Use of Machine Learning to predict the load bearing capacities of steel structural systems

Abstract

This study explored the effectiveness of machine learning (ML) and artificial intelligence (AI) methods to predict the load bearing capacities of steel portal frame structures. For that, the predictive potential of Deep Neural Network and two tree-based ML models (XGBoost and GBRT) were evaluated. Two portal frame configurations were taken into consideration, and dataset was created using automated parametric Finite Element Method (FEM) analysis conducted in ABAQUS software. Additionally, this parametric script was developed using python. Moreover, global dimensions of portal frames and cross sectional and geometric properties of member segments of these frames were taken as input features, while corresponding load proportionality factors (LPFs) obtained from FEM analysis were set as outputs. Tree-SHAP (Shapley Additive Explanation) analysis has been employed to explain the underlying reasoning behind predictions for tree-based ML models. Global SHAP analysis revealed the specific input features of the global dimension of the frame, columns, top chord, bottom chord, and braces that highly impact the output LPF. Additionally, contribution of each input feature for a specific instance was identified from the local SHAP explanation. To improve the DNN model's predictive performance, the applicability of the local feature embedding technique was also investigated along with the usage of transformation activation functions to overcome the research gap in previous literature. To identify the optimum architecture of the main neural network for the DNN model, the number of hidden layers and the neuron amount in each of these layers were taken as variables. According to the results, NN architecture with 4 hidden layers with 56 neurons in each layer was obtained as optimal to predict the LPF accurately. Additionally, combination of Swish activation function and Stochastic Gradient Decent (SGD) optimizer contributes to the prediction accuracy improvement. Considering hyperparameter optimization of tree-based ML models to improve the prediction accuracy, a subsampling rate of 0.5, a tree depth of 8, and a regularization λ of 1.2 were obtained as optimal hyperparameters of the XGBoost model, while a subsampling rate of 0.8 and a tree depth of 7 were found to be optimum for the GBRT model. Comparison revealed that the DNN model provides the best performance where overall prediction accuracy is greater than 95.5% and R² is greater than 0.97 for the 2 considered topologies of the portal frame. Additionally, XGBoost and GBRT offer acceptable accuracy (overall accuracy > 93% and R² > 0.96). Moreover, both DNN and XGBoost models indicated high competitive performances (overall accuracy > 93.7%) when predicting LPF of optimized arrangements of portal frames. Furthermore, the DNN model is a data-efficient solution and it can be adapted in data-scarcity environments since it demonstrated highly acceptable performance (overall accuracy > 94.2%, R² > 0.98) even at 1000 training data points. It was identified that the local feature embedding technique and adaptation of the Swish transformation activation function significantly improve the DNN model performances compared to previous literature. When dealing with tasks such as direct design method, optimization and reliability analysis, these ML models are suitable alternatives for replacing FEM analysis and provide faster solutions with acceptable accuracy.