Speaker
Description
Composite materials often exhibit complex anisotropic mechanical responses governed by their microstructural geometry and material phase distributions. Capturing these responses and their dependence on microstructural parameters is thus a critical challenge. In this work, we address this problem through a data-driven model-discovery framework that infers interpretable constitutive relationships directly from homogenized stress-strain data. By sparsely selecting relevant invariants of the Cauchy-Green deformation tensor, this approach aims to construct a strain energy density function that naturally encodes anisotropy while fulfilling thermodynamic constraints.
Previous works have introduced surrogate modeling techniques, such as neural networks, to approximate the homogenized response of a Representative Volume Element (RVE) as functions of various design parameters. These methods can learn anisotropy types and preferred directions but often yield “black-box” models that lack interpretability and require extensive hyperparameter tuning. In contrast, the proposed model-discovery approach delivers interpretable constitutive relationships by design and facilitates the direct inference of anisotropic features.
With the discovered model at hand, one obtains a surrogate that can be readily integrated into computational frameworks---potentially offering an alternative to more expensive concurrent multiscale methods like FE$^2$. We demonstrate the effectiveness of our approach on both synthetic and homogenized datasets, showing that it yields improved interpretability, reduced computational costs, and reliable material responses compared to existing methodologies.
