Speaker
Description
Machine Learning (ML) methods have proven their potential in material modeling with the goal to replace analytical constitutive models by machine-learned relationships [1]. Apart from the possible simplification of existing approaches, the application of ML is motivated by the goal to accurately characterize materials by a quick and integrated, automatized process [2]. The present talk particularly focuses on modeling of cyclic plasticity, assuming the von Mises flow rule and the associated plasticity framework as a basis. With the use of pseudo-experimental data, it is shown that lightweight Feed Forward Neural Networks can completely replace the assumptions on the mixed kinematic and isotropic hardening. The numerical investigations performed with different network architectures indicate that a direct training of a six-dimensional stress tensor is not feasible and show that the reduction of the problem complexity is key for a successful training and a low generalization error. One of the steps to achieve that is learning the plastic multiplier and the evolution of backstresses. The training data set is generated by random walks constructed by Gaussian Processes which then is optimized to reduce the overlap between its entries. The input strains are transformed into the principal axes to exploit isotropy for further efficiency gains and physics-informed regularization [3] is applied to ensure the accuracy and stability of the ML model. An optimal lightweight architecture is determined by systematically varying the number of hidden layers, neurons and activation functions. In terms of computational effort necessary for training and prediction, the final configuration has shown a significant improvement compared to the alternative approaches based on the application of recurrent Neural Networks with significantly reduced training data. The validation of the model has been performed on the example of dual-phase steels and indicates high accuracy and cyclic stability of the results.
[1] Hildebrand S., Klinge, S. Hybrid data-driven and physics-informed regularized learning ofcyclic plasticity with Neural Networks. Mach. Learn.: Sci. Technol. 5 045058 (2024). DOI 10.1088/2632-2153/ad95da
[2] Hildebrand S., Friedrich, J.G., Mohammadkhah, M., Klinge, S. Coupled CANN-DEM Simulation in Solid Mechanics, Machine Learning: Science and Technology, 2024 (accepted)
[3] Hildebrand, S., Klinge, S. Comparison of neural FEM and neural operator methods for applications in solid mechanics. Neural Comput \& Applic 36, 16657–16682 (2024). DOI 10.1007/s00521-024-10132-2
