Machine learning materials physics: Deep neural networks trained on elastic free energy data from martensitic microstructures predict homogenized stress fields with high accuracy

We present an approach to numerical homogenization of the elastic response of microstructures. Our work uses deep neural network representations trained on data obtained from direct numerical simulation (DNS) of martensitic phase transformations. The microscopic model leading to the microstructures is based on non-convex free energy density functions that give rise to martensitic variants, and must be extended to gradient theories of elasticity at finite strain. These strain gradients introduce interfacial energies as well as coercify the model, enabling the admission of a large number of solutions, each having finely laminated microstructures. The numerical stiffness of these DNS solutions and the fine scales of response make the data expensive to obtain, while also motivating the search for homogenized representations of their response for the purpose of engineering design. The high-dimensionality of the problem is reduced by training deep neural networks (DNNs) on the effective response by using the scalar free energy density data. The novelty in our approach is that the trained DNNs also return high-fidelity representations of derivative data, specifically the stresses. This allows the recapitulation of the classic hyperelastic response of continuum elasticity via the DNN representation. Also included are detailed optimization studies over hyperparameters, and convergence with size of datasets.

K. Sagiyama, K. Garikipati · 2019