TC-Next: Zero-Shot Multimodal Cyclone Forecasting

# TC-Next: Zero-Shot Multimodal Cyclone Forecasting
Source: [https://arxiv.org/html/2609.02085](https://arxiv.org/html/2609.02085)
Zhe Wang††thanks:Equal contribution\.Affiliation:Carnegie Mellon UniversityAffiliation:Pittsburgh, PA 15213Email:[zhew2@andrew\.cmu\.edu](mailto:)Sijie Chen11footnotemark:1Affiliation:Carnegie Mellon UniversityAffiliation:Pittsburgh, PA 15213Email:[sijieche@andrew\.cmu\.edu](mailto:)Yiming LuoAffiliation:Carnegie Mellon UniversityAffiliation:Pittsburgh, PA 15213Email:[yimingl1@andrew\.cmu\.edu](mailto:)Daehyun KimChien\-Yi ChangAffiliation:Durham UniversityAffiliation:Durham, United KingdomEmail:[chien\-yi\.chang@durham\.ac\.uk](mailto:)

###### Abstract

We present TropicalCycloneNext \(TC\-Next\), a multimodal deep learning model that forecasts tropical cyclone track and intensity at66–2424h leads by leveraging a foundation model’s forecast fields of atmospheric kinematic and thermodynamic fields and GridSat infrared satellite imagery\. Trained only on GraphCast forecasts over the Western Pacific \(WP\), yet reliant only on generic atmospheric variables, TC\-Next on GraphCast lowers track error by1515–44%44\\%and intensity error by a factor of33–66relative to a conventional, rule\-based tracker, TempestExtremes; applied without retraining to the forecast fields of Pangu\-Weather and IFS HRES, it stays ahead of TempestExtremes on both\. Applied zero\-shot to the generic weather fields of WeatherNext Cyclones on the 2025 WP season, TC\-Next attains lower intensity error at every lead time, and lower or comparable track error, compared to that model’s specialized direct tracker in a deterministic comparison\. Our ablation studies show that our multimodal model is able to utilize the additional modality to improve performance in tracking errors at every lead time and in intensity prediction at longer lead times\.

## 1Introduction

Tropical cyclones \(TCs\) are among the deadliest climate hazards, a threat accelerating under a warming climate\[[1](https://arxiv.org/html/2609.02085#bib.bib4),[2](https://arxiv.org/html/2609.02085#bib.bib5)\]\. Mass evacuations depend on accurate track and intensity forecasts, such as those that evacuated over two million people ahead of the145145kt Super Typhoon Ragasa in 2025\[[3](https://arxiv.org/html/2609.02085#bib.bib6)\]\. Improving them is therefore a concrete, high\-leverage form of climate adaptation\. Recent advances in deep learning have produced weather foundation models, including deterministic ones such as GraphCast\[[4](https://arxiv.org/html/2609.02085#bib.bib7)\], Pangu\-Weather\[[5](https://arxiv.org/html/2609.02085#bib.bib8)\], and Aurora\[[6](https://arxiv.org/html/2609.02085#bib.bib11)\], as well as probabilistic ones such as GenCast\[[7](https://arxiv.org/html/2609.02085#bib.bib9)\]and FGN\[[8](https://arxiv.org/html/2609.02085#bib.bib10)\], that forecast the global atmospheric states, matching or outperforming operational numerical weather prediction \(NWP\) while running orders of magnitude faster\. While they reproduce TC tracks well\[[4](https://arxiv.org/html/2609.02085#bib.bib7),[7](https://arxiv.org/html/2609.02085#bib.bib9)\], they struggle with intensity\. When intensity is read off their fields using TC trackers like TempestExtremes\[[9](https://arxiv.org/html/2609.02085#bib.bib17)\], the results are no better than persistence at short leads\[[10](https://arxiv.org/html/2609.02085#bib.bib18)\]\. Furthermore, the ERA5 reanalysis\[[11](https://arxiv.org/html/2609.02085#bib.bib3)\]they are trained on carries a1212–1515m/s intensity error against best\-track data\[[12](https://arxiv.org/html/2609.02085#bib.bib1)\]due to resolution constraints and assimilation processes\[[13](https://arxiv.org/html/2609.02085#bib.bib15)\]\.

Prior work addresses this issue in three main ways\. First, WeatherNext Cyclones \(WN\-C\)\[[14](https://arxiv.org/html/2609.02085#bib.bib12)\]extends the underlying weather model with TC\-specific feature channels trained using best\-track labels and extracts track and intensity through a specialized heuristic direct tracker\. This attains state\-of\-the\-art intensity, but at the cost of leaving the tracker and the TC head locked to the specialized model\. Second, Fuxi\-TC downscales the atmospheric variables resolution to resolve the storm’s tight inner core\[[13](https://arxiv.org/html/2609.02085#bib.bib15)\]by using a diffusion model to map coarse 0\.25° FuXi forecasts to 0\.1° WRF\[[15](https://arxiv.org/html/2609.02085#bib.bib25)\]\-simulated fields\. However, its intensity skill is below its teacher model, WRF\-0\.1\. Third, a number of end\-to\-end models that do not leverage foundation models have been proposed in recent years\[[16](https://arxiv.org/html/2609.02085#bib.bib13),[17](https://arxiv.org/html/2609.02085#bib.bib16),[18](https://arxiv.org/html/2609.02085#bib.bib14), e\.g\.,\]\. Although these models are relatively successful at intensity prediction, they cannot match the track skill achieved by foundation models\. Furthermore, no existing approach achieves zero\-shot transfer across diverse forecast models while also leveraging satellite observations of the storm core\.

We introduceTropicalCycloneNext\(TC\-Next\), a multimodal deep learning model designed to predict TC track and intensity around Western Pacific \(WP\) basin\. TC\-Next uses a foundation model’s forecast fields, which supply the large\-scale steering flow and thermodynamic environmental conditions that are known affect TC intensity, together with high\-resolution \(0\.07°\) GridSat satellite imagery\[[19](https://arxiv.org/html/2609.02085#bib.bib2)\]of the storm’s recent history, which provides convective structures associated with the storm\. Our contributions are three\-fold:1\. State\-of\-the\-Art Zero\-Shot Transfer:111We define*zero\-shot transfer*as replacing the training\-time forecast sourceF=GraphCastF=\\mathrm\{GraphCast\}\(the publicly released0\.25∘0\.25^\{\\circ\},3737\-level checkpoint trained on ERA5 1979–2017\) with an unseen sourceF′F^\{\\prime\}at inference without retraining or modifying any other input\. The only exception is the comparison against WN\-C in Table[1](https://arxiv.org/html/2609.02085#S3.T1): there, IFS HRES analysis frames replace the ERA5 history frames to match the analysis distribution WN\-C is fine\-tuned on, and the TC Vitals block further replaces the IBTrACS anchors with real\-time TC Vitals estimates\.We conduct comprehensive zero\-shot out\-of\-distribution \(OOD\) evaluations across both AI\-based and numerical forecasting models\. Despite being trained solely on GraphCast, TC\-Next achieved state\-of\-the\-art transfer performance: it outperformed the specialized direct tracker of the current SOTA \(WN\-C\) in a deterministic setup \(Table[1](https://arxiv.org/html/2609.02085#S3.T1)\), and consistently surpassed the TempestExtremes baseline in both track and intensity errors across multiple AI foundation models as well as the HRES numerical model at all lead times, lowering track error by1515–49%49\\%and intensity error by a factor of22–66\(Table[2](https://arxiv.org/html/2609.02085#S3.T2)\)\.2\. Multimodal Efficacy:Ablation studies on the satellite infrared \(IR\) branch demonstrate that multimodal inputs lower track error at every lead time and improve intensity error from\+18\+18h onward\. The benefit increases with lead time, peaking at\+24\+24h \(Table[3](https://arxiv.org/html/2609.02085#S3.T3)\)\.3\. Cross\-Model Portability:TC\-Next requires from the forecast source only generic atmospheric variables \(Table[5](https://arxiv.org/html/2609.02085#A1.T5); SST is persisted from the analysis\)\. Hence, the same trained model can be applied across forecast sources, providing a common learned tracker for evaluating TC forecast skill\.

