Trackformer

Overview

• Causal by construction — the inputs end at the issue time: no JMA or JTWC forecast route, no future analysis, no observation from after the issue. Later data is used only to score a forecast, never to make one — which is what makes a hindcast on a 1979 storm mean something
• An ensemble mean, and an honest band — benchmark and live runs average 50 seeded perturbations of the past and issue-time inputs, not 50 separately trained networks; the historical backfill is one member per issue. The Lab publishes the mean rather than the members, so the shaded band on the explorer is 1.2's published mean track error at each lead — how far off it typically is — not a spread invented for the run on screen
• Intensity is the honest weak spot — central-pressure error against JMA is 12.84 hPa for 1.2 and 13.53 hPa for 1.1 on the 134 starts where both can be scored, but the paired 95% interval of [−2.68, +1.20] hPa includes zero, so no intensity gain is claimed. The maximum-wind number is an auxiliary output whose averaging period is not validated and whose USA-wind benchmark is worse than 1.1's, and 1.2 has no wind-radius head at all — so the explorer draws no radius rings for it
• 32,230 forecasts you can look up — six-hourly starts for every western-Pacific storm from 1996 on, plus first issues back to 1970, each saved with its inputs' hashes and served read-only by the Trackformer Weather Lab, alongside a live run every six hours for whatever JMA is tracking
• Live & honest — it overlays its forecast on the JMA official track for active storms and lets you scrub any past storm to forecast forward against what actually happened. Where 1.2 has no run the overlay falls back to Trackformer 1.1 and says so; where neither has one it draws nothing, rather than quietly substituting another model. Clearly flagged experimental — research output, not an operational warning system

Example forecasts

At a glance

Model

Trackformer 1.2 — coupled pressure-field tropical-cyclone forecaster

Released

29 Sep 2026 as trackformer-1.2; paper revised 1 Oct 2026

Author

Yu Yao-Hsing

Status

Research model, not operational

Type

Recurrent convolutional encoder–decoders with Transformer attention, rolled out autoregressively

Parameters

21,452,595 — one shared set of weights

Domain

Western North Pacific: basin grid 0–60°N, 100–180°E at 2.5°; regional grid ±15° around the storm at 0.25°

Inputs

Nine six-hourly analyses (−48 to 0 h) of sea-level pressure, 500 hPa height and the winds at 850/500/200 hPa; static geography; the observed centre, motion and issue-time intensity

Outputs

Track, central pressure, basin and regional sea-level pressure fields, and an auxiliary maximum wind at +6, +12, …, +120 h

Ensemble

Benchmark, showcase and live-ensemble runs average 50 seeded perturbations of the inputs; the historical archive and predict.py give single clean forecasts

Training data

Whole storms by year: 2000–2021 training, 2022–23 validation, 2024–25 test; labels from IBTrACS

Licence

MIT

Weights

weights.pt, 86.0 MB (85,997,315 bytes); the released checkpoint is from epoch 4

Weights SHA-256

db49f36e85a3766defc4c172746897a1f783705d1ce8e6f9dfb8e87ae1d902cb

Checkpoint SHA-256

f194a23d3f91ea76ad776dfad942fabd669eeae8b3fd665815463095367e9ee0 — the hash every saved forecast carries, and the one the Typhoon Tracks overlay checks

Dataset SHA-256

04fd0d30620137aad20173dcd59e22da2a6c20fae2f17de593e91a59517459c1 (manifest)

Paper

Trackformer 1.2: Coupled pressure-field forecasting (PDF, 10 pages)

How it works

1. Initialise from the past. Nine six-hourly analyses, −48 h to the issue time, of sea-level pressure, 500 hPa height and the winds at 850, 500 and 200 hPa, with static geography, the storm's observed centre, recent motion, current intensity and validity masks. Finer regional pressure detail is used where it exists; where it does not, it stays masked rather than filled in.
2. Evolve the environment. A convolutional recurrent network advances the basin's pressure by steering-based advection plus learned pressure tendencies, with multiscale spatial tokens and attention conditioning its memory.
3. Move and evolve the core. A storm-centred pressure anomaly is carried relative to that environment and updated by its own learned tendencies; basin pressure plus the tapered core anomaly is the spatial forecast.
4. Read out the storm. The centre is the local low within 300 km and the central pressure is sampled there, so the route and the pressure forecast cannot disagree. A separate auxiliary head gives a single maximum-wind number — not a wind map.
5. Roll forward. Twenty autoregressive six-hour steps give +6, +12, …, +120 h. Future truth is a training target, never an inference input.

Under the hood, from the paper

• Two encoder–decoders. The basin network runs at widths 72/144/288/432 with depths 2/2/4/4 on the 25 × 33 grid; the core network runs at 64/128/256/384 on a 65 × 65 frame at 20 km spacing, ±640 km around the storm. Each keeps a 32-channel convolutional GRU memory between steps.
• Attention for context. The basin is pooled into 134 multiscale tokens (width 64, 4 heads, 2 pre-norm Transformer layers), and all 825 basin cells cross-attend to them, so a ridge or trough far from the storm can condition its motion.
• Steering transport. The steering flow is a learned blend of the 850, 500 and 200 hPa winds, starting at weights 0.4/0.4/0.2, plus a bounded learned correction. It advects the basin pressure, and the next step's motion proposal is 70% of that steering displacement (wind in m/s × 21.6 gives km per six hours) plus 30% of the previous step's motion, plus a learned correction bounded at ±100 km.
• Pressure-associated readout. The new centre is a soft minimum of pressure within 300 km of the proposal, weighted by a 180 km Gaussian; central pressure is sampled at that centre, not taken as the grid minimum.
• Geographic reconstruction. The map is the basin field plus the tapered core anomaly, registered by position. Once a storm leaves the ±15° regional patch — at +66 to +90 h in the films below — only the 2.5° basin field covers it.

