Thermal conductivity gradation design in alumina-polymer thermal interface materials based on DF-PINN and Bayesian optimisation

Thermal interface materials are central to thermal management in high-power electronics. However, optimisation of multimodal filler gradation remains largely empirical, because packing, percolation and processability are tightly coupled. Here we develop a physics-informed, data-efficient inverse-design framework for triple-graded alumina–polymer composites at a fixed total loading. Using six commercial alumina powders, with median sizes spanning 1 µm to 70 µm, we benchmark all twenty triplet combinations experimentally at a fixed total loading of 50 wt% and construct a Dinger–Funk-guided low-fidelity generator to provide dense synthetic supervision across composition space. A DF-PINN trained on approximately 16,000 synthetic samples, anchored by 20 high-fidelity measurements, jointly predicts packing efficiency and thermal conductivity while remaining consistent with effective-medium bounds. Moreover, the dimensionless parameter Lc serves as a practical criterion that thermal conductivity gains on the order of 10% when L c exceeds 1. Embedding the DF-PINN surrogate within Bayesian Optimisation yields recipes with thermal conductivity that outperform DF-optimal baselines by more than 15%, achieving up to 2.86 W/(m·K) for a 1–10–70 µm triplet. Meanwhile, this strategy enables scalable virtual screening from arbitrary powder libraries, thereby reducing experimental burden and development cost.