respecting causality is all you need for training physics\-informed neural networks\.Computer Methods in Applied Mechanics and Engineering421,pp\. 116813\.External Links:[Document](https://dx.doi.org/10.1016/j.cma.2024.116813)Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p1.1),[§3\.1](https://arxiv.org/html/2611.21933#S3.SS1.p6.1)\.
- \[61\]S\. Wang, S\. Sankaran, H\. Wang, and P\. Perdikaris\(2023\)An expert’s guide to training physics\-informed neural networks\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2308.08468)Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p1.1),[§3\.1](https://arxiv.org/html/2611.21933#S3.SS1.p6.1)\.
- \[62\]S\. Wang, Y\. Teng, and P\. Perdikaris\(2021\)Understanding and mitigating gradient flow pathologies in physics\-informed neural networks\.SIAM Journal on Scientific Computing43\(5\),pp\. A3055–A3081\.External Links:[Document](https://dx.doi.org/10.1137/20M1318043)Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p1.1),[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p3.1),[§3\.1](https://arxiv.org/html/2611.21933#S3.SS1.p6.1)\.
- \[63\]S\. Wang, X\. Yu, and P\. Perdikaris\(2022\)When and why pinns fail to train: a neural tangent kernel perspective\.Journal of Computational Physics449,pp\. 110768\.External Links:[Document](https://dx.doi.org/10.1016/j.jcp.2021.110768)Cited by:[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p1.1),[§2\.2](https://arxiv.org/html/2611.21933#S2.SS2.p2.1)\.
- \[64\]G\. Waxenegger\-Wilfing, K\. Dresia, J\. Deeken, and M\. Oschwald\(2021\)A reinforcement learning approach for transient control of liquid rocket engines\.IEEE Transactions on Aerospace and Electronic Systems57\(5\),pp\. 2938–2952\.External Links:[Document](https://dx.doi.org/10.1109/TAES.2021.3074134)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p1.1)\.
- \[65\]Y\. Xu, S\. Wandelt, X\. Sun, Y\. Yang, X\. Jin, S\. Karichery, and M\. Drwal\(2023\)Machine\-learning\-assisted optimization of aircraft trajectories under realistic constraints\.Journal of Guidance, Control, and Dynamics46\(9\),pp\. 1814–1825\.External Links:[Document](https://dx.doi.org/doi.org/10.2514/1.G007038)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p2.1)\.
- \[66\]S\. Yin, J\. Li, and L\. Cheng\(2020\)Low\-thrust spacecraft trajectory optimization via a dnn\-based method\.Advances in Space Research66\(7\),pp\. 1635–1646\.External Links:[Document](https://dx.doi.org/10.1016/j.asr.2020.05.046)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p2.1)\.
- \[67\]A\. Zavoli and L\. Federici\(2021\)Reinforcement learning for robust trajectory design of interplanetary missions\.Journal of Guidance, Control, and Dynamics44\(8\),pp\. 1440–1453\.External Links:[Document](https://dx.doi.org/10.2514/1.G005794)Cited by:[§1](https://arxiv.org/html/2611.21933#S1.p2.1)\.
- \[68\]C\. Zhang, Y\. Zhu, and L\. Yang\(2026\)Low\-thrust strategies in orbital pursuit\-evasion games with rendezvous constraints: a pinns\-based approach\.Proceedings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering240\(3\),pp\. 433–447\.External Links:[Document](https://dx.doi.org/10.1177/09544100251377497)Cited by:[1st item](https://arxiv.org/html/2611.21933#S2.I1.i1.p1.1),[§2\.3](https://arxiv.org/html/2611.21933#S2.SS3.p1.1),[§2\.3](https://arxiv.org/html/2611.21933#S2.SS3.p2.1)\.
- \[69\]X\. Zhong, S\. Liu, T\. Chen, and H\. Hu\(2025\)Multi\-objective trajectory planning for flexible spacecraft via physics\-informed neural network\.Aerospace Science and Technology,pp\. 110710\.External Links:[Document](https://dx.doi.org/10.1016/j.ast.2025.110710)Cited by:[§2\.3](https://arxiv.org/html/2611.21933#S2.SS3.p2.1)\.

Similar Articles

Curriculum Learning of Physics-Informed Neural Networks based on Spatial Correlation

arXiv cs.LG

This paper proposes a spatially correlated curriculum learning framework for Physics-Informed Neural Networks (PINNs) that improves training stability and solution accuracy by leveraging spatial correlations among subregions, addressing issues like high-dimensional non-convex loss landscapes and imbalanced multi-objective constraints.