- Jeremy Templeton, Sandia National Laboratories, Albuquerque, New Mexico
Material models are used in a wide variety of applications at Sandia to predict a system’s response to intended and unintended loading configurations to ensure that its requirements are met. Most current material models are predicated on a given functional form derived from physical knowledge, with adjustable parameters to account for different material behaviors. All models involve some degree of approximation and therefore incur some degree of model form error. Such errors are exacerbated when using novel materials for which a deep physical understanding is lacking, for example, in materials output from advanced manufacturing processes. Such materials are characterized by a crystalline morphology, so our goal is to understand if and how machine-learning approaches can be leveraged to represent this phenomenology. This talk will present a Tensor Basis Neural Network approach, which enforces physical realizability on the network, to model both the stress/strain and plastic flow relationships from synthetic data. In addition to network architecture, we will explore how the application drives meta-parameter selection, data sufficiency needs, and model use.