MechE Colloquium: Machine-learning based constitutive modeling

Abstract: 

Recurrent neural network (RNN) models are emerging as the most promising deep learning technique to describe the three-dimensional stress-strain behaviors of elasto-plastic materials under complex loading scenarios. At the same time, there is growing evidence that RNN formulations for language processing (e.g. LSTMs, GRUs) may lead to erroneous results when they are used in the context of constitutive modeling. We present the development of mechanics-specific RNN formulations that satisfy important mathematical requirements such as self-consistency by construction. The resulting Minimal State Cells (MSCs) are compact networks that allow for the choosing the number of state variables independently from the degree of mathematical flexibility. This decoupling enables the models to detect the optimal number of state variables for a given constitutive modeling problem. We show that MSCs convincingly reproduce a variety of material behaviors and recognize important governing mechanisms. In a first application, we make use of RNNs to come up with a computationally efficient surrogate model to represent crystal plasticity. In a second application, we also demonstrate their ability to capture rate-dependent material behavior. To be able to train RNNs directly from experiments, we discuss the potential of transfer- and multi-task learning approaches. Furthermore, we pursue the development of automated mechanical testing systems to generate “big” experimental data for the training of data-driven plasticity models. 


Biography: 

Professor Dirk Mohr currently holds the Chair of Artificial Intelligence in Mechanics and Manufacturing at ETH’s Department of Mechanical and Process Engineering. He joined the faculty of ETH in 2015 after heading the Experimental Dynamics Group at the Solid Mechanics Laboratory at Ecole Polytechnique (France). He was educated in Structural and Computational Mechanics at the University of Karlsruhe (Germany), the Ecole Nationale des Ponts et Chaussées (France) and the Massachusetts Institute of Technology (USA) where he received his PhD in Applied Mechanics in 2003. He is Associate Editor of the International Journal of Solids and Structures (IJSS) and the International Journal of Impact Engineering (IJIE). He also serves on the editorial boards of the journals Strain, Journal of Manufacturing and Materials Processing and the International Journal of Plasticity. His research focusses on developing experimentally-​validated computational models to enable the optimal design and manufacturing of lightweight materials and structures that are subject to extreme loading conditions in real-​life applications.