## 2Methodology

### Problem formulation\.

We forecast TC track and intensity at lead timesℒ=\{6,12,18,24\}​h\\mathcal\{L\}=\\\{6,12,18,24\\\}\\,\\mathrm\{h\}\. Given observations through timett, a networkfθf\_\{\\theta\}emits, for each leadℓ\\ell, raw center and extent updates𝐚ℓ,𝐬ℓ∈ℝ2\\mathbf\{a\}\_\{\\ell\},\\mathbf\{s\}\_\{\\ell\}\\in\\mathbb\{R\}^\{2\}and an intensity incrementuℓ∈ℝu\_\{\\ell\}\\in\\mathbb\{R\}:

\{\(𝐚ℓ,𝐬ℓ,uℓ\)\}ℓ∈ℒ=fθ\(Et−24:t,St−24:t,E^t\+6:t\+24F,𝐜t−24:t,𝐯t−24:t\)\.\\big\\\{\(\\mathbf\{a\}\_\{\\ell\},\\mathbf\{s\}\_\{\\ell\},u\_\{\\ell\}\)\\big\\\}\_\{\\ell\\in\\mathcal\{L\}\}=f\_\{\\theta\}\\\!\\left\(E\_\{t\-24:t\},S\_\{t\-24:t\},\\hat\{E\}^\{F\}\_\{t\+6:t\+24\},\\mathbf\{c\}\_\{t\-24:t\},\\mathbf\{v\}\_\{t\-24:t\}\\right\)\.\(1\)The center and intensity updates are applied autoregressively from timett:

𝐜^t\+ℓ=𝐜^t\+ℓ−6\+δ​tanh⁡\(𝐚ℓ\),v^t\+ℓ=v^t\+ℓ−6\+uℓ,\\hat\{\\mathbf\{c\}\}\_\{t\+\\ell\}=\\hat\{\\mathbf\{c\}\}\_\{t\+\\ell\-6\}\+\\delta\\tanh\(\\mathbf\{a\}\_\{\\ell\}\),\\qquad\\hat\{v\}\_\{t\+\\ell\}=\\hat\{v\}\_\{t\+\\ell\-6\}\+u\_\{\\ell\},\(2\)whereδ=4∘\\delta=4^\{\\circ\}bounds the motion per step along each axis;𝐜^t\+ℓ\\hat\{\\mathbf\{c\}\}\_\{t\+\\ell\}andv^t\+ℓ\\hat\{v\}\_\{t\+\\ell\}are the predicted center and intensity att\+ℓt\+\\ell\. Here,Et−24:t∈ℝ5×H×W×CE\_\{t\-24:t\}\\in\\mathbb\{R\}^\{5\\times H\\times W\\times C\}holds five historical atmospheric frames withH×W=280×420H\\times W=280\\times 420grid cells \(a70∘×105∘70^\{\\circ\}\\times 105^\{\\circ\}domain at0\.25∘0\.25^\{\\circ\}\) andC=12C=12channels \(Table[5](https://arxiv.org/html/2609.02085#A1.T5)\);St−24:t∈ℝ5×256×256×2S\_\{t\-24:t\}\\in\\mathbb\{R\}^\{5\\times 256\\times 256\\times 2\}holds five IR frames, each a16∘16^\{\\circ\}window of brightness temperature and its validity mask;𝐜t−24:t∈ℝ5×2\\mathbf\{c\}\_\{t\-24:t\}\\in\\mathbb\{R\}^\{5\\times 2\}denotes historical centers; and𝐯t−24:t∈ℝ5\\mathbf\{v\}\_\{t\-24:t\}\\in\\mathbb\{R\}^\{5\}is the historical intensity\. The four future atmospheric framesE^Ft\+6:t\+24∈ℝ4×H×W×C\\hat\{E\}^\{F\}\_\{t\+6:t\+24\}\\in\\mathbb\{R\}^\{4\\times H\\times W\\times C\}are forecasts from a weather foundation modelFFinitialized attt\. The extent𝐬ℓ\\mathbf\{s\}\_\{\\ell\}updates an auxiliary box around the storm \(fromUSA\_R34wind radius\) that lets the loss judge a position miss by both distance and storm size; see Eq\. \([3](https://arxiv.org/html/2609.02085#S2.E3)\) for its update rule\. The history window has is fixed at five steps, but past fixes are optional: only the anchor fix\(𝐜t,vt\)\(\\mathbf\{c\}\_\{t\},v\_\{t\}\)must be observed, so a forecast can issue from a storm’s first fix\.222History steps preceding the storm’s first record are zero\-filled and masked, in training and at inference alike\.

### Data\.

To train TC\-Next, we built a dataset combining atmospheric fields—ERA5 reanalysis for the storm’s observed history and GraphCast forecasts for the lead times ahead—with IR brightness temperatures near11​μ​m11\\,\\mu\\mathrm\{m\}from GridSat\-B1 and storm state from IBTrACS\. From each atmospheric source we crop a fixed, WP\-basin\-centered70∘×105∘70^\{\\circ\}\\times 105^\{\\circ\}domain;0\.07∘0\.07^\{\\circ\}IR is resampled to0\.0625∘0\.0625^\{\\circ\}and cropped to a16∘16^\{\\circ\}window centered on the storm’s observed position at each history step\. Storm centers follow IBTrACS’s best track, and intensity is the11\-minute maximum sustained wind \(USA\_WIND\)\.

A training sample is anchored to a single IBTrACS fix at timett\. Over the five66\-hourly steps ending attt, it gathers ERA5, IR, and TC positions and intensities, masking any steps that precede the storm’s first record\. It also includes the four GraphCast forecasts initialized attt\. As targets, we use the storm’s positions and intensities fromt\+6t\{\+\}6tot\+24t\{\+\}24h, masking any leads beyond its final record\. Every in\-domain fix that is followed by another in\-domain fix66h later is a sample; storms that never reach tropical\-storm strength \(peakUSA\_SSHS<0<0\) are dropped\. Details are in Appendix[A](https://arxiv.org/html/2609.02085#A1)\. We partition data chronologically to prevent temporal leakage: training \(1990–2017\), validation \(2018\), and testing \(2019–2020\), comprising 28,962, 1,171, and 1,815 samples from 801, 34, and 57 storms, respectively\. To avoid temporal overlap with WN\-C’s training data \(through 2024\), we evaluate against it on a separate 2025 test set of 867 samples from 29 storms\. At evaluation, we score track and intensity separately at each lead time: track error is the mean great\-circle distance \(km\), and intensity error is the mean absolute error \(kt\)\. Beyond GraphCast, we draw Pangu\-Weather\[[5](https://arxiv.org/html/2609.02085#bib.bib8)\]and IFS HRES\[[20](https://arxiv.org/html/2609.02085#bib.bib24)\]forecasts from the WeatherBench2 archive\[[21](https://arxiv.org/html/2609.02085#bib.bib20)\], and real\-time storm estimates from TC Vitals\[[22](https://arxiv.org/html/2609.02085#bib.bib19)\]\.

