Data-Driven Prediction of Thermal Conductivity from Short MD Trajectories: A GCNs-LSTM Approach

We propose a data-driven framework for rapid prediction of thermal conductivity in solids based on short-time molecular dynamics (MD) simulations. By converting atomic configurations into graph representations, a graph convolutional networks (GCNs) is used to extract spatial features, which are then processed by a long short-term memory (LSTM) network to capture the temporal evolution of physical properties. The framework is validated using equilibrium MD simulations of germanium at 1000 K across various system sizes. With size-specific normalization and optimized hyperparameters, the model accurately predicts the converged thermal conductivity, achieving results consistent with experimental data. Notably, the proposed method significantly reduces computational time by up to 800-fold at large system sizes, which demonstrates its potential to accelerate thermal transport simulations in solid-state systems.