Explanation of Dynamic Physical Field Predictions using WassersteinGrad: Application to Autoregressive Weather Forecasting

The core challenge WassersteinGrad addresses is not just interpretability in general, but the specific problem of attributing predictions across time steps in autoregressive models, where errors compound and a single saliency map no longer cleanly maps to a single decision.

This sits within a cluster of interpretability work Modelwire has been tracking. The ORCA paper from mid-April tackled post-hoc interpretability for SVMs by expanding decision functions into explicit feature coordinates, and WassersteinGrad is solving an analogous problem one layer of complexity higher: not a static classifier but a model that feeds its own outputs back as inputs across time. The nonlinear separation principle paper from the same period is also relevant background, since it addresses stability conditions in recurrent architectures, which is precisely the class of model where attribution becomes ambiguous. Neither of those papers, however, dealt with physical field data or the operational pressure of safety-critical deployment, which is where WassersteinGrad is staking its claim.

The real test is whether WassersteinGrad attributions hold up against established weather forecast verification benchmarks like WeatherBench 2, specifically whether the highlighted regions correspond to meteorologically meaningful features that domain experts independently identify as causal.