A Novel ANN Technique for Fast Prediction of Structural Behavior

Abstract

In recent decades, different concepts of machine learning (ML) have found applications in solving many engineering problems. Less time consumption in performing analyses, better optimization of the quality-price ratio and maintaining high model accuracy are just some ML advantages compared to traditional modeling procedures. There are currently a significant number of pre-trained machine learning models based on classification or regression tasks. However, there is a tendency to improve them through the implementation of the transfer learning (TL) approach. This article proposes an upgrade of the existing, pre-trained artificial neural network (ANN) model for the evaluation of the ultimate compressive strength of square concrete-filled steel tubular (CFST) columns. The aim of the improved TL model is to adapt to the problem of predicting the axial capacity of rectangular CFST columns in a more optimal way. The attractiveness of the TL is reflected through the possibility of overcoming certain shortcomings of classical models. Quick adaptation to the problem with small modifications of the existing surrogate model, better overcoming of potential overfitting even with a small dataset, and improved convergence towards the required solutions are some of the advanced TL strategies. The robustness of the proposed model was demonstrated through verification with experimental solutions and validation with the Eurocode 4 (EC4) design code. The application of such innovative paradigms can also be ensured for other research fields in a similar manner.