Bayesian-optimized interpretable surrogate model for seismic demand prediction of urban highway bridges

Urban highway bridges are crucial for regional economic development and population mobility, but they are vulnerable to seismic hazards. Accurately predicting the seismic demands of bridges is essential. Several commonly-used methods for seismic demand estimates, such as cloud analysis, incremental dynamic analysis, and multiple stripe analysis, are computationally expensive and subject to theoretical assumptions and convergence issues, especially for portfolios of bridges. Machine learning techniques can be effective, but their lack of interpretability hinders understanding of the surrogate model's predictions. To address these challenges, this study develops a Bayesian-optimized interpretable ensemble learning surrogate model for predicting and interpreting the critical seismic demands of urban highway bridges.Thousands of continuous prestressed concrete girder bridges with typical geometric configurations in China are generated, and 120 ground motion records with different frequency, waveform, and duration characteristics are chosen. The seismic demand dataset is established based on nonlinear time-history analyses for urban highway bridges, accounting for different types of uncertainties. The surrogate models are developed based on the XGBoost ensemble learning method, with hyperparameters automatically and efficiently determined by Bayesian optimization technique. Compared to the random search and grid search methods, Bayesian optimization is approximately 8 times and 29 times faster, respectively, and it also achieves the highest prediction accuracy. Meanwhile, the relationship of each hyperparameter in the surrogate model is also revealed. The prediction results of the proposed model are compared with other machine learning models. The SHAP algorithm is subsequently used to explain the model predictions regarding the feature importance, global interpretations, and feature dependency analysis. The study quantifies the contribution of each structural feature and seismic intensity measure to the selected critical seismic demand and identifies the interaction effects between different input variables in the prediction.