New wine in an old bottle

NeuralPOM: A Hybrid Ocean Forecast Model with Differentiable Dynamics and Staged Learning

Ruiqi Shu · Guangliang Liu · Ruijian Gou · Tianrun Gao · Haomin Wen · Hao Li · Xiaomeng Huang

arXiv:2611.02460 ↗ · Read the PDF ↗ · Supplement ↗ · Numerical source tables ↗

The scientific question

A coarse ocean model misses parts of the resolved evolution of a finer reference model. A learned endpoint correction can reduce local error, but its effect is then carried through pressure gradients, advection, and mixing. The paper examines how differentiable dynamics, staged learning, and constrained corrections shape that subsequent evolution. (Manuscript, Sections 1–2.)

The reference ocean

The primary experiment is a wind-driven double gyre. Each physical core has its own eight-member fine-grid reference ensemble, restricted consistently to the coarse forecast grid. The fine horizontal grid is 354 × 322, and the coarse grid is 90 × 82, with 15 active vertical layers. (Manuscript, Section 3.1.)

Developed reference surface fields, reproduced from the manuscript. The full panels retain their axes and color scales.

Three parts of the method

Part	Question it answers	Tutorial
Differentiable physics	How does an earlier correction change a later state?	Physical framework, adjoint solve
Staged learning	Which states and gradient paths train the network?	Pretrain → Midtrain → Posttrain
Constrained correction and DND	Which budgets and density responses should the update preserve or attenuate?	Correction, DND explorer

Density response and forecast error

For hourly coupling, the paper reports DND reductions relative to the same-weight Undamped update of roughly 5%–9% in zonal velocity and SSH error, 12%–13% in temperature, 38%–43% in salinity, and 9%–10% in density across the two physical cores. These are the paper's reported error comparisons; the projection's 43.75% squared-norm reduction follows separately from its retention factor. (Manuscript, Section 5.2.)

Projection and density-response diagnostics from the manuscript. Read Section 5 and Appendix C for the definitions of each plotted quantity.

Forecast tracer diagnostics from the manuscript. Numerical results are available in paper/data/ and paper/data/policy_comparisons/.

Read a forecast comparison correctly

The paired protocol uses 16 forecast origins and 100-day trajectories, initialized from reference member 7 at model days $8000+32r$, where $r=0,\dots ,15$. Model weights remain fixed during evaluation. The paper compares physical closures, learning stages, correction policies, and direct neural baselines. Their settings and comparison windows are documented in Section 3 and Appendix D.

For a specific comparison, match the physical core, correction interval, initial state, variable, spatial weighting, and lead-time window. The reproduction guide maps the saved source tables and evaluation commands to these choices.

Trace the paper to the repository

Evidence or method	Repository location
Model architecture	models/corrector_net.py
State constraints	hybrid/sparse_coupling.py
Training stages	v1/engine.py
Density-normal policy	evaluation/density.py
Forecast protocol	configs/examples/evaluation_100d.json
Forecast and stage metrics	paper/data/

Authors and availability

Ruiqi Shu, Guangliang Liu, and Ruijian Gou contributed equally, in this author order.

The repository README and manuscript contain the full author affiliations. Corresponding authors are Hao Li (lihao_lh@fudan.edu.cn) and Xiaomeng Huang (hxm@tsinghua.edu.cn).

Code is available on GitHub under Apache-2.0. Data are hosted in the NeuralPOM ModelScope repository. See its dataset documentation for available assets and access.