Holocene Rain-Belt Reorganization in the East Asian Summer Monsoon Identified by Physics-Informed Neural Networks

Holocene precipitation variability in the East Asian summer monsoon (EASM) region provides a baseline for separating externally forced trends from internal variability and for contextualizing modern extremes, yet quantitative reconstructions are limited by sparse proxies and biased simulations. This study develop a physics-informed neural network (PINN) fusing ERA5, PMIP simulations, speleothem δ¹⁸O, and topography. The objective function enforces water-balance closure, Clausius-Clapeyron scaling, monsoon coherence, and orographic forcing; thus, reanalysis constrains present-day relationships while PMIP priors and cave records anchor Holocene states. Validation against ERA5 (1981-2010) confirms skilful reproduction of regional climatology and interannual variability. SHAP analysis shows precipitation is controlled by low-level winds (moisture transport), surface energy, and topography, with wind-topography interactions shaping regional gradients. The reconstruction reveals basin-wide orbital drying as the leading mode and a secondary meridional mode reflecting rain-belt shifts intensifying near 4.2 ka. Cross-correlations indicate northward phase propagation with an ～century lead in the south, consistent with anomaly transmission along moisture pathways. Notably, speleothem δ¹⁸O exhibits a robust near-zero-lag inverse relation with reconstructed precipitation at millennial/centennial scales. These results imply the East Asian “4.2 ka event” reflects rain-belt reorganization rather than a uniform amplitude change.