ABSTRACT Reliable precipitation projections are essential for water resource planning, agriculture, and disaster risk management. However, coupled model intercomparison project phase 6 (CMIP6) models often exhibit substantial biases, limiting their direct use in local-scale studies. This study evaluated the capability of extreme gradient boosting (XGBoost), a machine learning-based bias correction method, to improve precipitation simulations from 15 CMIP6 models using climate hazards infrared precipitation with stations (CHIRPS) as reference. Empirical quantile mapping (EQM) was applied as a benchmark. XGBoost consistently outperformed EQM in mean-state correction, achieving larger error reductions. For instance, in GFDL-ESM4, RMSE decreased from 4.85 to 0.11 mm/year and PBIAS from −49.97 to 0.49% with XGBoost, compared to 0.36 mm/year and −0.81% with EQM. Similar improvements were observed for MIROC6. However, residual biases persisted in models, such as ACCESS-CM2, especially during March–May rainy season. For extremes, EQM better preserved high-intensity precipitation tails, while XGBoost tended to underestimate rare events when they were underrepresented in training data. Future projections indicate a 4.9–5.3% precipitation increase under SSP2-4.5 and SSP5-8.5 scenarios for mid (2025–2055) and late century (2056–2086). Overall, the findings confirm XGBoost's potential for improving mean precipitation, while highlighting EQM's strength for extreme critical to flood-risk management.
Tadase et al. (2026) studied this question.