Optimal Forecasts of the 2021 Pacific Northwest Heatwave from Deep-Learning Sensitivity Analysis

Forecast sensitivity has traditionally been quantified by linearizing perturbations around a control forecast using tangent-linear/adjoint models, or ensemble methods, to relate small changes in the initial conditions to a forecast metric. These approaches are limited in both the duration of the forecast, and the amplitude of the perturbations, by the linear approximation. Automatic differentiation facilities in machine-learning frameworks enable a new approach, which we call deep-learning sensitivity analysis, requiring no linear approximation. Combined with computationally efficient deep-learning models, this new approach allows fully nonlinear optimization of initial conditions to improve forecast accuracy at long lead times.

We apply this new approach to forecasts of the June 2021 Pacific Northwest heatwave using the GraphCast model and the JAX framework for automatic differentiation. Starting with the ERA5 reanalysis, we optimally perturb these initial conditions using deep-learning sensitivity analysis. We find that the optimal perturbations yield over 90\% reduction in 10-day forecast errors over the Pacific Northwest relative to a control forecast starting from the ERA5 reanalysis. Similar improvements are found for Pangu-Weather model forecasts initialized with the GraphCast-derived optimal, suggesting that model error is not an important part of the initial perturbations. The initial perturbations have amplitudes typical of analysis errors, and their spatial structure reveals a correspondence with features in the tropics and midlatitudes that previous studies have identified as important precursors to the heatwave.