Abstract
Climate models serve as critical tools for evaluating the effects of climatechange and projecting future climate scenarios. However, the reliance onnumerical simulations of physical equations renders them computationallyintensive and inefficient. While deep learning methodologies have madesignificant progress in weather forecasting, they are still unstable forclimate emulation tasks. Here, we propose PACER, a lightweight 684K parameterPhysics Informed Uncertainty Aware Climate Emulator. PACER emulates temperatureand precipitation stably for 86 years while only being trained on greenhousegas emissions data. We incorporate a fundamental physical law ofadvection-diffusion in PACER accounting for boundary conditions and empiricallyestimating the diffusion co-efficient and flow velocities from emissions data.PACER has been trained on 15 climate models provided by ClimateSetoutperforming baselines across most of the climate models and advancing a newstate of the art in a climate diagnostic task.