Microstructure-based Variational Neural Networks for Robust Uncertainty Quantification in Materials Digital Twins

Aleatoric uncertainties - irremovable variability in microstructure morphology, constituent behavior, and processing conditions - pose a major challenge to developing uncertainty-robust digital twins. We introduce the Variational Deep Material Network (VDMN), a physics-informed surrogate model that enables efficient and probabilistic forward and inverse predictions of material behavior. The VDMN captures microstructure-induced variability by embedding variational distributions within its hierarchical, mechanistic architecture. Using an analytic propagation scheme based on Taylor-series expansion and automatic differentiation, the VDMN efficiently propagates uncertainty through the network during training and prediction. We demonstrate its capabilities in two digital-twin-driven applications: (1) as an uncertainty-aware materials digital twin, it predicts and experimentally validates the nonlinear mechanical variability in additively manufactured polymer composites; and (2) as an inverse calibration engine, it disentangles and quantitatively identifies overlapping sources of uncertainty in constituent properties. Together, these results establish the VDMN as a foundation for uncertainty-robust materials digital twins.

💡 Research Summary

The paper introduces the Variational Deep Material Network (VDMN), a physics‑informed surrogate model that extends the deterministic Deep Material Network (DMN) to quantify aleatoric uncertainties arising from microstructural variability. In VDMN, key DMN parameters—phase volume fractions and interface orientations—are promoted to Gaussian variational distributions characterized by means and covariances. This probabilistic embedding transforms the hierarchical laminate homogenization process from a fixed mapping into a stochastic operator that outputs a distribution of effective material responses.

To propagate these uncertainties through the network, the authors develop an analytic scheme based on first‑order Taylor expansion for the mean response and second‑order corrections for the covariance, with Jacobians and Hessians obtained via automatic differentiation. The procedure is applied recursively across the binary‑tree architecture, yielding closed‑form expressions for the mean and covariance of the homogenized stiffness tensor.

Training is performed by minimizing a negative log‑likelihood loss that compares the predicted Gaussian distribution (mean and covariance) with observed elastic homogenization data. The loss includes a regularization term enforcing conservation of total phase fraction, thereby preserving physical interpretability of the laminate weights. Because the likelihood is evaluated on the tangent space of the Riemannian manifold of positive‑semi‑definite stiffness tensors, the method naturally handles heteroskedastic data.

VDMN supports two operational modes: (1) an analytic mode that directly uses the closed‑form propagation for rapid inference, and (2) a sampling mode that draws ensembles of deterministic DMNs from the learned variational priors, enabling Monte‑Carlo uncertainty estimation and seamless integration with existing nonlinear DMN implementations.

The authors validate VDMN in three stages. First, on a synthetic dataset of 30 stochastic two‑phase microstructures generated by spinodal decomposition, VDMN learns accurate uncertainty representations in the linear elastic regime and successfully extrapolates to nonlinear loading conditions without additional training data. Second, experimental validation on additively manufactured polymer‑matrix composite tensile bars demonstrates that VDMN can predict the full distribution of nonlinear stress‑strain responses, offering a forward‑prediction capability useful for reducing experimental campaign costs. Third, an inverse calibration study shows that VDMN can disentangle overlapping sources of aleatoric uncertainty—such as phase stiffness variability and volume‑fraction fluctuations—from noisy homogenized measurements, providing quantitative estimates of each contributor.

Overall, VDMN merges microstructure‑based physics with probabilistic inference, delivering a robust, generalizable platform for uncertainty‑aware materials digital twins. Its ability to learn from limited linear data, extrapolate to nonlinear regimes, and separate multiple uncertainty sources makes it attractive for high‑performance material design, real‑time digital‑twin monitoring, and reliability‑driven optimization in advanced manufacturing contexts.