Hybrid physics-data-driven modeling for sea ice thermodynamics and transfer learning

TL;DR

This work develops a state-dependent, neural-network bias-correction for the Icepack sea-ice column model by learning forecast error tendencies under perturbed snow-thermodynamics and radiative parameters. The Icepack–NN hybrids deliver stable online performance, outperforming a climatology-based benchmark across long lead times, and show strong robustness to initial-condition and atmospheric-forcing errors. Transfer learning is proposed as an efficient way to adapt pretrained corrections to updated model configurations, with a practical criterion (based on $\text{nMSE}_{\textrm{dir}}$) guiding when fine-tuning is preferable to retraining. A PI-RFE feature-importance analysis highlights ice-concentration/volume states and ice-layer enthalpies as the dominant predictors, suggesting a minimal yet physically consistent feature set for bias correction; the results offer concrete guidance for deploying ML-based corrections in evolving ice-physics frameworks.

Abstract

This study explores a physics-data driven hybrid approach for sea-ice column physics models, in which a machine learning (ML) component acts as a state-dependent parameterization of forecast errors. We examine how perturbations in snow thermodynamics and sea-ice radiative properties affect forecast errors, and train dedicated neural networks (NNs) for each model configuration. The performance of the hybrid models is evaluated for long lead-time forecasts and compared against a benchmark system based on climatological forecast-error estimates. The NN-based hybrids prove to be stable, robust to initial condition and atmospheric forcing errors, and consistently outperform their climatology-based counterpart. To derive guiding principles for efficiently handling possible physical model updates, we perform transfer learning experiments to test whether pretrained NNs optimized for one model configuration can be successfully adapted to another. Results indicate that direct evaluation of pretrained networks on the target task provides useful insights into their adaptability, recommending transfer learning whenever performance exceeds a trivial baseline. Finally, a feature-importance analysis shows that atmospheric forcing inputs have negligible influence on NN predictive skill, while ice-layer enthalpies play a key role in achieving satisfactory performance.

Figures (22)

• Figure 1: Locations of Icepack simulations, selected from the ERA5 grid downsampled to 1° resolution, at latitudes between 65° N and 90° N.
• Figure 2: Ice volume time-series over a 15-year period at the 6 selected locations: red dashed lines () indicate the truth; solid grey lines () show forecasts from individual models with perturbed parameters; black dashed lines () denote the forecast ensemble means.
• Figure 3: Schematic of the hybrid model loop.
• Figure 4: Scatter plots of total ice concentration and volume errors at lead time $\tau = 60$ days, as functions of selected forecast state variables and atmospheric forcings. Point colors represent density estimated via Gaussian KDE. Rows (a) and (b) show data aggregated across all model configurations, whereas panels (c) and (d) show results from model $\mathrm{m}\_05$ only.
• Figure 5: Ice volume (a) and concentration (b) RMSE and BIAS at a lead time of 60 days for all model configurations sorted by the RMSE.