This study presents a machine learning-based method to predict seismic responses for building structures by integrating structural and seismic properties. In the presented method, seismic intensity measures (SIMs) representing the characteristics of seismic waves are introduced to predict seismic responses. These SIMs include indicators that represent the time domain information, frequency domain information, and energy characteristics of seismic waves, and are used as the inputs of an artificial neural network (ANN) for response prediction. To reflect the structural response characteristics to seismic waves with various frequency properties in the ANN model, the structural properties are also incorporated as the input information in the prediction model. The maximum inter-story drift ratio of the structure is selected as the target seismic response and set as the output of the prediction model. Separate ANN models are built for linear and nonlinear systems, respectively. To train prediction models, datasets are generated via analysis of a large number of seismic waves for numerous linear and nonlinear systems with various characteristics. Then, the prediction performance of ANN models is evaluated using datasets that consist of the SIMs, the structural properties, and the corresponding seismic response. In addition, response prediction performance is comparatively examined according to changes in the types of properties used as the inputs of the prediction model. Furthermore, response prediction performance according to ANN architecture configuration changes is also investigated in detail.
Read full abstract