### Model\.

TC\-Next is a multimodal encoder\-decoder architecture with3\.63\.6M parameters, trainable on a TPU v4\-8 VM in less than 70 minutes\. Its two encoders process atmospheric fields and satellite imagery at their respective input resolutions\. A*macro*CNN maps each atmospheric frame to a2∘2^\{\\circ\}feature grid\. A finer\-stride*micro*CNN processes a16∘16^\{\\circ\}GridSat window centered on the observed storm position at each historical time step, together with its per\-pixel validity mask, producing features on a0\.25∘0\.25^\{\\circ\}grid\. The macro feature map goes into a region\-proposal network\[[23](https://arxiv.org/html/2609.02085#bib.bib21)\]\(RPN\) to predict a storm\-centered proposal box and the extent, supervised directly against the center and wind\-radii box at every history step, to give the macro CNN a gradient path that bypasses the recurrent encoder–decoder\. RoIAlign\[[24](https://arxiv.org/html/2609.02085#bib.bib22)\]then extracts geographically aligned10∘10^\{\\circ\}windows around the proposed center from both feature maps, and low\-rank multimodal fusion\[[25](https://arxiv.org/html/2609.02085#bib.bib23)\]combines the resulting representations; see Eq\. \([5](https://arxiv.org/html/2609.02085#A2.E5)\)\.

An encoder LSTM processes the five fused historical representations and initializes a decoder LSTM\. At each forecast leadℓ\\ell, the decoder extracts a10∘10^\{\\circ\}window from the macro\-encoded forecast frame att\+ℓt\+\\ell, centered on its previous predicted storm position\. Separate prediction heads produce the center and intensity updates of Eq\. \([2](https://arxiv.org/html/2609.02085#S2.E2)\) together with an extent update𝐬ℓ\\mathbf\{s\}\_\{\\ell\}that scales the predicted box extent𝐫^t\+ℓ∈ℝ2\\hat\{\\mathbf\{r\}\}\_\{t\+\\ell\}\\in\\mathbb\{R\}^\{2\}\(width and height\):

𝐫^t\+ℓ=clip\[2∘,10∘\]⁡\(𝐫^t\+ℓ−6⊙exp⁡\(tanh⁡\(𝐬ℓ\)/2\)\)\.\\hat\{\\mathbf\{r\}\}\_\{t\+\\ell\}=\\operatorname\{clip\}\_\{\[2^\{\\circ\},10^\{\\circ\}\]\}\\\!\\Big\(\\hat\{\\mathbf\{r\}\}\_\{t\+\\ell\-6\}\\odot\\exp\\\!\\big\(\\tanh\(\\mathbf\{s\}\_\{\\ell\}\)/2\\big\)\\Big\)\.\(3\)The decoder initializes𝐜^t\\hat\{\\mathbf\{c\}\}\_\{t\}and𝐫^t\\hat\{\\mathbf\{r\}\}\_\{t\}from the last RPN proposal attt, and setsv^t=vt\\hat\{v\}\_\{t\}=v\_\{t\}\. Each forecast therefore updates an existing estimate of storm position, extent, and intensity—the sense in which the forecast is anchored—the decoder then rolls forward autoregressively over the available forecast frames\. Architectural and training details, including the loss in Eq\. \([6](https://arxiv.org/html/2609.02085#A2.E6)\), appear in Appendix[B](https://arxiv.org/html/2609.02085#A2)\.

## 3Experiments and Results

We evaluate TC\-Next in three settings: \(1\) against the direct tracker of WN\-C on the WP 2025 season \(Table[1](https://arxiv.org/html/2609.02085#S3.T1)\); \(2\) against TempestExtremes\[[9](https://arxiv.org/html/2609.02085#bib.bib17)\], configured as in TCBench\[[10](https://arxiv.org/html/2609.02085#bib.bib18)\], on the pooled WP 2019–2020 seasons across three forecast sources \(Table[2](https://arxiv.org/html/2609.02085#S3.T2)\); and \(3\) with its satellite infrared branch ablated \(Table[3](https://arxiv.org/html/2609.02085#S3.T3)\)\. TC\-Next is trained solely on GraphCast fields; Pangu\-Weather, IFS HRES, and WN\-C are evaluated without retraining\. Baseline configuration, sample matching, pooling, and scoring conventions are given in Appendix[C](https://arxiv.org/html/2609.02085#A3)\.

Table 1:WP, 2025; TC\-Next against the direct tracker of WN\-C\. Errors as defined in Section[2](https://arxiv.org/html/2609.02085#S2.SS0.SSS0.Px2); bold marks the lower error in each column within an anchor block\. Scored samples: Table[6](https://arxiv.org/html/2609.02085#A3.T6)\.- aZero\-shot: applied to WN\-C without retraining, reading only WN\-C’s generic weather channels\.bThe direct tracker derives its position and intensity from WN\-C’s dedicated TC feature channels\.

### TC\-Next transfers zero\-shot to WN\-C and outperforms its direct tracker\.

Reading only generic weather channels, our frozen model outperforms the WN\-C direct tracker’s single\-member predictions on track and intensity\. WN\-C’s ensemble members are scored individually against best track and averaged, so Table[1](https://arxiv.org/html/2609.02085#S3.T1)reflects expected single\-member error, not ensemble\-mean error\. The IBTrACS block uses the same HRES atmospheric source and IBTrACS anchor convention used for WN\-C fine\-tuning: our track error is lower at every lead, and early\-lead intensity error is roughly halved\. The TC Vitals block\[[22](https://arxiv.org/html/2609.02085#bib.bib19)\]swaps the IBTrACS anchors for the operational, real\-time ones\. Here, track performance remains on par—leading at\+6\+6h and\+24\+24h, trailing by under0\.60\.6km in between—while our intensity advantage widens, holding between3\.943\.94and7\.297\.29kt against the direct tracker’s7\.957\.95to9\.499\.49kt across all leads\. These results suggest that the gains survive the shift to operational anchors; fine\-tuning directly on real\-time inputs \(TC Vitals and GOES\-R imagery\[[26](https://arxiv.org/html/2609.02085#bib.bib26)\]\) is a natural next step\.

### TC\-Next outperforms TempestExtremes across all forecast sources and lead times\.

On GraphCast, the source the model is trained on, track error is1515–44%44\\%lower than TempestExtremes’ \(Table[2](https://arxiv.org/html/2609.02085#S3.T2)\), and intensity error is3\.23\.2–6\.16\.1times lower\. The margin survives zero\-shot transfer: applied without retraining, TC\-Next leads TempestExtremes on track by1818–45%45\\%on Pangu\-Weather and by2222–49%49\\%on IFS HRES, and its intensity error stays below1111kt on every source at every lead, whereas TempestExtremes’ never falls below1515kt\. These results highlight that our learned model provides an alternative to rule\-based detection across forecast sources\.

Table 2:WP, 2019–2020; TC\-Next vs\. TempestExtremes\. Scored samples: Table[7](https://arxiv.org/html/2609.02085#A3.T7)\.- aSource: the model was trained on this forecast source\.bZero\-shot: applied without retraining\.
- †\\daggerPangu\-Weather used 2019 for validation, so that year is not strictly held out\.

Table 3:WP, 2019–2020; IR ablation study\. Scored samples: Table[8](https://arxiv.org/html/2609.02085#A3.T8)\.
### The satellite infrared branch lowers track error at every lead and intensity error at longer leads, with a benefit that grows with lead time\.

