Multiscale Analysis of Composites Using Surrogate Modeling and Information Optimal Designs
Steven M. Arnold, Matthew Piekenbrock, Trenton M. Ricks, Joshua Stuckner · AIAA Scitech 2020 Forum · 2020
In this effort, a general workflow which leverages machine learning and recent advances in optimal experimental design theory is introduced to infer hidden properties from experimental observations. Optimal design provides the most efficient and informative tests and is especially useful when resources only permit a limited number of tests. The potential applicability of this field herein is broadened by utilizing machine learning models as efficient, accurate, and generalizable surrogate models to optimize over the experimental design space. This workflow is demonstrated by inferring the hidden lower length scale constituent fiber and matrix properties of composite materials from macroscale mechanical properties derived from the response of “virtual” laminated composite experiments. This method requires a model to map the relationship between the hidden and observable properties and an efficient surrogate model to make the optimization tractable. The virtual data, produced by multiple applications of the Generalized Method of Cells (GMC), consists of ‘point-wise’ properties extracted from composite laminate tensile response curves under uniaxial loading. The surrogate model is an artificial neural network trained to predict these pointwise properties (e.g., modulus, proportional limit stress, ultimate tensile stress) to a sufficiently high accuracy. The optimal experimental design framework proposed is a statistically principled approach designed to uncover information about microstructure-properties-performance relationships. A priori knowledge of the explicit connections between the virtual data and the microstructure properties (e.g., laminate lay-up sequence, fiber volume fraction, etc.) are exploited to demonstrate the fidelity of our method. In doing so, a verifiable information theoretic explanation of the relationships between composite response, microstructure, and constitutive properties for an arbitrarily diverse set of laminate composites is established.