Efficiency and explainability of design-oriented machine learning models to estimate seismic response, fragility, and loss of a steel building inventory

Machine learning (ML) has recently been used as an efficient surrogate to estimate different steps of performance-based earthquake engineering (PBEE), from dynamic structural analysis to fragility and loss assessments. However, due to the varied data, models, and features in existing literature, the relative efficiency of ML models across different PBEE steps remains unclear. Additionally, the black-box nature of advanced ML algorithms limits their ability to provide design-oriented insights, hindering the broader application of ML in PBEE-based design. This study provides a comprehensive comparison of the accuracy and explainability of design-oriented ML models across different steps of PBEE using a consistent database of 621 steel moment frames with varying designs and geometry. Eight ML algorithms were used in a careful training workflow comprising feature selection, hyperparameter tuning, cross-validation, and model inference. The sensitivity of model accuracy to representative PBEE outputs—maximum responses, median fragility, and expected annual loss—was assessed using statistical measures. In addition, the explainability of the best models for each step was examined to explore the relationship between design parameters and the corresponding PBEE output. The results show that while ML models can reasonably map design parameters to all different PBEE outputs, models accuracy was higher for drift responses, median fragilities, and component-based loss metrics. In addition, the optimal algorithm remained the same across different PBEE steps, where support vector machines and random forests provided the highest accuracy with an average R² of 0.93 and 0.91 over different outputs on the test set. Although the selected feature sets varied across outputs and algorithms, height, number of stories, fundamental period, and the minimum of the beams’ moment of inertia were influential for both models and notably affected different PBEE outputs.