Table[3](https://arxiv.org/html/2609.02085#S3.T3)ablates the satellite infrared branch on the full test set: the*IR\-ablated*variant is a separately trained model with the IR input zeroed in training and evaluation, all else fixed\. The fed variant lowers track error at every lead, by0\.70\.7km at\+6\+6h rising to2\.22\.2km at\+24\+24h, and intensity error from\+18\+18h onward\. These results suggest that the benefit of multimodal input increases with lead time and is largest at\+24\+24h\.

## 4Conclusion

In this study, we introduce TC\-Next, a multimodal deep learning model that predicts TC track and intensity using a foundation weather model’s forecast fields and high\-resolution satellite imagery of the storm’s history\. Because the model’s forecast input relies exclusively on generic atmospheric variables, it can be applied across forecast sources without retraining\. Evaluated on our WP dataset, TC\-Next demonstrates robust zero\-shot transfer capabilities across diverse foundation models, numerical prediction systems, and real\-time input configurations, consistently outperforming traditional rule\-based detection methods\. Ongoing work focuses on expanding this dataset and extending the zero\-shot benchmark to additional TC basins toward a global benchmark\.

## Acknowledgments and Disclosure of Funding

We gratefully acknowledge the support of Google for providing computing resources via the TPU Research Cloud \(TRC\) program, which made the extensive training of our models possible\. This work contains modified Copernicus Climate Change Service information \(ERA5\); neither the European Commission nor ECMWF is responsible for any use of the Copernicus information it contains\. This work originated from a CMU course project\. We thank our original team member, Ziping He, for her initial contributions, and course staff Ralf Brown and Yonatan Bisk for their guidance\. DK was supported by Creative\-Pioneering Researchers Program through Seoul National University and the National Research Foundation of Korea\(NRF\) grant funded by the Korea government\(MSIT\) \(RS\-2025\-02310080\)\. CC was supported by the JustN0W Strategic Research Fund through Durham University and the project ELEVATE\-ProClima through British Academy \(EPG\\100778\)\.

