Physic‐Law Integrated Neural Network for Nonlinear Seismic Demand Prediction

Abstract

ABSTRACT Due to the structural parameters diversity, material attributes nonlinearity, and ground motions uncertainty, predicting the elastic‐plastic seismic response of columns is challenging and crucial, particularly in near‐fault areas where significant damage can occur. Traditional machine learning (ML) models have demonstrated powerful capabilities for predicting structural seismic demand. However, their difficulty in accurately capturing latent system nonlinearity and the challenges associated with quantifying the impact of pulse effects on structural seismic demand complicate their application in practical engineering. This study proposes an efficient, high‐precision, and highly interpretable physic‐law‐integrated neural network (PLNN) method. It introduces a novel physic‐law model (PLM), which establishes the relationship between the normalized period (pulse to structural fundamental period ratio) and seismic demand. In addition, this research examines and quantifies the effects of column properties and pulse attributes on the characteristic coefficients of the PLM using an artificial neural network model (ANN). This method provides a PLNN model for estimating seismic demand based on the structural parameters, material attributes, and seismic characteristics by integrating the ANN model with the PLM. The ability and stability of the PLNN are evaluated by comparing its prediction performance with that of a traditional ML model. The results indicated that the proposed PLNN model maintains high prediction accuracy and significantly enhances computational efficiency. The PLNN model in this research requires 10 neurons to achieve optimal fitting goodness, one‐third of the number required by the corresponding ANN model. In addition, the accuracy of the PLNN model is twice that of the ANN model in small sample settings, indicating the stability of the PLNN.

Topics & Concepts