Total: 1
Thermal transport in complex solids is governed by local structure, defects, and anisotropy, yet most continuum models still rely on oversimplified, homogenized conductivities. Here, we bridge atomistic and continuum descriptions by building finite element (FE) models directly from the site-projected thermal conductivity (SPTC), an atomic-level decomposition of the Green-Kubo thermal conductivity. We introduce a new toolkit, the Simulator Collection for Atomic-to-Continuum Scales (SCACS), which uses a graph neural network to predict SPTC on large atomic structures, coarse-grain these fields into anisotropic conductivity tensors, and embeds them into the heat-flow FE equation with a customized, anisotropy-aware adaptive mesh refinement scheme. Applied to silicon nanostructures, the resulting FE models act as representative volume elements, reproduce bulk conductivities, and capture interfacial and defect-driven anisotropy while maintaining thermodynamic consistency. Additionally, SCACS predicts experimental conductance trends and fields. This innovation demonstrates a novel and general route for transferring atomistic transport information into device-scale thermal simulations with physics-based approximations.