Inputs and outputs

The model reads one causal issue packet and writes one forecast, both NumPy .npz files (pickled content is refused). Every input ends at the issue time. The eight channels are sea-level pressure (hPa), 500 hPa height (m) and the u/v winds at 850, 500 and 200 hPa (m/s), each normalised by a mean and standard deviation fitted on the training split alone — sea-level pressure, for example, by 8.27 hPa. The full contract is in the model README.

Training

• Split by whole storm and year, so no storm is in two splits: 13,949 training windows from 2000–2021, 1,041 validation windows from 2022–23 and 1,195 test windows from 2024–25; 800, 100 and 100 of them carry native regional pressure detail. No separate scores on the test split are published — the evaluation below is the released benchmark.
• Normalisation is a per-channel mean and standard deviation fitted on the training split only (through 2021).
• Objective: a weighted sum of the basin-field and regional-field errors, the route (weight 0.2), central pressure (0.5), auxiliary wind (0.05), and the six-hour changes in pressure (0.25) and position (0.1). The innermost 150 km is left out of regional-field supervision. Future observations are targets only.
• Ensemble: 50 members from seeded, smoothed perturbations of the past and issue-time inputs — not 50 separately trained networks — averaged with equal weight. Noise is 0.025 standard deviations on the basin fields and 0.015 on the regional field, smoothed over 5 and 9 cells, with each member's seed fixed so every forecast can be replayed exactly. The spread is not calibrated uncertainty.
• Labels: IBTrACS positions and central pressure, from both the JMA (TOKYO_PRES) and USA (USA_PRES) fields.

Evaluation

The released daily benchmark scores 1.2 against Trackformer 1.1 on 1,473 daily forecasts of 270 western-Pacific storms — 230 from 1980–1999 and 40 from 2024 on — with the same starts, valid times and leads for both. Each lead's error is averaged per day and each storm is weighted equally. 1.2 is the mean of 50 members. The cohort lies outside 1.2's fitting and validation years, but it is a development comparison, not a certified untouched holdout.

• Wind is worse than 1.1: against USA one-minute wind on the same 134 starts, the auxiliary wind's error is 25.03 kt for 1.2 against 16.98 kt for 1.1, a paired gap of +8.05 kt [5.02, 11.31].
• Single cases: the four showcase forecasts are scored one by one under Example forecasts below; the paper works through Fung-wong (2025) in detail.
• Not yet compared: a matched run against DeepMind's WeatherNext Cyclones Mini is pending, so no score is claimed.
• Read the protocol: evaluation limits · daily benchmark · intensity · every daily score · DeepMind protocol

The forecast archive

• 32,230 historical forecasts: a start every six hours for western-Pacific storms from 1996 on (30,677 of them), plus each storm's first issue back to 1970. All are single members, each saved with the hash of its exact input tensor.
• Inputs by era: NCEP/NCAR Reanalysis 1 analyses until NOAA ended it on 17 March 2026, archived GFS f000 analyses after that, and for live runs the latest GFS f000 analysis with JMA's own analysis of the storm.
• Live: typhoon-predict runs 1.2 on GitHub-hosted CPUs every six hours for each storm JMA tracks; a 50-member live ensemble takes about 68 s. Completion counts and source errors are public in the status feed.
• Still filling in: the learned auxiliary wind and the recovered pressure cores are exported for the archive separately, so some saved forecasts carry pressure but no wind yet — missing, never zero. The panel below follows that backfill live.

Run it, or read its forecasts

With PyTorch and NumPy, on a causal issue packet you prepare to the input contract (the release has no bundled live downloader):

git clone https://github.com/yu314-coder/typhoon-predict
cd typhoon-predict
curl -L -o models/trackformer_1_2_field/weights.pt \
  https://huggingface.co/euler314/typhoon-predict/resolve/main/models/trackformer_1_2_field/weights.pt
python models/trackformer_1_2_field/predict.py causal_issue_packet.npz forecast.npz --device cpu   # or mps / cuda

Or read the 32,230 archived forecasts, plus the benchmark and live runs, from any web page — the Trackformer Weather Lab's API is public, read-only and CORS-enabled (documentation):

const base = "https://trackformer-weatherlab.rudin-euler-8253.chatgpt.site";
const storm = await (await fetch(base + "/api/history/v1/storms/2013306N07162")).json();   // Haiyan 2013
const fc = await (await fetch(base + "/api/history/v1/forecasts/" + storm.issues[0].id)).json();
// fc.route: lat, lon, pressure_hpa at lead_hours 0..120; fc.members, fc.kind, fc.source_note

Release package: trackformer_1_2_field_20260929.tar.gz (89.9 MB) · manifest with every hash

The paper

@misc{yu2026trackformer,
  author       = {Yu, Yao-Hsing},
  title        = {Trackformer 1.2: Coupled Pressure-Field Forecasting},
  howpublished = {Research technical report, revised 1 October 2026},
  year         = {2026},
  url          = {https://github.com/yu314-coder/typhoon-predict}
}

Versions