Prediction of plastic yield surface for porous materials by a machine learning approach. (December 2020)

• Record Type: Journal Article
• Authors: Shen, W.Q.; Cao, Y.J.; Shao, J.F.; Liu, Z.B.
• Is Part Of: Materials today communications. Volume 25(2020)
• Publication Date: 2020-12
• Subjects: Yield surface -- Artificial neuron network -- Machine learning -- Porosity -- Porous material -- Micro-mechanics; Materials science -- Periodicals; 620.11
• DOI: 10.1016/j.mtcomm.2020.101477
• Languages: English
• ISSNs: 2352-4928

Abstract

Highlights: The effect of porosity on the macroscopic behaviour of porous materials is investigated by a machine learning approach. The main existing strength criteria of the studied material are evaluated and compared. The artificial neuron network (ANN) approach used in this study overcomes the difficulties encountered in the different analytical methods. New finite element solutions are carried out with a wide range of porosity. The ANN predictions are validated by comparing with the FEM results, which ameliorates fundamentally the analytical criteria. The proposed approach is quite easy to apply and provides a sound background for various future works. Abstract: The present paper focuses on the prediction of effective plastic yield surface of porous materials having a von Mises type solid matrix. Some typical explicit yield criteria obtained by different analytical homogenization methods are briefly reviewed and evaluated by using numerical results obtained from direct finite element simulations with different values of porosity. Each criterion has its own advantage and weakness. In order to get a better prediction, the Artificial Neuron Network (ANN) algorithm is adopted specially for the prediction of macroscopic yield stress of porous materials, seen as a regression problem with two input parameters and one output value. For the training purpose which is a key step in the ANN approach, new numerical results are presented in the present work with a wide range of porosity and of macroscopic stress triaxiality. Based on these data, the ANN approach is trained and it converges quickly. Then the ANN predictions are compared with numerical test data, a good agreement is found for all loading cases. Comparing with the existing yield criteria, the prediction given by the ANN approach is much more accuracy and easy to apply.