Methodology for effective and efficient regional seismic retrofit using machine learning and stochastic optimization

Abstract

A major challenge in specifying design requirements for policy-based seismic retrofits is ensuring that the desired performance enhancements are achieved across a portfolio of buildings with diverse structural characteristics and hazard. To address this challenge, a modular optimization methodology is developed. A prediction module uses machine learning-based surrogate models as compact statistical links between building structural characteristics and seismic performance outcomes. The optimization module uses regionally targeted objective functions to determine the retrofit enhancements that achieve the most desirable aggregated performance outcome for the portfolio of buildings. Lastly, the evaluation module benchmarks the performance of the optimized retrofit scheme against the existing inventory as well as more conventional retrofit approaches. The ability of the framework to balance the need for effective (improved performance) and efficient (minimizing cost) retrofits is demonstrated by applying it to the inventory of buildings under the purview of the Los Angeles Soft-Story Ordinance.