This research develops a digital twin–based decision-support framework for real-time, impact-aware urban flood management. We introduce a reinforcement learning–driven system that integrates heterogeneous sensor data (e.g., FloodNet) and contextual vulnerability indicators to jointly predict localized flood impacts and optimize budget-aware alerting strategies. Moving beyond traditional forecast accuracy, RAFT enables actionable decision-making by prioritizing high-impact events and guiding emergency response. Empirical results show up to 28% improvement in precision over baseline models, with flexible gains in recall across configurations. Complementing our agent framework, we design an interactive digital twin environment that supports live data ingestion, simulation, and stakeholder-informed policy learning. Ongoing work extends this framework to a multi-agent setting for distributed, resource-constrained coordination, with a focus on hyperlocal inland flooding. Future directions include incorporating diverse data sources and enabling long-term infrastructure planning through simulation-driven resilience analysis.

Students: Umar Faruque, Jason Marquez, John Lee

J. Clayton, S. Ahearn, F.  Lotfi-Jam, A. Raja, C. Fisher, A. Townsend, W. Ju, et al. 2024. “Developing a Digital Twin for ClimateAdaptation.” Workshop Report. Cornell Mui Ho Center for Cities, Ithaca, NY.