beita6969 / scienceclaw-benchmark-for37

Global gridded 2 m temperature forecast, 24 h lead

Task

• Input: a training record fields (T, lon, lat) in K with ISO UTC times (6-hourly, consecutive steps); per initialisation context (n, 4, lon, lat) in K at init-18 h, -12 h, -6 h, 0 h and init_time (n ISO strings).
• Deliverable: float forecast (n, lon, lat) in K for init + 24 h, axes in the store's (longitude, latitude) order, finite and within a physical range (about 150-350 K, which catches degrees Celsius).
• Quality: WeatherBench 2 RMSE, lower is better: per initialisation sqrt of the area-weighted mean squared error (weights proportional to sin(lat + d/2) - sin(lat - d/2), mean 1), then the mean over initialisations (scilib.weather.lat_weighted_rmse). Natural references: persistence (the field at init) and seasonal cycle plus damped anomaly persistence.

Routes

• scilib.weather.fit_predict(fields, times, context, init_time) (operator gridded_field_patch_ridge_forecast): per grid cell and UTC-hour bin a seasonal cycle (constant plus 3 annual harmonics), then one ridge regression per latitude row of the anomaly at +24 h on the (2r+1)x(2r+1) anomaly patches of the four context fields (defaults radius 2, lam 30; longitude wraps). Building blocks: fit_seasonal_cycle, seasonal_cycle_at, anomalies (operator gridded_field_seasonal_anomalies), fit_patch_ridge, predict_patch_ridge, lat_weights.
• Validation: weather.blocked_cv(fields, times, gap_days=6) runs temporal block cross-validation inside the record (training steps farther than gap_days from the held-out block) and returns per-initialisation RMSE. Operator gridded_field_lat_weighted_rmse scores any forecast.
• The tool library has no pretrained global weather model asset (scienceclaw_tools(operation=weights) lists what is staged).

Pitfalls

• Respect temporal causality: fit only on record steps that precede the forecasts you issue, and compute row i of the forecast from context[i] and init_time[i] alone. Another initialisation's context can contain the future of item i, so never pool contexts across items to forecast a target.
• A seasonal cycle fitted on a short record (for instance one year) is extrapolated to other seasons or years; validate on a later or different period, not on random time steps (neighbouring steps are strongly autocorrelated, hence the block gap).
• Use one forecast function for validation initialisations and final initialisations, and score validation forecasts only against validation targets built from validation contexts.
• Keep units in K and the (lon, lat) axis order; longitude is periodic, latitude is not. Do not convert to degrees Celsius or transpose the output.