A mixture of experts approach to handle ambiguities in parameter identification problems in material modeling

Abstract

To simulate the mechanical behavior of a material, it is essential to calibrate the internal parameters of the used material model to experimental measurements. This is typically done in a trail-and-error approach by hand or automatically using optimization algorithms. As an alternative to trial-and-error, neural network-based approaches can be used to imitate the inverse mapping. This is usually realized in a grey-box model, combining neural networks, deterministic models, and domain knowledge. However, the proposed neural network-based approaches found in literature do not address the challenge that is posed when the parameter identification problem is non-unique. In the present paper this problem is discussed and an improved approach is proposed using a mixture of experts model. Mixture of experts is an ensemble technique based on a dynamically structured framework of submodels aiming to partition the non-unique problem into unique subtasks.