Optimizing artificial neural networks for mechanical problems by physics-based Rao-Blackwellization: example of a hyperelastic microsphere model

Abstract

The Rao-Blackwell scheme provides an algorithm on how to implement sufficient information into statistical models and is adopted here to deterministic material modeling. Even crude initial predictions are improved significantly by Rao-Blackwellization, which is proven by an error inequality. This is first illustrated by an analytical example of hyperelasticity utilizing knowledge on principal stretches. Rao-Blackwellization improves a 1-d uniaxial strain-energy relation into a 3-d relation that resembles the classical micro-sphere approach. The presented scheme is moreover ideal for data-based approaches, because it supplements existing predictions with additional physical information. A second example hence illustrates the application of Rao-Blackwellization to an artificial neural network to improve its prediction on load paths, which were absent in the original training process.