|About this Abstract
||NUMISHEET 2022: The 12th International Conference on Numerical Simulation of 3D Sheet Metal Forming Processes
||Towards Machine-learning based Constitutive Modeling
||Colin Bonatti, Christian C. Roth, Vincent Grolleau, Dirk Mohr
|On-Site Speaker (Planned)
||Christian C. Roth
Machine learning provides a potentially powerful means to derive constitutive models from experimental data. Among the wide spectrum of deep learning models, recurrent neural networks (RNNs) are particularly suitable for plasticity modeling due to their built-in history variables. The shortcomings of standard RNNs are discussed before presenting a new mechanics-specific RNN formulation and their integration into explicit finite element software. It is shown that mechanics-based RNNs may act as universal material model. In particular, it will be shown that they are able to provide an accurate surrogate model of crystal plasticity (CP). CP-RNN models bring an extraordinary computational speed-up as compared to CP-FTT which enables the use of crystal plasticity models in industrial practice. In addition to developing mechanics-specific neural network architectures, new robot-assisted experimental procedures are presented that generate “big data” for the identification of machine-learning based plasticity and failure models from experiments.
||Definite: At-meeting proceedings