Answers are grounded in this paper and cite their sources.
Heterogeneous microstructures such as woven composites, sintered materials, and additively manufactured parts exhibit unavoidable morphological variability that can produce a significant spread of effective thermal responses. The goal of this work is a physics-based surrogate that learns the distribution of the effective conductivity over a family of statistically equivalent microstructures and propagates it to the temperature field of a component, without solving a microscale boundary value problem for each realization. To that effect, we extend the Variational Deep Material Network (VDMN) framework, originally developed for mechanical homogenization, to thermal conduction and integrate it into a multiscale scheme where a macroscale finite element (FE) solver is supplied with stochastic effective conductivity information by VDMN (FE-VDMN). The thermal VDMN represents variability through probability distributions over the geometric parameters of its hierarchical laminate building blocks and analytically propagates the first two cumulants of these distributions to the mean and covariance of the effective conductivity tensor. In this way, stochastic thermal homogenization is recast as inference and cumulant propagation on a reduced network representation. For spinodal and superalloy-like microstructure families, the thermal VDMN reproduces the mean and a calibrated joint distribution of the effective conductivity components over a broad range of phase contrast. Demonstrated at the component scale, FE-VDMN propagates this distribution to the temperature field at a fraction of the cost of classical FE 2 schemes, which require a new microscale solve for every microstructure realization. This framework provides a computationally tractable route for uncertainty-aware multiscale thermal simulation, with FE-FFT reference comparisons included to assess accuracy and computational cost.
No takes yet. Share an insight, caveat, or question.
Francis et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: