Machine-Learning-Based Thermal Conductivity Prediction in Two-Dimensional TiS2/MoS2 Van Der Waals Heterostructures

Abstract Two-dimensional (2D) materials and heterostructures display unique thermal characteristics compared to their bulk counterparts. However, the accurate estimation of the thermal conductivity of 2D materials, particularly of 2D van der Waals heterostructures, presents significant challenges for both computational and experimental methods. In this study, we propose a computationally efficient approach to investigate the thermal conductivity of 2D TiS2/MoS2 van der Waals heterostructures. Our approach utilizes machine-learning interatomic potentials (MLIPs) to predict the thermal conductivity of the heterostructure. This approach effectively incorporates intralayer interactions by utilizing moment tensor potentials (MTP) trained with computationally inexpensive density functional theory (DFT)-based datasets. These datasets are generated from ab-initio molecular dynamics (AIMD) trajectories over less than 1 ps, while the interlayer van der Waals interactions are calibrated using the D3-dispersion correction method. By explicitly incorporating the missing dispersion contribution into the MTP, this method provides greater accuracy in predicting interlayer interactions than the widely applied Lennard-Jones (LJ) potential. Finally, molecular dynamics (MD) simulations are conducted to determine the thermal conductivity of the TiS2/MoS2 heterostructures using the derived potential parameters. This study enhances our understanding of thermal transport in van der Waals (vdW) heterostructures, leveraging MLIPs to explore new nanostructured materials with superior thermal conductivity.