Core Concepts
S2S prediction challenges require physics-based benchmarks like ChaosBench for improved forecasting.
Abstract
ChaosBench introduces a large-scale, physics-based benchmark for S2S climate prediction, addressing existing limitations in forecasting range, operational baselines, and explainability. The dataset includes 45 years of observations and 8 years of physics-based forecasts, establishing tasks for full and sparse dynamics prediction. Existing models perform poorly compared to climatology, emphasizing the need for physics-based constraints.
Stats
44 days lead-time
60 variables (channels)
45 years
48 variable-channels
Quotes
"Our benchmark is one of the first to incorporate physics-based metrics to ensure physically-consistent and explainable models."