Using gridpoint relaxation for forecast error diagnostics in neural weather models

Abstract: Deep-learning (DL) weather prediction models offer notable advantages over traditional physics-based models, including auto-differentiability and low computational cost, enabling detailed diagnostics of forecast errors. Using our convolutional encoder-decoder model, ConvCastNet, we systematically relax selected subdomains of the forecast fields towards "true" weather states (ERA5 analyses) and monitor the forecast skill gain in other regions. Our results show that an 8-day midlatitude forecast improves substantially when the stratosphere and boundary layer are relaxed, while relaxation of the tropical atmosphere has negligible effect. This highlights the importance of accurately representing the stratosphere and the planetary boundary layer for medium-range weather prediction. Additionally, we investigate the relationship between model error sensitivity to initial conditions with relaxation experiments. By utilising auto-differentiability, we analyse saliency maps, i.e. the gradients of the forecast errors with respect to input fields, to identify overlapping regions of large error sensitivity and high impact of relaxation to truth. This combined approach provides valuable heuristics for diagnosing neural model errors and guiding targeted model improvements.