A multi-scale loss formulation for learning a probabilistic model with proper score optimisation

TL;DR

This paper evaluates a multi-scale loss function for training probabilistic weather models, demonstrating improved small-scale variability modeling without sacrificing overall forecast accuracy, and highlights potential for scale-aware training improvements.

Contribution

It introduces and tests a multi-scale loss formulation within a weather forecasting model, enhancing small-scale variability representation while maintaining forecast skill.

Findings

Multi-scale loss improves small-scale variability modeling.

No negative impact on overall forecast skill.

Potential for future scale-aware model training.

Abstract

We assess the impact of a multi-scale loss formulation for training probabilistic machine-learned weather forecasting models. The multi-scale loss is tested in AIFS-CRPS, a machine-learned weather forecasting model developed at the European Centre for Medium-Range Weather Forecasts (ECMWF). AIFS-CRPS is trained by directly optimising the almost fair continuous ranked probability score (afCRPS). The multi-scale loss better constrains small scale variability without negatively impacting forecast skill. This opens up promising directions for future work in scale-aware model training.