Predicting Structural System Behavior via a Deep Surrogate Modeling Framework With Imbalanced Feature Learning

Abstract In recent years, deep learning has experienced rapid advancement and has been applied across numerous disciplines. Predicting the structural responses using a deep learning method involves several intricate challenges since structural datasets often suffer from data imbalance, where certain stress responses are more prevalent, potentially skewing the deep learning method’s training process and accuracy. To meet these challenges, we propose a deep surrogate modeling framework for predicting systems behavior of complex engineering systems that integrates imbalanced feature learning. This research develops a deep surrogate modeling framework by using a convolutional neural network (CNN) for predicting structural response behavior, which might have data imbalance issues. To mitigate data imbalance issues, statistical sampling and label transformation methods are examined to improve prediction accuracy. Specifically, stress responses in cantilever beams are considered to highlight the efficiency and applicability of the proposed framework. The proposed deep surrogate modeling framework not only demonstrates enhanced accuracy of predictions in complex engineering systems but also addresses the challenge of imbalanced data distributions. The proposed deep learning framework aims to overcome the challenges associated with predicting structural responses in complex engineering systems.