Physics-informed active learning for the design of thermally conductive composite microstructures with controlled anisotropy

The increasing power density of modern electronic devices demands advanced thermal interface materials (TIMs) with improved heat dissipation. Conventional material discovery approaches, based on extensive experiments or computationally intensive simulations, struggle to efficiently explore the vast design space of polymer-filler composites. This work presents an active learning framework for the accelerated discovery of thermally conductive composites with controlled thermal anisotropy and percolation behavior. The framework integrates a conditional variational autoencoder for generative modeling of 3D voxelized microstructures, a physics-informed finite-difference solver to evaluate anisotropic thermal conductivity along three orthogonal directions, and an ensemble surrogate model for rapid property prediction with uncertainty estimation. An adaptive candidate selection strategy balances exploration and exploitation, guiding the search toward high-performance and previously unexplored regions of the design space while preserving structural diversity. Iterative refinement of the training dataset enables the identification of novel non-percolating microstructures exhibiting enhanced thermal conductivity and low anisotropy, reaching mean conductivities of approximately 1.4 W m -1 K -1 . The proposed methodology provides a scalable protocol for the data-driven design of composite materials with tailored thermal transport properties and offers a systematic pathway for TIM development. • Data-driven framework for thermally conductive polymer composites. • Finite-difference solver quantifies anisotropic thermal conductivity. • Conditional VAE generates 3D non-percolating composites at 20 vol%. • Surrogate model trained on ∼20,000 microstructures with uncertainty. • Conductivity (∼1.4 W m -1 K -1 ) achieved with low anisotropy.