Supervision

Corrected endpoints

The staged trainer scores corrected fields against matching reference endpoints. The default objective calls src/neuralpom/training/daily_objective.py; v1/losses.py composes configured losses and records their contributions.

Let $K$ be the set of configured loss components. For component $k\in K$, let $w_k$ be its weight, $s_k$ its normalization scale, and ${\mathcal{L}}_k$ its raw loss. The combined objective $\mathcal{L}$ is

$$\mathcal{L}=\sum _{k\in K}\frac{w_k}{s_k^2}{\mathcal{L}}_k.$$

The portable templates select weighted_mse. The composition interface also includes velocity spectral-energy and physical-mean losses. The configuration stores each component's name, weight and scale.

Spatial weighting and calibration

The objective uses wet-cell geometry and the physical quantities defined by the daily objective. Calibration uses training inputs. Pretrain excludes the sea-surface contribution according to the stage settings; the physical runtime and later stages retain the complete state.

The data and gradient checks are in tests/test_v1_data_losses.py. See Configuration for the template fields and Training for stage horizons.