Modelling of QSTE340TM steel FCG curve by neural network

Abstract

Abstract. This study deals with the modelling of fatigue crack growth rate of QSTE340TM steel by means of neural network. This material is a thermomechanically rolled high strength, low alloy steel with yield strength at least 340 MPa, combining strength, weldability, and excellent cold-forming properties. It’s a go-to material for automotive and structural applications where durability and lightweight design are critical. The classical deterministic methods of fatigue crack growth rate assessment are often quite expensive and require a well-equipped testing facility. In contrast, in recent decades, the methods of machine learning became widespread thanks to their ability to discovery unobvious data-driven dependencies. Machine learning is a part of artificial intelligence. It is trained on the existing data and improves with time without being explicitly programmed. The experimental dataset was taken from open scientific sources. It contained the fatigue crack growth rate data for four stress ratios of 0.1, 0.3, 0.5, and 0.7. The input data comprised of the following features: stress intensity factor range DK (MPa√m), fatigue crack growth rate da/dN, and stress ratio R. The target feature was da/dN, and the two rest were treated as input features. The model was shown only the data with stress ratio R that was equal to 0.1, 0.3, and 0.7. The model was tested on fatigue crack growth rate data with stress ratio R equal to 0.5. The neural network model in the form on multilayer perceptron was built. The model contained two hidden layers with 100 and 80 neurons, respectively. The activation function was RELU. The solver was chosen as Adam, and maximum iterations parameter was 500. The errors of the model were as follows: MSE = 1.238e-10, MAE = 8.594e-06, and R² = 0.95346. From the obtained results, neural network gives quite accurate prediction results and can solve such kinds of problems.