PhysMetrics.Weather: An Evaluation Framework for Physical Consistency in ML Weather Models
Researchers have developed PhysMetrics.Weather, a new evaluation framework designed to assess the physical consistency of machine learning weather prediction models. Unlike traditional models, MLWP models are often data-driven and evaluated with pixel-wide error metrics, which do not guarantee adherence to physical laws. This framework introduces metrics for conservation, spectral, and dynamical realism, aiming to guide the development of physics-informed architectures and determine the reliability of MLWP models for operational use. AI
IMPACT This framework could improve the reliability and operational use of AI-driven weather forecasting models by ensuring their outputs align with physical laws.