Architecture
Package structure
| Package | Responsibility |
|---|---|
cores/explicit | Explicit free-surface POM |
cores/implicit_p0_fixed | Semi-implicit free-surface POM |
cores/case_adapters | Case boundaries, compatibility operators and tidal input data |
cases | Shared Double-gyre settings |
models | Residual U-Net and Triton architecture |
hybrid | State packing, constrained correction and physical execution |
v1 | Data loading, caches, staged training, restart and tracking |
evaluation | Frozen inference, deployment policies and diagnostics |
physics_baselines | Physical closure comparisons |
ai_baselines | Direct full-state neural forecasts |
training | Shared daily objective |
The production core registry in src/neuralpom/core_factory.py exposes explicit and implicit_p0_fixed.
State and correction
The physical runtime restores the complete numerical state and advances it to a correction endpoint. The network input contains 101 channels from six three-dimensional fields over 16 storage levels and five two-dimensional fields. The network predicts velocity, temperature and salinity increments on 15 active layers, giving 60 output channels.
hybrid/state.py packs and restores the numerical state. hybrid/sparse_coupling.py applies column and layer constraints, updates the required history fields, and refreshes density before physical integration continues. v1/runtime.py connects this operation to data, calibration and the physical core.
Entry points
The neuralpom console command dispatches to v1/cli.py; neuralpom-ai dispatches to ai_baselines/cli.py. The short scripts under scripts/v1/ expose plan, preparation, training, evaluation and reporting from a checkout. Reproduction describes the complete workflow.