A well-posed variational approach to the identification and convergent approximation of material laws from boundary data

TL;DR

This paper introduces a variational approach for identifying material laws from boundary data in elasticity, establishing a well-posed framework, proving existence of solutions, and demonstrating neural network-based approximations with convergence guarantees.

Contribution

It formulates a novel variational control framework for material identification, proving existence and uniqueness conditions, and shows neural network approximations converge to true material laws.

Findings

Existence of optimal controls is proven within the proposed framework.

Maxout neural networks can approximate energy densities densely.

The approach requires only boundary measurements, not full-field data.

Abstract

We formulate the problem of material identification as a problem of optimal control in which the deformation of the specimen is the state variable and the unknown material law is the control variable. We assume that the material obeys finite elasticity and that the deformation of the specimen is in static equilibrium with prescribed boundary displacements. We further assume that the attendant total energy of the specimen can be measured, e.g., with the aid of the work-energy identity. In particular, no full-field measurements, such as DIC, are required. The cost function measures the maximum discrepancy between the total elastic energy corresponding to a trial material law and the measured total elastic energy over a range of prescribed boundary displacements. The question of material identifiability is thus reduced to the question of existence and uniqueness of controls. We propose a specific functional framework, prove existence of optimal controls and show that the question of material identifiability hinges on the separating properties of the boundary data. The proposed framework naturally suggests and supports approximation by maxout neural networks, i.e., neural networks of piecewise affine or polyaffine functions and a maximum, or join, activation function. We show that maxout neural networks have the requisite density property in the space of energy densities and result in convergent approximations as the number of neurons increases to infinity. Simple examples are also presented that illustrate the minimax structure of the identification problem.