Representative feature contributions when predicting the seismic collapse of steel buildings using an ensemble neural network

This study implements an ensemble neural network (ENN) to obtain representative and stable feature importance contributions to collapse prediction of steel moment resisting frames (SMRFs). The feature importance assessment includes global sensitivity analyses (GSAs) and feature extraction techniques. To construct the ENN, hundreds of neural network (NN) architectures are generated and an elite set of 50 NNs is initially obtained using a multi-criteria decision analysis (MCDA). A final elite set of 50 NNs is generated after applying a genetic algorithm to the initial elite set, which undergoes several iterations of crossover and mutation. To generate the dataset of SMRF collapse status, thousands of nonlinear time history analyses are carried out on frame systems ranging from 2 to 20 stories. The frames are based on five SMRF baseline systems with variability in input parameters that are randomly selected for each system to incorporate uncertainties in mass, structural damping, nonlinear spring backbone parameters, and ground motion parameters. Then, single NNs and the ENN generate surrogate models that are used to predict the SMRF collapse status. Thereafter, five feature analysis techniques are implemented on the ENN to determine the most important contributors to global collapse of the evaluated SMRFs, including two GSAs (Sobol and delta methods) and three feature importance extraction methods (permutation, integrated gradients, and Shapley adaptive explanation values). The GSA results were obtained for a single-instance (one SMRF) for each NN and, in the most comprehensive approach, as the average of applying this technique to 100 instances (i.e., 100 SMRFs). The study found that GSA results can fluctuate significantly based on the NN architecture and the data instances used in sampling. The use of a multi-instance-based ENN offered a more stable collapse predictor for SMRFs. The results consistently highlight ground motion parameters and elastic system parameters as the most critical contributors to collapse prediction of modern SMRF buildings.