## References

- \[1\]K\. Emanuel\(2005\)Increasing destructiveness of tropical cyclones over the past 30 years\.Nature436\(7051\),pp\. 686–688\.External Links:[Document](https://dx.doi.org/10.1038/nature03906)Cited by:[§1](https://arxiv.org/html/2609.02085#S1.p1.1)\.
- \[2\]T\. Knutson, S\. J\. Camargo, J\. C\. L\. Chan, K\. Emanuel, C\. Ho, J\. Kossin, M\. Mohapatra, M\. Satoh, M\. Sugi, K\. Walsh, and L\. Wu\(2019\)Tropical cyclones and climate change assessment: part I: detection and attribution\.Bulletin of the American Meteorological Society100\(10\),pp\. 1987–2007\.External Links:[Document](https://dx.doi.org/10.1175/BAMS-D-18-0189.1)Cited by:[§1](https://arxiv.org/html/2609.02085#S1.p1.1)\.
- \[3\]Guy Carpenter\(2025\)Post\-event report: 2025 western north pacific typhoon ragasa\.Technical reportGuy Carpenter & Company, LLC\.Note:[https://www\.guycarp\.com/insights/2025/10/post\-event\-report\-2025\-western\-north\-pacific\-typhoon\-rRagasa\.html](https://www.guycarp.com/insights/2025/10/post-event-report-2025-western-north-pacific-typhoon-rRagasa.html)Cited by:[§1](https://arxiv.org/html/2609.02085#S1.p1.1)\.
- \[4\]R\. Lam, A\. Sanchez\-Gonzalez, M\. Willson, P\. Wirnsberger, M\. Fortunato, F\. Alet, S\. Ravuri, T\. Ewalds, Z\. Eaton\-Rosen, W\. Hu, A\. Merose, S\. Hoyer, G\. Holland, O\. Vinyals, J\. Stott, A\. Pritzel, S\. Mohamed, and P\. Battaglia\(2023\)Learning skillful medium\-range global weather forecasting\.Science382\(6677\),pp\. 1416–1421\.Cited by:[§1](https://arxiv.org/html/2609.02085#S1.p1.1)\.
- \[5\]K\. Bi, L\. Xie, H\. Zhang, X\. Chen, X\. Gu, and Q\. Tian\(2023\)Accurate medium\-range global weather forecasting with 3D neural networks\.Nature619,pp\. 533–538\.Cited by:[§1](https://arxiv.org/html/2609.02085#S1.p1.1),[§2](https://arxiv.org/html/2609.02085#S2.SS0.SSS0.Px2.p2.1)\.
- \[6\]C\. Bodnar, W\. P\. Bruinsma, A\. Lucic, M\. Stanley, A\. Vaughan, J\. Brandstetter, P\. Garvan, M\. Riechert, J\. A\. Weyn, H\. Dong, J\. K\. Gupta, K\. Thambiratnam, A\. T\. Archibald, C\. Wu, E\. Heider, M\. Welling, R\. E\. Turner, and P\. Perdikaris\(2025\)A foundation model for the Earth system\.Nature641\.Note:Aurora; arXiv:2405\.13063Cited by:[§1](https://arxiv.org/html/2609.02085#S1.p1.1)\.
- \[7\]I\. Price, A\. Sanchez\-Gonzalez, F\. Alet, T\. R\. Andersson, A\. El\-Kadi, D\. Masters, T\. Ewalds, J\. Stott, S\. Mohamed, P\. Battaglia, R\. Lam, and M\. Willson\(2025\)Probabilistic weather forecasting with machine learning\.Nature\.Note:GenCast; arXiv:2312\.15796Cited by:[§1](https://arxiv.org/html/2609.02085#S1.p1.1)\.
- \[8\]F\. Alet, I\. Price, A\. El\-Kadi, D\. Masters, S\. Markou, T\. R\. Andersson, J\. Stott, R\. Lam, M\. Willson, A\. Sanchez\-Gonzalez, and P\. Battaglia\(2025\)Skillful joint probabilistic weather forecasting from marginals\.Note:FGN; arXiv:2506\.10772Cited by:[§1](https://arxiv.org/html/2609.02085#S1.p1.1)\.
- \[9\]P\. A\. Ullrich, C\. M\. Zarzycki, E\. E\. McClenny, M\. C\. Pinheiro, A\. M\. Stansfield, and K\. A\. Reed\(2021\)TempestExtremes v2\.1: a community framework for feature detection, tracking, and analysis in large datasets\.Geoscientific Model Development14\(8\),pp\. 5023–5048\.External Links:[Document](https://dx.doi.org/10.5194/gmd-14-5023-2021)Cited by:[Appendix C](https://arxiv.org/html/2609.02085#A3.SS0.SSS0.Px1.p1.1),[§1](https://arxiv.org/html/2609.02085#S1.p1.1),[§3](https://arxiv.org/html/2609.02085#S3.p1.1)\.
- \[10\]M\. Gomez, S\. Ganesh S\., M\. McGraw, F\. I\. Tam, I\. Azizi, S\. Darmon, M\. Feldmann, S\. Bourdin, L\. Poulain\-Auzéau, S\. J\. Camargo, J\. Lin, D\. Chavas, C\. Lee, R\. Gupta, A\. Jenney, and T\. Beucler\(2026\)TCBench: a benchmark for tropical cyclone track and intensity forecasting at the global scale\.Note:arXiv:2601\.23268Cited by:[Appendix C](https://arxiv.org/html/2609.02085#A3.SS0.SSS0.Px1.p1.1),[§1](https://arxiv.org/html/2609.02085#S1.p1.1),[§3](https://arxiv.org/html/2609.02085#S3.p1.1)\.
- \[11\]H\. Hersbach, B\. Bell, P\. Berrisford, S\. Hirahara, A\. Horányi, J\. Muñoz\-Sabater, J\. Nicolas, C\. Peubey, R\. Radu, D\. Schepers,et al\.\(2020\)The ERA5 global reanalysis\.Quarterly Journal of the Royal Meteorological Society146\(730\),pp\. 1999–2049\.External Links:[Document](https://dx.doi.org/10.1002/qj.3803)Cited by:[§1](https://arxiv.org/html/2609.02085#S1.p1.1)\.
- \[12\]K\. R\. Knapp, M\. C\. Kruk, D\. H\. Levinson, H\. J\. Diamond, and C\. J\. Neumann\(2010\)The international best track archive for climate stewardship \(IBTrACS\): unifying tropical cyclone data\.Bulletin of the American Meteorological Society91\(3\),pp\. 363–376\.External Links:[Document](https://dx.doi.org/10.1175/2009BAMS2755.1)Cited by:[§1](https://arxiv.org/html/2609.02085#S1.p1.1)\.
- \[13\]S\. Guo, L\. Chen, Y\. Zhao, Y\. Lin, Z\. Niu, X\. Zhang, Z\. Sun, X\. Zhong, and H\. Li\(2025\)FuXi\-TC: a generative framework integrating deep learning and physics\-based models for improved tropical cyclone forecasts\.Note:arXiv:2508\.16168Cited by:[§1](https://arxiv.org/html/2609.02085#S1.p1.1),[§1](https://arxiv.org/html/2609.02085#S1.p2.1)\.
- \[14\]F\. Alet, T\. R\. Andersson, I\. Price, S\. Markou, A\. El\-Kadi, D\. Masters, A\. Li, S\. Merchant, N\. Williams, G\. Thornton, K\. MacKay, O\. Graham, A\. Uddin, B\. Gaiarin, D\. Shah, E\. Kruse, W\. Hogsett, D\. Zelinsky, J\. Cangialosi, J\. Martinez, J\. Franklin, M\. DeMaria, K\. Musgrave, C\. L\. Bain, H\. Titley, J\. Stott, R\. Lam, A\. Bell, P\. Komarek, M\. Willson, A\. Sanchez\-Gonzalez, and P\. Battaglia\(2026\)Operational tropical cyclone forecasting with AI\.Nature\.Note:WeatherNext CyclonesExternal Links:[Document](https://dx.doi.org/10.1038/s41586-026-10953-2)Cited by:[§1](https://arxiv.org/html/2609.02085#S1.p2.1)\.
- \[15\]W\. C\. Skamarock, J\. B\. Klemp, J\. Dudhia, D\. O\. Gill, Z\. Liu, J\. Berner, W\. Wang, J\. G\. Powers, M\. G\. Duda, D\. M\. Barker, and X\. Huang\(2021\)A description of the advanced research WRF model version 4\.3\.Technical reportTechnical ReportNCAR/TN\-556\+STR,National Center for Atmospheric Research,Boulder, CO\.External Links:[Document](https://dx.doi.org/10.5065/1dfh-6p97)Cited by:[§1](https://arxiv.org/html/2609.02085#S1.p2.1)\.
- \[16\]L\. Boussioux, C\. Zeng, T\. Guénais, and D\. Bertsimas\(2022\)Hurricane forecasting: a novel multimodal machine learning framework\.Weather and Forecasting37\(6\),pp\. 817–831\.Cited by:[§1](https://arxiv.org/html/2609.02085#S1.p2.1)\.
- \[17\]R\. Hang, Z\. Xu, J\. Zhao, R\. Yu, L\. Cheng, and Q\. Liu\(2026\)CycloneMAE: a scalable multi\-task learning model for global tropical cyclone probabilistic forecasting\.Note:arXiv:2604\.12180Cited by:[§1](https://arxiv.org/html/2609.02085#S1.p2.1)\.
- \[18\]Z\. Lin, J\. Chu, and Y\. Ham\(2025\)Enhancing tropical cyclone track and intensity predictions with the OWZP\-Transformer model\.npj Artificial Intelligence\.External Links:[Document](https://dx.doi.org/10.1038/s44387-025-00037-3)Cited by:[§1](https://arxiv.org/html/2609.02085#S1.p2.1)\.
- \[19\]K\. R\. Knapp, S\. Ansari, C\. L\. Bain, M\. A\. Bourassa, M\. J\. Dickinson, C\. Funk, C\. N\. Helms, C\. C\. Hennon, C\. D\. Holmes, G\. J\. Huffman, J\. P\. Kossin, H\. Lee, A\. Loew, and G\. Magnusdottir\(2011\)Globally gridded satellite observations for climate studies\.Bulletin of the American Meteorological Society92\(7\),pp\. 893–907\.External Links:[Document](https://dx.doi.org/10.1175/2011BAMS3039.1)Cited by:[§1](https://arxiv.org/html/2609.02085#S1.p3.1)\.
- \[20\]ECMWF\(2024\)IFS documentation CY49R1 – part IV: physical processes\.Technical reportEuropean Centre for Medium\-Range Weather Forecasts,Reading, UK\.Note:[https://www\.ecmwf\.int/en/elibrary/81626\-ifs\-documentation\-cy49r1\-part\-iv\-physical\-processes](https://www.ecmwf.int/en/elibrary/81626-ifs-documentation-cy49r1-part-iv-physical-processes)Cited by:[§2](https://arxiv.org/html/2609.02085#S2.SS0.SSS0.Px2.p2.1)\.
- \[21\]S\. Rasp, S\. Hoyer, A\. Merose, I\. Langmore, P\. Battaglia, T\. Russell, A\. Sanchez\-Gonzalez, V\. Yang, R\. Carver, S\. Agrawal, M\. Chantry, Z\. Ben Bouallegue, P\. Dueben, C\. Bromberg, J\. Sisk, L\. Barrington, A\. Bell, and F\. Sha\(2024\)WeatherBench 2: a benchmark for the next generation of data\-driven global weather models\.Journal of Advances in Modeling Earth Systems16\(6\),pp\. e2023MS004019\.External Links:[Document](https://dx.doi.org/10.1029/2023MS004019)Cited by:[Appendix C](https://arxiv.org/html/2609.02085#A3.SS0.SSS0.Px2.p1.1),[§2](https://arxiv.org/html/2609.02085#S2.SS0.SSS0.Px2.p2.1)\.
- \[22\]S\. Trahan and L\. Sparling\(2012\)An analysis of NCEP tropical cyclone vitals and potential effects on forecasting models\.Weather and Forecasting27\(3\),pp\. 744–756\.External Links:[Document](https://dx.doi.org/10.1175/WAF-D-11-00063.1)Cited by:[Appendix C](https://arxiv.org/html/2609.02085#A3.SS0.SSS0.Px2.p1.1),[§2](https://arxiv.org/html/2609.02085#S2.SS0.SSS0.Px2.p2.1),[§3](https://arxiv.org/html/2609.02085#S3.SS0.SSS0.Px1.p1.1)\.
- \[23\]S\. Ren, K\. He, R\. Girshick, and J\. Sun\(2015\)Faster R\-CNN: towards real\-time object detection with region proposal networks\.InAdvances in Neural Information Processing Systems,Cited by:[Appendix B](https://arxiv.org/html/2609.02085#A2.SS0.SSS0.Px1.p2.1),[§2](https://arxiv.org/html/2609.02085#S2.SS0.SSS0.Px3.p1.1)\.
- \[24\]K\. He, G\. Gkioxari, P\. Dollár, and R\. Girshick\(2017\)Mask R\-CNN\.InIEEE International Conference on Computer Vision \(ICCV\),Cited by:[Appendix B](https://arxiv.org/html/2609.02085#A2.SS0.SSS0.Px1.p2.1),[§2](https://arxiv.org/html/2609.02085#S2.SS0.SSS0.Px3.p1.1)\.
- \[25\]Z\. Liu, Y\. Shen, V\. B\. Lakshminarasimhan, P\. P\. Liang, A\. Zadeh, and L\. Morency\(2018\)Efficient low\-rank multimodal fusion with modality\-specific factors\.InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics \(ACL\),Cited by:[Appendix B](https://arxiv.org/html/2609.02085#A2.SS0.SSS0.Px1.p3.2),[§2](https://arxiv.org/html/2609.02085#S2.SS0.SSS0.Px3.p1.1)\.
- \[26\]T\. J\. Schmit, P\. Griffith, M\. M\. Gunshor, J\. M\. Daniels, S\. J\. Goodman, and W\. J\. Lebair\(2017\)A closer look at the ABI on the GOES\-R series\.Bulletin of the American Meteorological Society98\(4\),pp\. 681–698\.External Links:[Document](https://dx.doi.org/10.1175/BAMS-D-15-00230.1)Cited by:[§3](https://arxiv.org/html/2609.02085#S3.SS0.SSS0.Px1.p1.1)\.

