Abstract

Author list: Tom Beucler (UNIL), David Neelin (UCLA), Hui Su (HKUST), Ignacio Lopez-Gomez (Google Research), Chris Bretherton, Oliver Watt-Meyer (AI2), Tapio Schneider, Costa Christopoulos (Caltech), Will Chapman (UC Boulder), Laure Zanna (NYU), Aditya Grover (UCLA), Adam Subel (NYU) Deep learning emulates atmospheric reanalyses with high fidelity, enabling increasingly well-calibrated ensemble weather forecasts at progressively longer lead times. To extend these gains to climate-relevant horizons, AI prediction systems must produce credible forced responses to drivers of interest (e.g., greenhouse gases, land-use change). We propose a minimal, testable framework for AI climate modeling: (i) represent external forcings explicitly and restrict them to physically appropriate state tendencies; and (ii) stress-test robustness in out-of-distribution regimes, including extremes and counterfactual trajectories. Using leading climate emulators and hybrid physics-AI models, we identify coupling and development challenges and compare scaling with resolution and effective complexity. AI models do not appear intrinsically more efficient than GPU -ported dynamical models once complexity is accounted for, yet they can directly predict target variables at the desired grid without integrating the full high-frequency, multivariate state. Diverse ML downscaling strategies can partially substitute for explicit fine-scale resolution when observations are available, paving the way towards inexpensive, local risk assessment across prediction horizons.