Causal Disaster System Modeling and Inference with Multi-resolution Score-Based Variational Graphical Diffusion (Papers Track)
Xuechun Li (Johns Hopkins University); Shan Gao (Johns Hopkins University); Susu Xu (Johns Hopkins University)
Abstract
Complex systems with intricate causal dependencies present significant challenges for accurate estimation, particularly in disaster modeling where multiple physical processes interact simultaneously. In earthquake scenarios, for example, accurately inferring the true states of cascading hazard latent variables while capturing their causal dependencies is crucial yet challenging. Existing methods struggle to handle varying data resolutions while capturing physical relationships and causal dependencies, especially when data comes from diverse sources with inconsistent sampling. Therefore, we introduce SVGDM: Score-based Variational Graphical Diffusion Model, which addresses these challenges through a novel integration of score-based diffusion models and causal graphical models. Our framework constructs individual stochastic differential equations (SDEs) for each variable at its corresponding native resolution, then couples these SDEs through a causal score mechanism where parent nodes inform the evolution of the child nodes. The framework enables a unified modeling of causal effects in causal disaster system, such as earthquake-induced cascading hazards, including ground shaking, landslides, liquefaction, and building damage. Through experiments on three major earthquakes (2020 Puerto Rico, 2021 Haiti, and 2023 Turkey-Syria earthquakes), we demonstrate improved prediction accuracy (> 0.93 AUC) and causal understanding compared to existing methods, while maintaining robust performance under varying levels of background knowledge availability.