yanxingjianken/ki-era5-tracks

Conditional Kinematic Interpolant for ERA5 air-parcel tracks

Predicts the acceleration and a scaled velocity score of an air parcel from its state (polar-stereographic x, y and log-pressure height z = -7 km ln(p/1000 hPa), velocity) and the observed tracks of up to 192 other parcels of its patch over 48 h (every 3 h), after Jiang, Marzouk, Bose & Pooladian, Learning smooth dynamics via Kinematic Interpolants, arXiv:2610.22519. Integrated forward (forecast) or backward (backcast) up to 24 h, deterministically or as a stochastic ensemble (ensemble.json: eps_rel 1.5, chosen so that the 24-h spread matches the error of the ensemble mean on validation patches).

• Training data: LAGRANTO tracks in hourly ERA5 (0.25°, 37 pressure levels), releases every 5 days of 2015-2017, 20 start levels, 16 patches of 480 parcels per level, both hemispheres (southern mirrored); 2.5e7 training tracks.
• Test (2021, unseen year): median 24-h error of the deterministic forecast 335 km, backcast 254 km; about 45 % below straight motion at 12-24 h.
• context_ki_era5.pt: EMA weights, training arguments (args, width d_model 512) and the per-axis scales (scales, also in scales.json). Code: github.com/yanxingjianken/ki-lagrangian-tracks (03_era5).

In-browser graphs (onnx/)

encoder.int8.onnx (context tokens (1, 3264, 8) -> latents) and decoder.int8.onnx (latents, queries (1, Q, 8) -> acceleration and scaled score) with the weights of the matrix products stored as 8-bit integers per output channel; config.json holds the scales, the ensemble settings and the token layout. On 2021 tracks their 24-h ensemble-mean errors match the full-precision model within the scatter between sets of random kicks. Used by the page yanxingjianken.github.io/blocking-plots/rwb-prediction/ with onnxruntime-web.