Harnessing compound flood risk assessment Utilizing a physics-informed GeoAI surrogate framework considering social vulnerability sources.

Compound flooding (CF), driven by the interaction of extreme storm surge, rainfall, and river discharge, poses escalating risks to society, especially in densely populated megadeltas. Most assessments still treat these drivers separately, and few link flood hazards to social vulnerability. This study develops a physics-informed GeoAI (PI-GeoAI) surrogate framework that couples a hydrodynamic model with supervised machine learning for spatiotemporal risk assessment. We demonstrate the framework using Typhoon Mangkhut in the Greater Bay Area, China. The Random Forest surrogate outperforms seven other algorithms, reaching 72% within-one-class accuracy across seven flood-damage classes (depth-derived categories) and reproducing the broad spatial pattern of flooding. Building on this surrogate, we conducted a factorial perturbation experiment with 120 plausible mid-century scenarios, generated by combining six storm-surge, five precipitation, and four discharge-intensification levels. CF risk is quantified by the cube-root geometric mean of population exposure, PCA-derived Social Vulnerability Index (SoVI), and hazard severity index. The study finds that baseline CF risk concentrates in the low-to-moderate range, and the perturbation primarily reshapes risk magnitudes rather than redistributing populations across risk classes. The perturbation experiment reveals a near-additive, surge-dominated amplification under combined high-end forcing, with flood depth increasing by 1.11× and risk by 1.01×. Correspondingly, 23.8% of pixels shift upward by at least one damage class and 10.6% by at least one risk class. Storm-surge perturbations dominate the compound-risk gradient. This framework provides a scalable, spatially explicit approach for evaluating CF risk and identifying socially differentiated hotspots in urbanizing megadeltas.