Intensity · central pressure
JMA central-pressure MAE · hPa
Pressure-based intensity. Track position. Track direction.
Track uses all 1,473 starts / 270 storms; pressure uses 134 matched starts / 40 storms. The other 1,339 starts lack the frozen 1.1 intensity inputs, not zero error. DeepMind Mini has not run on this daily cohort. Development evidence, not an untouched test.
JMA central-pressure MAE · hPa
Mean position error · km
Six-hour heading error · degrees
1.1 · frozen native ensemble1.2 · 50 distinct input-perturbed membersDeepMind Mini · not evaluated here, not zero
The mean JMA pressure error is slightly lower for 1.2, but its paired 95% whole-storm interval [-2.68, 1.20] hPa includes zero. No universal intensity improvement is established.
2024141N03142 · issue 2024-05-22 00:00 UTC. All 134 matched starts are selectable; the first example is not cherry-picked.
Same physical hPa scale and exact valid UTC; no shifting or stretching. Reference gaps stay unavailable, not zero. Hover the graph or move the slider to inspect each time.
| Lead / valid UTC | Observed JMA | 1.1 | 1.2 · mean of 50 |
|---|---|---|---|
| +6h2024-05-22 06:00 UTC | Unavailable | 1,005.57 hPa | 1,000.25 hPa |
| +12h2024-05-22 12:00 UTC | Unavailable | 1,007.00 hPa | 999.96 hPa |
| +18h2024-05-22 18:00 UTC | Unavailable | 1,007.40 hPa | 999.77 hPa |
| +24h2024-05-23 00:00 UTC | Unavailable | 1,007.08 hPa | 999.97 hPa |
| +30h2024-05-23 06:00 UTC | Unavailable | 1,006.35 hPa | 999.16 hPa |
| +36h2024-05-23 12:00 UTC | Unavailable | 1,005.47 hPa | 998.19 hPa |
| +42h2024-05-23 18:00 UTC | 1,004.00 hPa | 1,004.47 hPa | 996.20 hPa |
| +48h2024-05-24 00:00 UTC | 1,006.00 hPa | 1,003.20 hPa | 992.93 hPa |
| +54h2024-05-24 06:00 UTC | 1,004.00 hPa | 1,001.87 hPa | 989.02 hPa |
| +60h2024-05-24 12:00 UTC | 1,006.00 hPa | 1,000.60 hPa | 984.68 hPa |
| +66h2024-05-24 18:00 UTC | 1,004.00 hPa | 999.30 hPa | 981.35 hPa |
| +72h2024-05-25 00:00 UTC | 1,006.00 hPa | 997.62 hPa | 978.94 hPa |
| +78h2024-05-25 06:00 UTC | 1,004.00 hPa | 996.08 hPa | 978.35 hPa |
| +84h2024-05-25 12:00 UTC | 1,002.00 hPa | 994.54 hPa | 977.83 hPa |
| +90h2024-05-25 18:00 UTC | 998.00 hPa | 993.49 hPa | 977.34 hPa |
| +96h2024-05-26 00:00 UTC | 994.00 hPa | 992.52 hPa | 977.05 hPa |
| +102h2024-05-26 06:00 UTC | 990.00 hPa | 991.79 hPa | 976.80 hPa |
| +108h2024-05-26 12:00 UTC | 980.00 hPa | 990.99 hPa | 976.68 hPa |
| +114h2024-05-26 18:00 UTC | 975.00 hPa | 990.50 hPa | 976.74 hPa |
| +120h2024-05-27 00:00 UTC | 970.00 hPa | 990.81 hPa | 976.65 hPa |
| Measure | 1.1 | 1.2 | Coverage |
|---|---|---|---|
| JMA pressure-curve similaritysimilarity · higher is better | 0.7074 | 0.7118 | 134 starts / 40 storms |
| USA pressure-curve similaritysimilarity · higher is better | 0.7237 | 0.6637 | 131 starts / 40 storms |
| Mean track errorkm · lower is better | 798.4 | 471.2 | 1473 starts / 270 storms |
| +120 h track errorkm · lower is better | 1,646.4 | 1,031.6 | 1473 starts / 270 storms |
| Six-hour direction errordegrees · lower is better | 51.58 | 34.96 | 1473 starts / 270 storms |
| Centred route-shape similaritysimilarity · higher is better | 0.7544 | 0.8837 | 1473 starts / 270 storms |
| Geographic path similaritysimilarity · higher is better | 0.5345 | 0.6560 | 1473 starts / 270 storms |
| Frechet path distancekm · lower is better | 1,658.1 | 1,045.8 | 1473 starts / 270 storms |
| JMA central-pressure MAEhPa · lower is better | 13.53 | 12.84 | 134 starts / 40 storms |
| USA central-pressure MAEhPa · lower is better | 12.84 | 12.55 | 134 starts / 40 storms |
All 1,473 original forecast hashes and equal-storm route scores were reproduced. All 134 actual 50-member intensity replays passed their original-pressure and member/hash checks. No weights, forecasts or calibration were changed.
