Design of Transition Metal Dichalcogenides Using General-Purpose Machine-Learned Interatomic Potentials

Exploring the general-purpose interatomic potential for describing multiple compounds is challenging. In this study, we confirm the applicability of machine learning methods. By fitting the density-functional theory (DFT) data set using a feedforward neural network, the general-purpose machine-learned interatomic potentials (GMIPs) for transition metal dichalcogenides (TMDCs) are obtained, which can effectively describe the interatomic interactions within MX2 (M = Mo and W; X = S, Se, and Te) compounds and MoX2/WX2 heterojunctions. GMIPs demonstrate a high degree of accuracy across a wide range of temperatures and strains. Subsequently, we confirm the effectiveness of the potential through GMIP-based molecular dynamics (MD) simulations. The mechanical behavior of these compounds and heterojunctions is analyzed in large supercells containing 30,000 and 48,000 atoms. Some key parameters, including Poisson’s ratio, elastic modulus, critical strain, and tensile strength, are revealed. The lattice thermal conductivity is predicted by both nonequilibrium molecular dynamics (NEMD) and homogeneous nonequilibrium molecular dynamics (HNEMD). Among all compounds, WS2 exhibits the highest thermal conductivity (168–207 W/mK), while MoTe2 shows the lowest thermal conductivity (21–35 W/mK). Moreover, the presence of VS, VMo, and VW vacancies in MoS2 and WS2 results in the degradation of mechanical properties and a decrease in thermal conductivity.