Machine learning‐based peak ground acceleration models for structural risk assessment using spatial data analysis

Abstract

AbstractPredicting peak time‐domain ground‐motion parameters, such as peak ground acceleration (PGA), peak ground velocity, and peak ground displacement at a specific location, is challenging because of the limited number of recorded ground motions and the complexity of ground‐motion prediction equations. This study presents a novel approach that integrates a geographic information system with a spatial data analysis‐based machine learning PGA prediction model to overcome these challenges and predict PGA classes as a function of the PGA of the respective seismic stations, interstation distance of the seismic stations, and time‐average shear‐wave velocity in the upper 30 m of the target station. The proposed spatial data analysis‐based machine learning approach demonstrated the ability to generate satisfactory results in a short period. To account for the spatial dependencies of the variables, a feature selection method for spatial data using mutual information‐based feature selection was proposed, which provides a well‐prepared spatial matrix for machine learning algorithms. This study evaluates the performance of the model using various machine learning algorithms, including Random Forest, Naïve Bayes, K‐Nearest Neighbors, Decision Tree, AdaptiveBoost, Random Undersampling Boost, Extreme Gradient Boost (XGBoost), and Categorical Boost (CatBoost). Among these, XGBoost and CatBoost performed better than the other methods and yielded fairly accurate results. The models were validated using K‐Fold cross‐validation, and the Wilcoxon signed‐rank test was used for comparison. The spatial data analysis‐based machine learning models, particularly XGBoost and CatBoost, achieved high‐accuracy rates in classifying the PGA levels of 99.1% and 98.9%, respectively. Hyperparameters for the XGBoost model were tuned through GridSearchCV. Tree‐based models outperformed parametric models, indicating complex non‐linear spatial relationships, and by combining spatial feature selection with machine learning models demonstrated improved performance. Additionally, real‐time applications of spatial data analysis‐based machine learning PGA prediction models were used to estimate the seismic vulnerability of postulated concrete box‐girder bridges in San Fernando, thus providing insights into damage probabilities based on predicted PGA values.