Prediction of Seismic collapse behavior of deep steel columns using Machine learning

Abstract

Steel Column Net (SCNet), a dataset of more than nine hundred experimental and numerical results of deep wide flange (W-shape) columns with different attributes is compiled. Three failure modes are distinguished in SCNet: local, global, and coupled modes while column rotation capacity is expressed in terms of its cumulative inelastic rotation until failure. The efficiency of five machine learning (ML) classification models is explored to identify the failure modes of columns subjected to combined axial and lateral loading in a randomly assigned test set from SCNet. Among the ML classification techniques used, support vector machine and decision trees provide the best performance with a prediction accuracy of 89%. The efficiency of four ML regression models is explored to predict the cumulative inelastic rotation until failure and categorize highly ductile behavior of the columns in the test set. Of these, the gaussian process regression exhibits superior performance with an accuracy of 87%. The performance of the ML regression models is compared with the current AISC highly ductile limits for W-shape columns and found to provide a 30% improvement in classification accuracy with respect to the number of correctly classified highly and non-highly ductile columns in the test set. Based on these results, it is suggested that machine learning algorithms that are continually updated with new experimental and computational data could inform future generations of design specifications.