Because they are highly interdependent, most infrastructure systems are vulnerable to disasters such as severe floods. An amazing amount of data has been made available to better anticipate the impact of severe floods and become more sustainable and resilient. Yet, integrating these data sources and modeling the impact of floods remains a challenge. Here, we develop a transferable, fact-calibrated, weakly supervised, physically constrained probabilistic modeling framework. During a flood, we estimate facility-level flood risk by integrating XGBoost and depth–damage functions. Furthermore, we leverage Granger tests, distributed-lag models, and spatiotemporal Durbin specifications to examine cross-system predictive temporal dependencies and finite-horizon lag patterns. We have three main findings. First, flood-induced damage exhibits a threshold-driven structure: damage across all infrastructure types is activated within a narrow depth band (~0.20–0.26m), suggesting universal hydraulic tipping points for sustainable infrastructure management. Second, at the predictive temporal-dependence level, cascading effects are not as extensive as their physical connections suggest. Rather than spreading everywhere, these effects are highly selective, focusing on a limited set of bridging nodes, which provides a roadmap for cost-effective resilience interventions in urban contexts. Third, resilience is shaped by the network’s propagation structure, with systematic differences in response latency and finite-horizon risk-signal persistence across infrastructure types.