MACHINE LEARNING FOR SEISMIC-INDUCED DAMAGE ESTIMATION OF STEEL TANKS

Abstract

Aboveground steel storage tanks are large vessels employed to store various liquids, including water, food, fertilizers, oil, and other hazardous chemicals. The damage and collapse of storage tanks can generate long-lasting consequences on built environment and communities. The seismic behavior of storage tanks is often evaluated employing fragility functions obtained simulating the tank under a variety of ground motion records. However, the computational demand of high-fidelity simulation models makes risk assessment a burdensome task. The use of data-driven surrogate models could represent a suitable solution to evaluate the seismic vulnerability of steel tanks rapidly. In this context, this paper presents an open dataset composed of 204 aboveground cylindrical steel liquid storage tanks with different geometric properties. The dataset was assembled based on past earthquake reconnaissance reports. In the dataset, different types of damage experienced by the steel tanks are divided into four classes, ranging from no damage to complete failure. Eight different machine learning algorithms are trained to predict the damage class of a steel tank as a function of its geometric properties and seismic excitation parameters. Stratified 5-fold cross-validation is used to split the dataset into training and testing subsets and to assess the prediction capability of the machine learning models. Results showed that the Support Vector Machine algorithm yielded the most accurate predictions, followed by Random Forest, XGBoost, and LightBoost. Overall, the paper demonstrated the feasibility of using machine learning models to predict the damage level of steel liquid storage tanks subjected to seismic hazard.