Predicting ground motion intensity measures in Japan using a hybrid artificial neural network integrated with Taguchi optimization

Despite the availability of extensive strong-motion recordings in Japan, machine learning (ML)-based ground-motion models (GMMs) capable of consistently predicting a broad range of intensity measures (IMs) remain limited. This study develops a hybrid artificial neural network (ANN) integrated with Taguchi optimization (TO) to predict significant duration ( D s5-75 and D s5-95 ), Arias intensity ( AI ), cumulative absolute velocity ( CAV ), peak ground acceleration ( PGA ), peak ground velocity ( PGV ), and 5% damped pseudo-spectral acceleration ( PSA ) over periods from 0.01 to 10 s. The TO framework efficiently identified near-optimal ANN hyperparameters using only 25 experiments, and the resulting model achieved lower test-set root-mean-square-error ( RMSE ) than models optimized using Bayesian optimization and random search under the same computational budget. A mixed-effects framework was incorporated to quantify between-event and within-event variability, with residual analyses showing no notable systematic bias. The proposed model generally exhibited comparable to lower variability than existing Japan-specific GMMs, while a partially non-ergodic formulation further reduced aleatory variability by separating repeatable site effects. Despite having no predefined functional form, the model reproduced physically consistent magnitude scaling, distance attenuation, and site amplification behavior. Inter-IM residual correlations were quantified and compared with previous studies, while Shapley Additive exPlanations (SHAP) analysis provided physical interpretation of the learned relationships and relative importance of input parameters. Overall, the proposed approach provides a unified, computationally efficient, uncertainty-aware, and interpretable ML framework for predicting multiple ground-motion IMs.