Determination of ductile fracture properties of 16MND5 steels under varying constraint levels using machine learning methods

The current paper presents a machine learning method based on artificial neural network (ANN) model for the determination of ductile fracture properties of 16MND5 bainitic forging steel with various three-dimensional (3D) constraint conditions. A series of fracture test data with clamped single edge notched tension (SENT) specimens were used for model training and test. With the comprehensive analysis of prediction accuracy and extrapolation ability, a training strategy for ANN model was proposed including an artificially divided training set and the introduction of dropout layer. The artificial division makes the experimental samples in training set reduced by 40.7%, while the dropout layer prevents ANN model from overfitting caused by reduction of training data. Moreover, the deep nonlinear relationship between geometric dimensions (H/W, B/W, a/W) and ductile fracture properties was well learned by the ANN model. The average error of prediction is less than 11%. Finally, the proposed training strategy was extended to solve the fracture behaviors under varying thermal aging duration with saving training experimental samples by 53.8%. The results showed that the comprehensive interaction of in-plane constraint, out-of-plane constraint and thermal aging on the ductile fracture behaviors are well reproduced. Due to the good prediction performance, generalization and low training cost, the proposed training strategy can make the ANN model much more helpful for the solution of ductile fracture properties of different geometric dimensions in harsh environments.

Automated summary

The current paper presents a machine learning method based on artificial neural network (ANN) model for the determination of ductile fracture properties of 16MND5 bainitic forging steel with various three-dimensional (3D) constraint conditions. The proposed training strategy, which includes an artificially divided training set and the introduction of dropout layer, effectively reduces the training data and prevents overfitting. The ANN model shows good prediction performance and can be helpful in solving ductile fracture properties of different geometric dimensions in harsh environments.