A Comparative Study of ML and FEM Models for the Prediction of Seismic Structural Behavior

Abstract

During the last several decades, the finite element method (FEM) is the most commonly used numerical method for performing seismic structural analysis. It requires careful structural modeling, but also the adjustment of small computation steps, especially for highly complex nonlinearities, which leads to significant time consumption, to ensure the stability and accuracy of the results. The development of new techniques based on machine learning (ML) models attracts considerable attention, due to the possibility to overcome the limitations of traditional techniques. This study presents an implementation of the long short-term memory (LSTM) deep-learning recurrent neural network (RNN) for the estimation of nonlinear seismic structural response. It is established in a data-driven manner. The LSTM model has shown considerable success in capturing structural response during nonlinear dynamic time-history (NDTH) analysis. Even in the case of an insufficient number of data, it shows better performance and greater adaptation to experimental results than the robust FEM model. In order to secure the consistency of the dataset for different ground motion records and increments, resampling and filtering of data is recommended. Such an innovative approach can enable the prevention of catastrophic consequences from devastating earthquakes. That can be achieved by fast and accurate pre-earthquake response prediction, damage state forecasting, and accelerated development of fragility curves based on previously conducted incremental dynamic analyses (IDA) or experimental tests. The predictive capabilities of the developed model were demonstrated through a comparative analysis of the behavior of two adjacent interacting unreinforced masonry structures (URM), tested on a shaking table. Even with the relative lack of experimental data, LSTM showed superiority over SAP2000 software. In all sequences of the time history records, the LSTM model gave closer results to the experimental ones than the SAP2000. The lack of a massive amount of data was observed as the main reason for the deviation of the results obtained by the LSTM model in certain time intervals, so the expansion of datasets is proposed for upcoming simulations. In future research, to model the dynamic response of different material and structural types, the formation of such models in a physics-driven fashion is recommended.