Machine Learning From Internal Variability: Simulating Weather Extremes and Climate Feedbacks

February 23, 2026

Ankur Mahesh

Hosted by Jim Hurrell

Abstract

In a warmer climate, how will extreme weather events change? And how will average rainfall shift as the climate responds to greenhouse gases? Answering these central scientific questions is difficult because physics-based climate models are computationally expensive and uncertain. In this talk, I will show how machine learning (ML) weather emulators trained on historical data can answer these questions. The unifying theme is that these emulators will learn from the rich natural variability in the historical dataset: they do not require retraining on future climate scenarios.

First, I will demonstrate how the FourCastNet ML weather emulator can generate huge ensembles (7,500 members) to capture the long tails of climate hazards that traditional 100-member ensembles miss. Validated against operational weather systems, these ML ensembles maintain realistic error growth and physical fidelity. With this petabyte-scale huge ensemble, we can characterize plausible extreme weather events that could have occurred in summer 2023, the hottest summer on record.

Second, I will show how historically-trained emulators can quantify climate feedbacks to elevated greenhouse gases. The climate feedbacks to carbon dioxide are divided into slow and fast timescales. Fast feedbacks emerge from atmospheric radiative and convective processes operating on weekly timescales, and these processes are already fully operative in the present-day climate in response to natural variability. I will present a new design of FourCastNet that successfully encodes the physics of fast feedbacks. It simulates future changes in the global hydrological cycle with results that align with physics-based earth system models.