Challenges of learning multi-scale dynamics with AI weather models: Implications for stability and one solution
Year: 2023
Type: preprint
Abstract: Long-term stability and physical consistency are critical properties for AI-based weather models if they are going to be used for subseasonal-to-seasonal forecasts or beyond, e.g., climate change projection. However, current AI-based weather models can only provide short-term forecasts accurately since they become unstable or physically inconsistent when time-integrated beyond a few weeks or a few months. Either they exhibit numerical blow-up or hallucinate unrealistic dynamics of the atmospheric variables, akin to the current class of autoregressive large language models. The cause of the instabilities is unknown, and the methods that are used to improve their stability horizons are ad-hoc and lack rigorous theory. In this paper, we reveal that the universal causal mechanism for these instabilities in any turbulent flow is due to \textit{spectral bias} wherein, \textit{any} deep learning architecture is biased to learn only the large-scale dynamics and ignores the small scales completely. We further elucidate how turbulence physics and the absence of convergence in deep learning-based time-integrators amplify this bias, leading to unstable error propagation. Finally, using the quasi-geostrophic flow and European Center for Medium-Range Weather Forecasting (ECMWF) Reanalysis data as test cases, we bridge the gap between deep learning theory and numerical analysis to propose one mitigative solution to such unphysical behavior. We develop long-term physically-consistent data-driven models for the climate system and demonstrate accurate short-term forecasts, and hundreds of years of time-integration with accurate mean and variability.
Source: arXiv (Cornell University)
Authors: Chattopadhyay, Ashesh, Sun, Y. Qiang, Hassanzadeh, Pedram
Institutions: University of California, Santa Cruz, University of Chicago
Language: English
Subfield: Atmospheric Science
Field: Earth and Planetary Sciences
Domain: Physical Sciences
SDG: Climate action
Open Access status: green
Awards: CAIG: Stability and Physical Consistency of AI-based Climate Emulators for Estimating Forced Responses, G1900090525, G4805370682, G5781760129, ATMOSPHERIC BLOCKING EPISODES ARE IMPORTANT CONTRIBUTORS TO HIGH-IMPACT EXTREME WEATHER AND ARE A PRIME TARGET OF SUB-SEASONAL PREDICTION. INTRINSIC PREDICTABILITY OF BLOCKING AND SOURCES OF PREDICTION ERRORS HAVE NOT BEEN WELL CHARACTERIZED WHICH REPRESENT AN OPPORTUNITY FOR IMPROVED MEDIUM RANGE WEATHER FORECASTS. WE PROPOSE A MULTI-PRONGED APPROACH TO BETTER UNDERSTAND PREDICTABILITY AND TO IMPROVE PREDICTION OF ATMOSPHERIC BLOCKING EVENTS AND THE ASSOCIATED EXTREME WEATHER EVENTS. TWO SUITES OF MODELS BASED ON THE GODDARD EARTH OBSERVING SYSTEM MODEL (GEOS) AND THE COMMUNITY ATMOSPHERE MODEL (CAM) WILL BE USED. WE WILL FIRST EVALUATE THE PERFORMANCE OF THE MODELS AGAINST THE MERRA2 AND ERA INTERIM REANALYSES AND ADDITIONAL SATELLITE OBSERVATIONS FROM THE GLOBAL PRECIPITATION ACTIVE PASSIVE (SMAP) GLOBAL PRECIPITATION MISSION (GPM) AND THE TROPICAL RAINFALL MEASURING MISSION (TRMM). WE WILL EVALUATE THE MODELS ABILITY TO SIMULATE BLOCKING AND WEATHER EXTREMES CLIMATOLOGY TO FAITHFULLY CAPTURE BLOCKING DYNAMICS AT A PROCESS LEVEL AND IN THE CASE OF THE GEOSBASED MODELS THE ABILITY TO FORECAST BLOCKING EVENTS AND THE RESULTING WEATHER EXTREMES WITH THE MODELS DATA ASSIMILATION SYSTEM. WE WILL THEN CARRY OUT INTRINSIC PREDICTABILITY STUDIES USING A PERFECT MODEL APPROACH AS WELL AS CROSS-MODEL PREDICTION STUDIES. THE SUITES OF MODELS THAT WE USE HAVE VERSIONS THAT DIFFER ONLY IN THEIR REPRESENTATIONS OF THE STRATOSPHERE MOIST PROCESSES OR SOIL MOISTURE AND THEREFORE ALLOW US TO ISOLATE THE CONTRIBUTION OF THESE MODEL COMPONENTS ON PREDICTABILITY AND PREDICTION ERRORS. THE EFFECT OF NUMERICAL RESOLUTION WILL ALSO BE EXPLORED. THE USE OF BOTH THE GEOS- AND CAM-BASED MODELS WILL ALLOW CROSS COMPARISONS AND ASSESSMENTS OF THE ROBUSTNESS OF THE RESULTS ACROSS MODELS. LASTLY WE WILL GENERALIZE OUR RESULTS TO A BROADER RANGE OF MODELS IN THE SUBSEASONAL-TO-SEASONAL PREDICTION PROJECT DATASET. THE PROPOSED RESEARCH WILL ENHANCE BOTH NASA MODELING CAPABILITY AND THE SCIENCE RETURN ON NASA S INVESTMENT IN MODELING AND SPACE-BASED OBSERVATIONS IN THE CRITICAL AREA OF PREDICTING EXTREME WEATHER., G8876996369
OpenAlex ID: W4366198063