A field guide to NeuralPOM
NeuralPOM couples a differentiable ocean model with a learned correction. The physical model advances the ocean; a residual U-Net updates its endpoint; the corrected state then enters the next physical interval. The central question is how to make a useful local correction remain useful as the ocean evolves. (Manuscript, Sections 2–5)
Choose your starting point
| Your goal | Start with | Then continue to |
|---|---|---|
| Understand the scientific idea | Hybrid model | Staged learning → Density-normal damping |
| Run the published workflow | Installation | Quick start → Reproduction |
| Study the physical model | Governing equations | Mode splitting → Two dynamical cores |
| Work on the implementation | Architecture | Differentiation and restart → API |
Four ideas to keep in view
The numerical state includes its history. A leapfrog solver needs more than the current velocity, temperature, and salinity fields. A correction updates the appropriate current and backward fields together, and density is recomputed before the physical model continues. See Time stepping.
A correction interval is not a physical time step. The coarse model advances at 600 s. Corrections arrive every hour or every day. The Explicit core also resolves the fast external mode with shorter substeps. See Configuration.
Training changes the trajectory presented to the network. Pretrain learns from cached physical endpoints. Midtrain learns through connected physical intervals. Posttrain adapts on states generated over ten-day trajectories. See Staged training.
Tracer corrections change density as well as tracer error. Temperature and salinity enter the equation of state, which controls buoyancy and baroclinic pressure gradients. Density-normal damping acts on a constrained density-modifying projection of the tracer increment. See the interactive explanation.
These definitions follow the manuscript, hybrid/sparse_coupling.py, and v1/engine.py.
A useful first session
Install the package and run the plan command in Quick start. This checks the configuration and exposes the sampling and update budgets before you prepare data. Then read the reference format, match one core and correction interval to its dataset, and follow the preparation and training sequence.
To inspect results immediately, open the paper reading guide. The manuscript figures and numerical source tables are included in the code repository.