Prediction of ductile damage evolution based on experimental data using artificial neural networks. (5th December 2022)

• Record Type: Journal Article
• Authors: Schowtjak, A.; Gerlach, J.; Muhammad, W.; Brahme, A.P.; Clausmeyer, T.; Inal, K.; Tekkaya, A.E.
• Is Part Of: International journal of solids and structures. Volume 257(2022)
• Publication Date: 2022-12-05
• Subjects: Damage evolution -- Machine learning -- Artificial neural networks -- Metal forming; Mechanics, Applied -- Periodicals; Structural analysis (Engineering) -- Periodicals; Elastic solids -- Periodicals; Mécanique appliquée -- Périodiques; Constructions, Théorie des -- Périodiques; Solides élastiques -- Périodiques; Elastic solids; Mechanics, Applied; Structural analysis (Engineering); Periodicals; 624.18
• DOI: 10.1016/j.ijsolstr.2022.111950
• Languages: English
• ISSNs: 0020-7683

Abstract

Abstract: A model for the evolution of ductile damage in the sense of void fractions using artificial neural networks (ANN) is proposed. In contrast to constitutive damage models, the damage prediction is solely based on experimental data and has no underlying assumptions for the damage evolution law. This guarantees that the experimental observations are captured correctly. High resolution experimental void data obtained by scanning electron microscopy with a minimum single void area of 0.02 μ m 2 are used as training data. The loading state is obtained from finite element simulations. The equivalent plastic strain is used to describe the load amplitude and the triaxiality as well as the normalised Lode angle are utilised to characterise the loading type. Different strategies for the training of the model as well as the prediction are analysed. The model is used to visualise the loading state-dependent damage evolution. Furthermore, it is applied to two different bending processes. It is shown that the prediction quality highly depends on the experimental and numerical data used for training. If the loading states of the application problem are within the domain of the training data, the prediction quality is good and even better than constitutive models used in literature.