Classification of buildings' potential for seismic damage using a machine learning model with auto hyperparameter tuning

The research on the application of machine learning (ML) methods in the field of earthquake engineering shows a continuous and rapid progress in the last two decades. ML methods and models belonging to the category of supervised, unsupervised and semi-supervised learning are applied for the assessment of seismic vulnerability of structures and estimation of the expected level of seismic damage. These models lead to the classification of seismic damage into predefined classes through the extraction of patterns from data collected from various sources. However, the lack of detailed knowledge can affect their performance and ultimately reduce their reliability, as well as the generalizability that should characterize them. Towards this direction, the present paper attempts to compare and evaluate for the first time the ability of an extensive number of ML methods in the correct classification of R/C buildings at the first stage pre- and post-seismic inspection considering three seismic damage categories. A database consisting of 5850 training samples is used for this evaluation. This database is generated by solving 90 R/C buildings for 65 actual seismic excitations applying nonlinear time history analyses. For each one of the training samples the maximum interstory drift ratio is calculated as a damage index. In addition, a major contribution of this paper is the presentation and extensive documentation of the procedures required for the preprocessing of the data. Finally, an auto hyperparameter tuning method for the winning algorithm is proposed, so that the hyperparameters are automatically optimized utilizing Bayesian Optimization. The most significant conclusion extracted is that the studied ML algorithms extract very different classification results. In addition, the Support Vector Machine - Gaussian Kernel algorithm extracted the most accurate results of all the studied ML algorithms.