Deep learning-based prediction of the hysteretic behavior of buckling-restrained braces for seismic design using analysis-of-mean-based optimal hyperparameters

This study aims to develop a recurrent neural network (RNN)-based framework for predicting the nonlinear hysteretic behavior of phase-change-material-filled buckling-restrained braces (PCM-filled BRBs) in order to reduce reliance on repeated large-scale experiments and associated testing costs. Representative RNN architectures, including Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks, are investigated. To systematically configure the RNN models, an orthogonal array–based design of experiments is employed, incorporating key hyperparameters such as input window size, batch size, number of layers, and number of neurons. Model prediction accuracy is evaluated using mean absolute error (MAE) and root mean squared error (RMSE), and hyperparameter sensitivity is assessed through analysis of means and range analysis. Optimal hyperparameter combinations are identified for both LSTM and GRU models. Using these optimal settings, the models predict hysteresis responses for both horizontally installed and inclined BRBs. The results indicate that the GRU model outperforms the LSTM model in terms of MAE and RMSE, and that both models can reproduce the complex force–displacement hysteresis of PCM-filled BRBs without relying on explicitly defined phenomenological constitutive models or the iterative parameter calibration typically required in nonlinear finite element analysis.