Seismic damage prediction and simplified model of mega composite frame structures based on machine learning

To improve the efficiency of selecting structural input parameters for machine learning‐based earthquake damage prediction of mega composite frame structures (MCFSs), a 1/15‐scale structural model was designed and tested using a shaking table. The dynamic characteristics, displacements, strains, and structural damages of the model under different earthquake levels were analyzed. The correlations among the structural member damage, local damage, and overall damage under different peak ground accelerations were investigated. Based on the finite element analysis results, three algorithms, random forest, back propagation neural network (BPNN), and support vector machine, were used to predict the seismic damage indices of the structure. The study results show that the BPNN model can predict the seismic damage indices of a MCFS more accurately. The sensitivity of each input parameter to a seismic damage index of the structure was also studied. The research results show that plane‐size characteristic parameters such as the number of spans in the X and Y directions of the subframe are less sensitive to seismic damage index, while the cross‐sectional size of mega columns and the height of mega layers are more sensitive to seismic damage index. Lastly, with the parameters with lower sensitivity removed, the simplified prediction model for the input parameters still gives high prediction results for MCFSs, making it suitable for studying the seismic performance of such structures.