## Appendix AData Details

Table[4](https://arxiv.org/html/2609.02085#A1.T4)gives the sample and storm counts per split and the rule behind them; Table[5](https://arxiv.org/html/2609.02085#A1.T5)lists the channels extracted per frame\.

### Data sources and preprocessing\.

GraphCast’s forecasts use the same0\.25∘0\.25^\{\\circ\}grid and the same variables as the ERA5 history frames; both are read into one array with no regridding\. GridSat\-B1 is the NOAA climate data record of intercalibrated geostationary imagery, native to a0\.07∘0\.07^\{\\circ\}grid at33\-h cadence; we take its single infrared\-window channel at the 00, 06, 12 and 18 UTC frames that coincide with the reanalysis cadence and resample it by bilinear interpolation\. Pixels without a valid source observation, and whole frames absent from the record \(4343in 1990–2018,2525in 2020\), are zero\-filled and carried as a per\-pixel validity mask that the model receives as a second input channel\. Each of the1212atmospheric channels is normalized by its mean and standard deviation, computed in one pass over all five frames of every one of the 28,962 training samples, the statistics are then fixed for validation, testing, and every forecast source\. Infrared brightness temperature is scaled by fixed constants,\(Tb−250​K\)/40​K\(T\_\{b\}\-250\\,\\mathrm\{K\}\)/40\\,\\mathrm\{K\}, and multiplied by the validity mask\.

The best\-track records are post\-season reanalyses for 1990–2020, whereas the 2025 records remain provisional\. Storms that never reach tropical\-storm strength are excluded, and fixes outside the domain are discarded\. Storm positions are kept in one coordinate system throughout: each best\-track position \(longitude wrapped to00–360∘360^\{\\circ\}, so the domain stays continuous across the antimeridian\) is mapped linearly onto the0\.25∘0\.25^\{\\circ\}latitude–longitude grid and expressed as a fractional \(row, column\) location in\[0,1\]2\[0,1\]^\{2\}of the domain zero\-padded to288×448288\\times 448cells \(72∘×112∘72^\{\\circ\}\\times 112^\{\\circ\}extent\)\. The model reads, proposes, and predicts positions in these units; multiplying by that extent converts them to the degrees used by the box loss, and forecasts are mapped back to latitude and longitude only for scoring\. A fix with no reported wind is masked out of the intensity term only\. The extent target is a square box centered on the best\-track position whose side is twice the mean of the four quadrant3434kt wind radii reported by JTWC \(IBTrACSUSA\_R34\), converted from nautical miles to degrees and clipped so that the side lies in\[2∘,10∘\]\[2^\{\\circ\},10^\{\\circ\}\]\. Where no radii are reported, the side is fixed at4∘4^\{\\circ\}\. When a storm’s record spans less than2424h at the anchor time, the unavailable history steps are masked rather than dropping the sample\.

### Array dimensions and storage\.

Atmospheric fields are cached as one file per initialization timett: five frames×280×420×12\\times\\,280\\times 420\\times 12channels in half precision \(about14\.114\.1MB\), holding the ERA5 analysis atttand the four GraphCast forecasts initialized attt\(leads66–2424h\)\. A sample anchored atttreads the analysis frame from the five files att−24,…,tt\{\-\}24,\\dots,tand the four forecast frames from the file attt\. Each analysis frame is therefore stored once and reused by every sample whose history window covers it—about1\.51\.5samples per file, with history windows straddling split boundaries—so we report the dataset as total files rather than per\-split counts\. GridSat infrared frames are stored separately, one per66\-hourly step \(1120×16801120\\times 1680at resampled0\.0625∘0\.0625^\{\\circ\}, compressed, about1\.71\.7MB each\)\.

### Dataset statistics\.

The WP atmospheric cache \(ERA5 and GraphCast\) comprises19,36019\{,\}360files for 1990–2018,770770for 2019 and491491for 2020, or20,62120\{,\}621files and271271GiB in total; the 2025 WN\-C comparison adds a further601601files \(7\.97\.9GiB\)\. The infrared imagery totals34\.734\.7GiB for 1990–2020 and1\.001\.00GiB for 2025\.

Table 4:Dataset summary\. The view is a fixed70∘×105∘70^\{\\circ\}\\times 105^\{\\circ\}domain at0\.25∘0\.25^\{\\circ\}resolution spanning5∘5^\{\\circ\}S–65∘65^\{\\circ\}N and95∘95^\{\\circ\}E–160∘160^\{\\circ\}W\. A sample is a66\-hourly fix inside the view followed by another fix66h later, also inside it; storm counts are of the storms those samples come from\. Because the view is a fixed geographic box rather than a basin mask, it also holds storms that IBTrACS codes to adjacent basins: 1,108 of the 28,962 training samples, most of them East Pacific\. The test set pools two held\-out seasons, 2019 \(1,069 samples, 32 storms\) and 2020 \(746 samples, 25 storms\)\. The 2025 counts come from the per\-basin fileIBTrACS\.WP\.v04r01\.nc\(snapshot of 2026\-08\-18\) and are snapshot\-dependent while the 2025 records remain provisional\.Train \(1990–2017\)Val \(2018\)Test \(2019–2020\)Test \(2025\)BasinSamplesStormsSamplesStormsSamplesStormsSamplesStormsWP28,9628011,171341,8155786729Table 5:Atmospheric channels extracted per frame from the analysis \(history\) and GraphCast \(forecast\)\. Only three pressure levels are used, chosen for the environmental controls on TC intensity:850850hPa \(low\-level circulation and moisture\),500500hPa \(mid\-tropospheric environmental flow and core ascent\), and200200hPa \(outflow; paired with850850hPa it yields the deep\-layer vertical wind shear\)\. Sea\-surface temperature is not forecast by GraphCast and is taken from the ERA5 analysis at initialization timettin every forecast frame\.

