What Drives General Circulation Model Biases in Precipitation Extremes? A Generalised Pareto Distribution (GPD)-Based Diagnostic of CMIP6 Tail Behaviour

Abstract

This study uses a point-scale Peaks-Over-Threshold (POT) Generalised Pareto Distribution (GPD) framework to evaluate extreme precipitation simulated by 32 CMIP6 models from 2015 to 2024. Daily rainfall exceeding high thresholds is fitted to the GPD using three estimators: Maximum Likelihood Estimation, L-moments, and Bayesian inference. Return levels for 2, 5, 10, 20, 50, and 100 year events are computed. Two climatically distinct sites are analysed, one in tropical South India and one in mid-latitude Northern Italy, with observational benchmarks from IMD4 and E-OBS datasets. Results show systematic model biases. CMIP6 models generally overestimate extreme rainfall in India and underestimate it in Italy. Biases are closely linked to convection parameterisation schemes. Zhang-McFarlane models simulate Indian monsoon extremes within about 30 percent accuracy but underpredict Italian 100-year rainfall by 50 to 65 percent. Gregory-Rowntree scheme models show mixed results. Models with simple bulk convection schemes tend to have the largest errors, often exceeding 100 percent in both regions. Notably, the finely tuned Gregory-Rowntree model KACE-1-0-G performed well in both domains, highlighting the importance of targeted calibration regardless of scheme type. Variability in threshold selection, driven by convection scheme differences and drizzle bias, affects return level estimates more than GPD shape or scale parameter variations. Differences across estimation methods are minor compared to inter-model spread, with Bayesian inference showing the most stable results. Increasing model resolution alone does not guarantee improved simulation of extreme precipitation. Overall, findings underscore that model physics, particularly convection parameterisation, is the primary determinant of extreme precipitation biases, suggesting that advancements in physical processes offer greater potential for reducing biases than resolution enhancements alone.