An efficient deep learning approach for predicting the ultimate load and maximum lateral web deformation of unstiffened steel plate girders under patch loading

Abstract

The maximum web deformation, along with the ultimate load, plays a crucial role in estimating the behavior of unstiffened steel plate girders subjected to patch loading. While various machine learning (ML) approaches have been applied to predict the ultimate load of the girders based on experimental or numerical datasets, no research has been conducted to predict both the ultimate load and the maximum lateral web deformation simultaneously. Therefore, this study introduces an efficient method for predicting both factors using a multi-layer perceptron (MLP) model. Initially, a Python-based program was developed to create a finite element (FE) model of unstiffened steel plate girders under patch loading, validated against experimental data. This validated model was used to generate a dataset of 500 FE simulations with various girder geometries and material strengths. Subsequently, this dataset was employed to train an MLP model. The Differentiable Architecture Search (DARTS) method is then applied to optimize the MLP architecture to enhance the performance of the MLP model. The proposed DARTS-MLP model demonstrated high accuracy in predicting the ultimate load and maximum lateral web deformation, closely matching the true values in the test set. The DARTS-MLP’s predictions were compared with existing design codes and empirical formulae, further confirming its prediction accuracy. Finally, a web-based application was developed to apply the proposed DART-MLP model in practical design.