## Appendix BModel and Training Details

### Architecture\.

A*macro*CNN encodes each atmospheric frame with three stride\-22blocks of3232,6464, and9696channels, reducing the0\.25∘0\.25^\{\\circ\}input to a2∘2^\{\\circ\}feature map\. Its input is the1212atmospheric channels of Table[5](https://arxiv.org/html/2609.02085#A1.T5)plus four static planes: latitude, the sine and cosine of longitude, and the ERA5 land–sea mask\. A finer\-stride*micro*CNN processes the16∘16^\{\\circ\}GridSat window centered on the observed storm position at each history step, together with its per\-pixel validity mask, mapping the resampled0\.0625∘0\.0625^\{\\circ\}IR input to a0\.25∘0\.25^\{\\circ\}feature grid; the two maps then sit on scales RoIAlign can align, but the micro features are drawn from imagery four times finer than the environment fields\.

A region\-proposal network\[[23](https://arxiv.org/html/2609.02085#bib.bib21)\]reads the macro map alongside a Gaussian prior rendered at the storm’s known fix for that step and emits a box: a center offset from the prior, clamped to2\.5∘2\.5^\{\\circ\}, plus a side length in\[2∘,10∘\]\[2^\{\\circ\},10^\{\\circ\}\]\. The proposal decides where the model looks: RoIAlign\[[24](https://arxiv.org/html/2609.02085#bib.bib22)\]samples a fixed10∘10^\{\\circ\}window from both feature maps at the proposed center \(8×88\\times 8points on the macro map,16×1616\\times 16on the micro map\), and separate convolutional heads flatten each crop to20482048dimensions\.

Writezτm∈ℝ2048z^\{m\}\_\{\\tau\}\\in\\mathbb\{R\}^\{2048\}for the crop vector of streamm∈\{env,ir\}m\\in\\\{\\mathrm\{env\},\\mathrm\{ir\}\\\}at history stepτ\\tau\. Each is projected to128128dimensions, layer\-normalized with its own learned scale and offset, and extended by a constant:

z~τm=\[LNm​\(Pm​zτm\);1\]∈ℝ129\.\\tilde\{z\}^\{m\}\_\{\\tau\}=\\big\[\\mathrm\{LN\}\_\{m\}\(P\_\{m\}z^\{m\}\_\{\\tau\}\);\\,1\\big\]\\in\\mathbb\{R\}^\{129\}\.\(4\)The two streams are combined by low\-rank multimodal fusion\[[25](https://arxiv.org/html/2609.02085#bib.bib23)\],

Vτ=∑r=1Rwr​\(Wenv\(r\)​z~τenv\)⊙\(Wir\(r\)​z~τir\)\+b,V\_\{\\tau\}=\\sum\_\{r=1\}^\{R\}w\_\{r\}\\,\\big\(W^\{\(r\)\}\_\{\\mathrm\{env\}\}\\tilde\{z\}^\{\\mathrm\{env\}\}\_\{\\tau\}\\big\)\\odot\\big\(W^\{\(r\)\}\_\{\\mathrm\{ir\}\}\\tilde\{z\}^\{\\mathrm\{ir\}\}\_\{\\tau\}\\big\)\+b,\(5\)withR=8R=8and⊙\\odotthe element\-wise product\. The appended11retains the unimodal terms, so Eq\. \([5](https://arxiv.org/html/2609.02085#A2.E5)\) is strictly more expressive than passing the concatenation of the two streams through a linear layer\. The per\-stream normalization is load\-bearing: in a product each stream’s gradient is scaled by the other, so without it the infrared branch is starved by the much larger environment signal\.

The storm’s own scalars—the prior position, current intensity, availability flags, and a step index—are concatenated toVτV\_\{\\tau\}, then passes through a two\-layer MLP of width256256with ReLU, and an encoder LSTM of width256256reads the five resulting vectors in order; its final state initializes a decoder LSTM of the same width\. At each leadℓ\\ell, the decoder input is the20482048\-dimensional macro crop head applied to the RoIAlign window of the forecast frame att\+ℓt\+\\ell, taken at the previous predicted center, concatenated with that center, the previous extent \(divided by10∘10^\{\\circ\}\), the previous intensity, and a one\-hot lead index; a linear layer with ReLU maps this to256256dimensions before the LSTM\. Two linear heads read the LSTM output: a four\-dimensional box head giving𝐚ℓ\\mathbf\{a\}\_\{\\ell\}and𝐬ℓ\\mathbf\{s\}\_\{\\ell\}of Eqs\. \([2](https://arxiv.org/html/2609.02085#S2.E2)\) and \([3](https://arxiv.org/html/2609.02085#S2.E3)\), and a scalar head givinguℓu\_\{\\ell\}\. Both heads are initialized near zero, so the untrained model forecasts persistence\. The IR data enters the forecast only through the encoder state: the decoder reads the forecast environment fields alone, since no IR imagery exists for future times\.

### Training\.

The total loss has three terms of two kinds: a box loss, which scores position and extent, applied to both the forecast boxes and the history proposals; and an intensity loss, which scores maximum sustained wind\. Writingbℓ=\(𝐜ℓ,𝐫ℓ\)b\_\{\\ell\}=\(\\mathbf\{c\}\_\{\\ell\},\\mathbf\{r\}\_\{\\ell\}\)for the storm’s box in degrees, the extent target𝐫ℓ\\mathbf\{r\}\_\{\\ell\}being the best\-track gale\-force \(3434kt\) wind radius \(R34, the mean of the four reported quadrants; a fixed4∘4^\{\\circ\}where unreported\), training minimizes

ℒbox​\(𝒯F\)\+β​ℒint\+γ​ℒbox​\(𝒯H\),ℒbox\(𝒯\)=1\|𝒯\|∑ℓ∈𝒯\[λ∥b^ℓ−bℓ∥1\+1−GIoU\(b^ℓ,bℓ\)\],ℒint=1\|𝒯v\|∑ℓ∈𝒯v\(v^ℓ−vℓ\)2,\\begin\{array\}\[\]\{c\}\\mathcal\{L\}\_\{\\mathrm\{box\}\}\\big\(\\mathcal\{T\}\_\{F\}\\big\)\+\\beta\\,\\mathcal\{L\}\_\{\\mathrm\{int\}\}\+\\gamma\\,\\mathcal\{L\}\_\{\\mathrm\{box\}\}\\big\(\\mathcal\{T\}\_\{H\}\\big\),\\\\\[3\.0pt\] \\mathcal\{L\}\_\{\\mathrm\{box\}\}\(\\mathcal\{T\}\)=\\dfrac\{1\}\{\|\\mathcal\{T\}\|\}\\displaystyle\\sum\_\{\\ell\\in\\mathcal\{T\}\}\\Big\[\\lambda\\,\\\|\\hat\{b\}\_\{\\ell\}\-b\_\{\\ell\}\\\|\_\{1\}\+1\-\\mathrm\{GIoU\}\(\\hat\{b\}\_\{\\ell\},b\_\{\\ell\}\)\\Big\],\\qquad\\mathcal\{L\}\_\{\\mathrm\{int\}\}=\\dfrac\{1\}\{\|\\mathcal\{T\}\_\{v\}\|\}\\displaystyle\\sum\_\{\\ell\\in\\mathcal\{T\}\_\{v\}\}\(\\hat\{v\}\_\{\\ell\}\-v\_\{\\ell\}\)^\{2\},\\end\{array\}\(6\)withλ=0\.2\\lambda=0\.2andγ=0\.3\\gamma=0\.3\. Here𝒯F⊆\{t\+6,…,t\+24\}\\mathcal\{T\}\_\{F\}\\subseteq\\\{t\{\+\}6,\\dots,t\{\+\}24\\\}and𝒯H⊆\{t−24,…,t\}\\mathcal\{T\}\_\{H\}\\subseteq\\\{t\{\-\}24,\\dots,t\\\}are the forecast and history steps at which the storm has a best\-track fix, and𝒯v⊆𝒯F\\mathcal\{T\}\_\{v\}\\subseteq\\mathcal\{T\}\_\{F\}those at which wind is also reported, so a storm with a partial record contributes only its observed fixes\. Forℓ∈𝒯H\\ell\\in\\mathcal\{T\}\_\{H\},b^ℓ\\hat\{b\}\_\{\\ell\}is the region\-proposal box at that step\. The intensity weightβ\\betaopens at20%20\\%of its final value of2\.52\.5and reaches it after one epoch, which lets the network localize before intensity competes for the shared trunk\. We optimize with AdamW at a peak learning rate of3×10−43\\times 10^\{\-4\}, warmed up over300300steps and cosine\-decayed thereafter, with weight decay10−410^\{\-4\}, gradients clipped to global norm11, and a batch of3232sharded data\-parallel over a TPU v4\-8\. The model trains on the 1990–2017 split for1515epochs in about6565minutes; we keep the epoch with the best validation score on 2018, track error in km over100100plus intensity MAE in knots over1010, which selects epoch66\. No held\-out season enters selection at any point\. Under zero\-shot transfer the input normalization statistics remain those of the training source; none are re\-estimated\.

