Applicability of ML Weather-Forecasting Strategies to Long-Term Climate Projections

Determine whether machine-learning strategies that have achieved success in short-time weather predictions can be applied to improve climate projections by estimating changes in the statistics of weather events (e.g., return periods of heat waves or tropical cyclones) over decadal to centennial timescales.

Background

Recent machine-learning models for weather forecasting (e.g., nowcasting and medium-range forecasts trained on reanalysis) have demonstrated strong predictive skill at short lead times, benefiting from abundant and relatively stationary data streams. In contrast, climate projections involve non-stationary changes and the small-data regime for future conditions, raising concerns about whether strategies effective for short-term weather prediction can generalize to long-term climate statistics.

This open question highlights a central challenge in translating the successes of ML-based emulators from initial-value, short-horizon weather prediction to boundary-value, long-horizon climate projection tasks that require reliable estimates of changes in event statistics under evolving forcings.