Machine learning for thermal conductivity prediction in graphene/hexagonal boron nitride van der Waals heterostructures

High-efficient thermal interface materials (TIMs) with tunable thermal conductivity (TC) and excellent insulation are crucial for effective thermal management in advanced electronic components. Graphene/hexagonal boron nitride (h-BN) (GBN) van der Waals (vdW) heterostructure is an ideal candidate for such TIMs since h-BN is insulative and the TC of this heterostructure is highly tunable by strain engineering and defect engineering (e.g., hydrogenation, in-plane strains, 13C isotope, and interlayer coupling strength). Herein, we exploit the joint effects of the factors mentioned above on the TC of GBN vdW heterostructures by using molecular dynamics (MD) simulation and machine learning (ML) models. When combined with other factors, hydrogenation has a dominant influence on the TC of GBN heterostructures. Among the ML models trained by MD results, the genetic programming (GP) model reliably predicts TC with a maximum error of 12.7%. A mathematical formula derived by the GP model connects the TC with the influencing factors to accelerate the targeted structure design. The results suggest that the GP model provides accurate TC predictions for GBN vdW heterostructures and facilitates efficient inverse design of GBN-based TIMs with desired TC.