## Appendix CExperimental Setup Details

### TempestExtremes\.

For every forecast source we read the storm’s position and intensity directly off its forecast fields with TempestExtremes\[[9](https://arxiv.org/html/2609.02085#bib.bib17)\], using the detection parameters of TCBench\[[10](https://arxiv.org/html/2609.02085#bib.bib18)\]: a sea\-level\-pressure minimum is accepted only if the pressure rises by200200Pa within6\.5∘6\.5^\{\\circ\}great\-circle distance \(a closed low\) and the300300–500500hPa thickness falls by58\.858\.8m2s\-2within5\.5∘5\.5^\{\\circ\}\(a warm core\), with candidate nodes closer than6\.0∘6\.0^\{\\circ\}merged\. Intensity is measured at each node as the minimum sea\-level pressure at the node itself and the maximum1010m wind speed within2\.0∘2\.0^\{\\circ\}great\-circle distance\. The baseline receives the same anchor as TC\-Next\. Where no node falls inside the gate the tracker returns no forecast: it produces a forecast for5454–70%70\\%of samples depending on source, whereas TC\-Next covers all samples by construction\. All comparisons are therefore restricted to the samples both methods produce\. Two caveats apply\. First, the2\.0∘2\.0^\{\\circ\}wind radius is the default of the standard TempestExtremes recipe; all baseline intensities are conditional on this choice\. Second, a gridded wind maximum over a finite radius is not the same quantity as the11\-minute sustained wind of the best\-track record, so part of the baseline’s intensity error is definitional rather than forecast error\.

### Forecast sources\.

TC\-Next is trained on GraphCast forecast fields only, together with ERA5 history frames, GridSat IR frames, and best\-track anchors from IBTrACS\. Every other input appears at evaluation time alone: the model is applied without retraining to the forecast fields of Pangu\-Weather, IFS HRES, and WN\-C, with the Pangu\-Weather and IFS HRES fields taken from the WeatherBench2 archives\[[21](https://arxiv.org/html/2609.02085#bib.bib20)\]\. Pangu\-Weather emits no vertical velocity, so theω500\\omega\_\{500\}input channel is zero\-filled for its rows; TempestExtremes does not use this field and is unaffected\. Table[1](https://arxiv.org/html/2609.02085#S3.T1)likewise substitutes inputs at evaluation time only\. Throughout the table, IFS HRES analysis frames replace the ERA5 history, matching the analyses WN\-C is fine\-tuned on\. In the TC Vitals block, TC Vitals\[[22](https://arxiv.org/html/2609.02085#bib.bib19)\]—the real\-time storm estimates issued by the operational forecast centers—additionally replace the best\-track anchors; that configuration therefore measures zero\-shot robustness to the inputs available in real time, not the behavior of a retrained variant\.

### Scoring\.

Table[2](https://arxiv.org/html/2609.02085#S3.T2)pools the 2019 and 2020 seasons, each cell weighted by its own per\-season matched count, and uses the0000:0000and1212:0000UTC initialization times, the only times at which Pangu\-Weather and IFS HRES are archived\. Within each forecast source, all rows are scored on the samples that source’s tracker matched, so each comparison is homogeneous\. Table[7](https://arxiv.org/html/2609.02085#A3.T7)counts, under the Track columns, the forecast samples for which TempestExtremes found a storm near the anchored position at that lead time \(TC\-Next forecasts every sample with a best\-track target\); under the Intensity columns, the subset of those matched samples whose best\-track record includes the11\-minute wind\. In Tables[1](https://arxiv.org/html/2609.02085#S3.T1)and[2](https://arxiv.org/html/2609.02085#S3.T2), track and intensity are therefore scored on different samples: samples without a best\-track11\-minute wind \(USA\_WIND\) are dropped from intensity scoring only\. In Table[1](https://arxiv.org/html/2609.02085#S3.T1), WN\-C is fine\-tuned on IFS HRES analyses, so TC\-Next is given HRES analysis frames in place of the ERA5 history to keep it on the same analysis distribution\. The WN\-C tracker can predict that a storm has dissipated while the IBTrACS still reports it; those rows are dropped for both models, so the IBTrACS block scores fewer than the867867samples Table[4](https://arxiv.org/html/2609.02085#A1.T4)counts for 2025\. Table[6](https://arxiv.org/html/2609.02085#A3.T6)lists the scored samples per lead\. Each of WN\-C’s four ensemble members is scored separately and its errors averaged, rather than the ensemble being collapsed first: this measures the two models rather than the error cancellation an ensemble mean buys\. TC\-Next is trained against a deterministic backbone, and adapting it to an ensemble would require a different loss, which we leave to future work\. Table[3](https://arxiv.org/html/2609.02085#S3.T3)is scored on every held\-out sample at all synoptic times; Table[8](https://arxiv.org/html/2609.02085#A3.T8)lists those counts\.

Table 6:Scored samplesnnat each lead for Table[1](https://arxiv.org/html/2609.02085#S3.T1)\(WP, 2025; TC\-Next against the direct tracker of WN\-C\)\. Each row counts the samples on which both models of that anchor block are scored, namely those the WN\-C tracker produced: where the WN\-C tracker predicts that the storm has dissipated, the sample is dropped for both models\. Intensity counts are lower because they also need a reportedUSA\_WIND\.Table 7:Scored samplesnnat each lead for Table[2](https://arxiv.org/html/2609.02085#S3.T2)\(WP, pooled 2019–2020; TC\-Next against TempestExtremes\)\. Each row counts the samples on which both models of that forecast source are scored: the0000:0000and1212:0000UTC samples \(the only initialization times at which WeatherBench2 archives Pangu\-Weather and IFS HRES forecasts\) that have a best\-track target at that lead and for which TempestExtremes found the storm; TC\-Next forecasts every sample with a target, so the counts are limited only by TempestExtremes\. Intensity counts are lower because they also need a reportedUSA\_WIND\.Table 8:Scored samplesnnat each lead for Table[3](https://arxiv.org/html/2609.02085#S3.T3)\(WP, 2019–2020; ablation of the IR branch\)\. Both variants are scored on every held\-out sample at all synoptic times, so the counts exceed those of Table[7](https://arxiv.org/html/2609.02085#A3.T7)and the two tables are not directly comparable\. Intensity counts are lower because they also need a reportedUSA\_WIND\.