Integrating ML-Based CGE Surrogate Models with Optimization for Enhancing Community Resilience

Abstract Natural hazards pose substantial risks to economic stability and community well-being. Understanding and mitigating these risks requires robust tools that can capture both the physical impacts of hazards and their economic consequences. This study presents a framework integrating machine learning (ML) and optimization with a computable general equilibrium (CGE) model to enhance community resilience to natural hazards, focusing on earthquake impacts. A penalized regression model serves as a surrogate to address CGE’s computational complexities. These predictive insights are embedded in an integer linear programming (ILP) optimization model, which identifies retrofit strategies that minimize the modeled economic disruption resulting from building damage. Applied to Salt Lake County (SLC), Utah, a seismically vulnerable region, the framework uses 5,000 earthquake simulations to assess economic impacts across 10 economic sectors and seven geographic regions, giving 70 sector-region combinations. The contribution of this work is the operational linkage of these components into a single decision-support workflow that connects physical earthquake damage scenarios, CGE-based economic impact modeling, an interpretable ML surrogate trained on CGE outputs, ML-derived sensitivity coefficients, and an ILP retrofit allocation model. This integration makes CGE-informed economic resilience analysis computationally tractable for city-scale retrofit planning, a use that a full CGE model cannot support directly within a combinatorial optimization over roughly 280,000 buildings. Using this workflow, we report a sensitivity analysis showing that retrofit priorities are not stable across economic outcomes: the sector-region losses that most influence domestic supply and employment differ from those that most influence household income and migration, and these differences are spatially uneven across Salt Lake County. A retrofit model driven only by direct physical loss or retrofit cost would not reveal this dependence. The framework therefore provides an interpretable and computationally feasible tool for outcome-specific retrofit prioritization.