Seismic vulnerability modelling of building portfolios using artificial neural networks
Petros Kalakonas, Vítor Silva
Abstract
ABSTRACT The incorporation of machine learning (ML) algorithms in earthquake engineering can improve existing methodologies and enable new frameworks to solve complex problems. In the present study, the use of artificial neural networks (ANNs) for the derivation of seismic vulnerability models for building portfolios is explored. Large sets of ground motion records (GMRs) and structural models representing the building stock in the Balkan region were used to train ANNs for the prediction of structural response, damage and economic loss conditioned on a vector of ground shaking intensity measures. The structural responses and loss ratios (LRs) generated using the neural networks were compared with results based on traditional regression models using scalar intensity measures in terms of efficiency, sufficiency, bias and variability. The results indicate a superior performance of the ANN models over traditional approaches, potentially allowing a greater reliability and accuracy in scenario and probabilistic seismic risk assessment.