TC-Next: Zero-Shot Multimodal Cyclone Forecasting
Zhe Wang, Sijie Chen, Yiming Luo, Daehyun Kim, Chien-Yi Chang
cs.LG · cs.AI · cs.CV
Quick verdict
Claims a 3.6M-param multimodal tracker beating TempestExtremes and WN-C zero-shot across sources; results credible but lack CIs; ablation gains marginal.
Summary
TC-Next is a 3.6M-parameter multimodal encoder-decoder that forecasts Western Pacific tropical cyclone track and intensity at 6–24 h leads. It fuses generic atmospheric fields from a weather foundation model (GraphCast at training), ERA5 history frames, 0.07° GridSat infrared imagery of the storm's recent history, and best-track anchors, using a region-proposal network, RoIAlign crops, low-rank multimodal fusion, and autoregressive LSTMs that update position, extent, and intensity. Trained only on 1990–2017 GraphCast fields, it cuts track error 15–44% and intensity error by a factor of 3.2–6.1 versus the rule-based TempestExtremes tracker on pooled 2019–2020 WP seasons (Table 2). Without retraining it transfers to Pangu-Weather and IFS HRES (track error 18–45% and 22–49% lower) and to WeatherNext Cyclones' generic fields on the 2025 season, where it beats WN-C's specialized direct tracker on intensity at every lead, with track error lower or comparable, in a deterministic comparison (Table 1). A separately trained IR-ablated control shows the satellite branch lowers track error by 0.7–2.2 km at all leads but intensity only marginally from +18 h onward (0.06–0.15 kt). Main limitations: no confidence intervals or significance tests, a single basin with short test windows, a conditional baseline (TempestExtremes covers only 54–70% of samples, with partly definitional intensity error), and no code-release statement.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Trained only on 1990–2017 GraphCast fields over the Western Pacific, TC-Next lowers track error by 15–44% and intensity error by a factor of 3.2–6.1 versus the rule-based TempestExtremes tracker on pooled 2019–2020 WP test seasons (Table 2).
- Without retraining, TC-Next transfers to Pangu-Weather and IFS HRES fields (track error 18–45% and 22–49% lower than TempestExtremes) and to WeatherNext Cyclones' generic channels, beating WN-C's specialized direct tracker on intensity at every lead and matching or beating it on track in a deterministic 2025 comparison (Tables 1–2).
- The GridSat infrared branch, ablated via a separately trained control, lowers track error at every lead (by 0.7–2.2 km) and intensity error from +18 h onward, with a benefit that grows with lead time (Table 3).
- Because the model consumes only generic atmospheric variables plus satellite imagery and an anchor, a single trained tracker can serve as a common instrument for evaluating TC skill across heterogeneous forecast sources.
Strengths
Genuine cross-source zero-shot transfer
One model trained only on GraphCast fields is applied unchanged to Pangu-Weather, IFS HRES, and WN-C generic channels, staying ahead of TempestExtremes everywhere (Table 2: track 18–45% and 22–49% lower on Pangu/HRES) and beating WN-C's specialized direct tracker on intensity at every lead in Table 1's deterministic comparison.
Unusually transparent baseline accounting
Appendix C discloses TempestExtremes' 54–70% forecast coverage, restricts scoring to matched samples, flags the definitional mismatch between a 2.0° gridded wind maximum and 1-min sustained wind, scores WN-C ensemble members individually rather than collapsed, and adds an operational TC Vitals anchor block; Table 2 footnotes that Pangu's 2019 is not strictly held out.
Leakage discipline and honest ablation design
Cheap, fully specified method
Weaknesses
No uncertainty quantification anywhere
Conditional, partially definitional baseline
Anchor asymmetry and missing persistence baseline
Narrow evaluation scope and MAE-only reporting
No code or checkpoint release stated
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
TC-Next (arXiv:2609.02085) is a method paper proposing a 3.6M-parameter multimodal tracker for Western Pacific tropical cyclones at 6–24 h leads. Its claims: (1) trained only on GraphCast forecast fields (1990–2017), it beats the rule-based TempestExtremes (TE) tracker by 15–44% in track MAE and a factor of 3.2–6.1 in intensity MAE on pooled 2019–2020 WP seasons (Table 2); (2) without retraining it transfers to Pangu-Weather and IFS HRES (track error 18–45% and 22–49% lower than TE) and to the generic weather channels of WeatherNext Cyclones (WN-C), beating WN-C's specialized direct tracker on intensity at every lead, with track error lower or comparable, in a deterministic 2025-season comparison (Table 1); (3) the GridSat IR branch reduces track error at all leads and intensity error from +18 h onward (Table 3). The portability framing — one learned tracker, fed only generic atmospheric variables (Table 5), usable to compare forecast sources — is the paper's most durable contribution.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is TC-Next and what does it predict?
TC-Next is a 3.6-million-parameter multimodal neural network that forecasts the track (position) and intensity (1-minute maximum sustained wind) of Western Pacific tropical cyclones at 6, 12, 18, and 24 hour lead times, using atmospheric fields from AI weather models, infrared satellite images of the storm's recent history, and a latest observed storm fix as the starting anchor.
What does 'zero-shot' mean in this paper?
The model was trained only on GraphCast forecast fields (1990–2017), then applied completely unchanged to atmospheric fields from other sources — Pangu-Weather, ECMWF's IFS HRES, and Google's WeatherNext Cyclones — without any retraining, and still outperformed the rule-based TempestExtremes tracker on all of them. (For the WeatherNext comparison the paper does swap in HRES analysis history frames and real-time TC Vitals anchors, as disclosed in its footnote 2.)
How much better is TC-Next than existing trackers?
On pooled 2019–2020 Western Pacific test seasons, it cuts track error by 15–44% and intensity error by a factor of 3.2–6.1 versus TempestExtremes. Applied to WeatherNext Cyclones fields on the 2025 season, it achieves lower intensity error than that system's own specialized tracker at every lead time, with track error lower or comparable in a deterministic comparison. These are point estimates; the paper reports no confidence intervals or significance tests.
Does the satellite imagery actually help?
Modestly. In an ablation where the infrared branch was removed and the model retrained from scratch, removing it raises track error by 0.7–2.2 km at all leads, while the intensity effect appears only from +18 h onward and is tiny (0.06–0.15 kt). The paper reports these as improvements but provides no statistical significance testing, so the intensity benefit in particular is uncertain.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- No figure content appears in the provided extraction; any figures in the published PDF could not be reviewed (all quantitative claims were checkable against the extracted tables).
- Equation and subscript formatting is flattened by PDF extraction (Eqs. 1–6), but every numeric value quoted in this review was legible and checked against its table row/column label.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.