Interventionally Consistent Surrogates for Urban Flood Resilience Planning

Abstract

Policy evaluation for urban flood resilience requires high-fidelity simulations of interdependent flood and transportation dynamics, but the computational cost of these models—the “Simulation Bottleneck”—prevents their use for scenario analysis and policy optimization. Standard machine learning surrogates offer the needed speed but are typically trained on observational data and fail in interventional scenarios, i.e. they are not interventionally consistent. This paper uses an Interventionally Consistent Surrogate framework grounded in causal abstraction theory. Using Batched Multi-Task Gaussian Processes structured to respect the simulator’s causal topology, we train a surrogate that approximately preserves the simulator’s response under policy interventions. We demonstrate the framework on Copenhagen’s inner city using a flood-transport simulator, achieving R2 > 0.92 across all three prediction tasks while providing calibrated uncertainty estimates.