Daily plan SHA-256 965184f1e5ab3fcb52e304f39b38583a14f753870aa2295f2d01942a441dcfd6
1.2 source checkpoint SHA-256 f194a23d3f91ea76ad776dfad942fabd669eeae8b3fd665815463095367e9ee0
Wind/radius diagnostics remain in the report, not the main charts. Basin-wide pressure-field error is a separate 1.2-only diagnostic (2.72 hPa); no missing 1.1 field score is fabricated. Curve-similarity diagnostics remove mean and amplitude; the actual hPa examples above do not.
The predicted route. The pressure behind it.
Physical model-generated pressure fields · 1.2.82-R2 E1 is a preliminary two-member candidate trained through +24h; v1.2.48 Small E12 and Large E3 are fixed single saved replays.
White dashed: observed · Solid: TrackFormer clean one-forecast passes · v1.2.48 replays are fixed single saved forecasts in either mode · SteeringFlow V40: deterministic V39 route blend with aligned V10 pressure maps · Dashed DeepMind: 1-member reference · Dots: selected lead
hPa · Fixed scale · 2 hPa isobars · bold and labeled every 4 hPa · 2.5° model broad-scale field · Issue-time analysis Interpolated contours preserve the saved grid resolution.
22 TrackFormer checkpoints + deterministic algorithm + DeepMind reference. Choose overlays; the pressure selection is independent.
Mean forecast TrackFormer routes and central pressure average member-derived tracks and pressures. 1.2.78 E2, 1.2.79 E1/E2, 1.2.80 E1 and 1.2.82-R2 E1 use two distinct latent members; these are not 50-member results. Most older checkpoints use saved 50-member means. v1.2.48 Small E12 and Large E3 are individual archived forecasts, not ensembles; the mean/single switch leaves those saved runs unchanged. Fields are averaged on the same grid only for actual ensembles. The minimum of a mean map need not equal mean member central pressure. +0 is analysis, not a forecast.
One clean forecast The clean option is a saved unperturbed zero-latent pass where available. v1.2.48 entries are stored single causal replays from the named-case archive, not freshly rerun live forecasts.
Traditional forecast SteeringFlow V40 uses a deterministic V39 route blend and transports the V10 pressure map onto that route; central pressure is sampled from the displayed map. It is not a stochastic ensemble or a full weather model.
DeepMind and interpretation WeatherNext Cyclones Mini is a fixed one-member reference. TIP’s ten overlapping starts are a historical diagnostic, not proof of no overfitting. The 1979 v1.2.48 replay uses a later-trained checkpoint and is not evidence of operational skill or independent generalization.
Checkpoint: 1.2.82-R2 E1 · 1 forecast. Saved clean, zero-latent pass; model-produced field and route. E1 trained through +24h; saved autoregressive forecasts continue to +120h. Two-member results are provisional, not 50-member means.
Checkpoint SHA-256: 4bfec19806bda8ee16702958ccbbcee76b9ef9eda243d1908836ca640b2f4e1a
Prediction artifact SHA-256: ba3ecb7f9932ae37ffd781e6e47be0f05f02ad80dfa084f74874376dc3c5b744Compare observed routes with saved TrackFormer forecasts and SteeringFlow V40.
50-member mean · hourly cloud checks for new JMA analyses · page follows new forecasts every minute
Reading the latest complete released-model run…
The site precomputes supported models twice daily from current causal GFS analyses and refreshes this view every 30 minutes. Model selection switches among completed forecasts for the selected issue.
Map playback, original issue times, and observed tracks on the same coordinates.
Checking hosted backfill status; previously verified forecasts remain available.
Moving-core field recovery: a separate verified cloud backfill is being prepared. Original routes, core pressure numbers and basin fields are preserved. Until a core is exported, the map explicitly shows coarse basin pressure only. Core-field status ↗
Learned wind recovery: joining the same bounded cloud backfill. Each verified export enables this issue’s +6…+120 h wind graph automatically. Missing and outside-domain wind remains unavailable. Wind averaging period is unvalidated. Wind recovery status ↗
Cloud live checks run hourly at :17; the catalogue refreshes every five minutes. New forecasts require new valid JMA analyses; scheduler and source delays are possible. Western Pacific only. Older observed records remain in the full archive. Live job status ↗ · Coverage & exclusions ↗
The source has no wind-radius reports for this storm. Missing does not mean zero.
These are observations at the selected valid time, not Trackformer 1.2 predictions.
Interface adapted from your GitHub Pages typhoon explorer. No 1.1 output is relabelled 1.2. This is a research visualization, not an operational warning service.