Machine Learning for Constitutive Modeling on an Exascale Computing Platform

Michael Homel | 19-FS-050


We developed a versatile software tool to assess the feasibility of applying machine learning and artificial neural networks (ANN) to constitutive modeling. Specifically, we studied the performance of ANN models, trained with a variety of existing continuum constitutive models, to 1) address challenges in transitioning existing constitutive models to a form suitable for graphics processing units (GPUs), and 2) understand the application space where machine learning and ANN can replace or augment traditional constitutive models. Our results showed that a black-box treatment of the constitutive response is only feasible for models with a very low-dimensional input space. However, hybrid formulation coupling ANN and traditional approaches have significant potential. We presented implications for GPU-ready constitutive modeling and direct numerical constitutive modeling from mesoscale or experimental data.

Impact on Mission

Our work leveraged and advanced Lawrence Livermore National Laboratory's core competencies in high-performance computing, simulation, and data science. Our results expanded expertise within the Laboratory's computational geosciences group and can be applied broadly to Laboratory modeling projects in support of all mission focus areas.