Rapid seismic loss assessment of school buildings using adaptive basis vectors sampling for support vector regression

Accurate and rapid assessment of earthquake-induced economic losses across varying seismic intensities and the exploration of their relationship are crucial for quantifying structural seismic resilience. Traditional seismic loss assessments using the finite element method require computationally intensive nonlinear time-history analyses, while conventional machine learning methods demand substantial resources for training on large seismic response datasets. This paper proposes Adaptive Basis Vectors Sampling Guided Support Vector Machines for Regression (ABVS-SVMR) to overcome these limitations. ABVS-SVMR reduces computational complexity by adaptively sampling an optimal subset to construct a low-rank kernel matrix approximating the full matrix, improving training efficiency. Bench-marking against Least Squares SVMR (LS-SVMR) and Artificial Neural Networks (ANNs) using 18,438 seismic loss records from reinforced concrete frame school buildings demonstrated excellent predictive accuracy ($R^2>0.97$) for all models. Crucially, ABVS-SVMR achieved a 13-fold speedup over LS-SVMR and a 27-fold speedup over ANNs, demonstrating exceptional potential for rapid seismic loss assessment.