Long-term trends in regional moisture and heat extremes in AMIP and NeuralGCM simulations

Reliable climate model simulations are crucial for informing mitigation and adaptation policies in response to future climate change. However, state-of-the-art global climate model simulations continue to show important regional biases relative to observations in their long-term trends relative to observations—biases that have persisted across generations of models, even in those with prescribed historical sea surface temperatures. These biases influence the simulation of key aspects like changes in the frequency and intensity of moisture and heat extremes, which are important for projecting the growing societal risk associated with climate change. Recent advances have produced hybrid dynamical and machine learning climate simulations, which integrate a differentiable dynamical core with neural network learned physics, that are stable on decadal timescales when forced by observed sea surface temperatures and may not be subject to the same biases as traditional climate models. In this study, we examine decadal trends in extreme humidity and temperature characteristics over land using NeuralGCM, one of the first of these hybrid climate models, and compare with global climate models also forced by observed sea surface temperatures from the Atmospheric Model Intercomparison Project Phase 6 (AMIP6). We assess the relative contributions of NeuralGCM's learned physics and dynamical core to enhancing the fidelity of long-term regional trend simulations compared to conventional models.