Daily full-state AI baselines
Forecast task
The direct neural models predict a complete daily state without POM integration during prediction. StateCodec encodes 43 spatial fields as 313 channels on the 90 × 82 × 16 storage grid. Frozen training-input means and standard deviations normalize each channel. Static geometry remains external.
The registry includes Triton, ResUNet, SimVP, ConvLSTM and PastNet. src/neuralpom/ai_baselines/upstream.py records the external source identities; scripts/fetch_ai_upstream.py fetches the pinned architecture files required by the corresponding adapters.
Training
| Stage | Trajectory | Update |
|---|---|---|
| Pretrain | One next-day forecast | One standardized next-state loss |
| Posttrain | Ten-day model-generated forecast | One daily update with detached carry |
The sample lists use neuralpom.v1.sampling.stage_samples. Pretrain selects validation-best weights. Posttrain carries its actual generated state forward after each parameter update. The full-state objective uses standardized channel errors and wet-grid weights.
Use neuralpom-ai --help to inspect preparation, planning, training and frozen evaluation commands. The matching core configuration identifies the reference ensemble and geometry. See Configuration and Reproduction for assets and outputs.
Diagnostics
Frozen AI forecasts use the common state diagnostics for physical-unit field errors, density, tracer decomposition, profiles and velocity spectra. The evaluation runner records reference pairing, weight identities and completed origins with the forecast outputs. The implementation is in src/neuralpom/ai_baselines/.