WDS-Modality: A Physics-Informed Multi-Modal Agent Framework for Resilient Water Distribution System Assessment Using...

Water Distribution Systems (WDS) resilience assessment remains dependent on specialized knowledge and complex simulations, limiting accessibility for operational decision-making. Data-driven surrogate models reduce computational complexity but often lack physical consistency, while large language models (LLMs) cannot directly interpret hydraulic and topological information. This study proposes WDS-Modality (WDSM), a physics-informed multi-agent framework that integrates LLM reasoning with hydraulic simulation without fine-tuning. The framework combines a topology-aware knowledge representation, a physics-informed verification mechanism, and a multi-objective optimization module. WDSM is evaluated through district metered area (DMA) partitioning and boundary valve optimization, a representative task requiring the joint consideration of hydraulic constraints, network topology, and reliability trade-offs. Experiments on five benchmark networks show that WDSM adaptively identifies DMAs with 8-34 partitions and 12-80 candidate boundary pipes with 3-17 closed pipes. The optimized configurations reduce flow entropy in most networks while producing physically feasible trade-offs among hydraulic reliability, mechanical reliability, and energy-based resilience. In diagnostic evaluation, WDSM achieves an overall success rate of 85.71%, with 97.50% performance in both partitioning and visualization tasks, while maintaining zero physics violations. These results demonstrate that coupling language-based reasoning with physics-based simulation enables reliable, interpretable, and physically grounded WDS analysis.