Teaching Molecular Dynamics to a Non-Autoregressive Ionic Transport Predictor

Researchers have developed a novel non-autoregressive learning framework to predict ionic transport properties in materials, addressing the limitations of current methods. Traditional molecular dynamics simulations are computationally expensive, while existing autoregressive models suffer from slow sequential inference and error accumulation. Conversely, standard non-autoregressive models often lack accuracy by ignoring dynamic behaviors. This new approach utilizes auxiliary modality learning, treating atomic trajectories as an auxiliary input during training but not requiring them for inference. This allows the model to learn dynamic properties efficiently without the computational burden of sequential processing. The framework effectively leverages datasets both with and without atomic trajectories. Experimental results demonstrate a speedup of over 200 times compared to autoregressive models on trajectory-inclusive datasets, alongside significantly reduced prediction errors relative to non-autoregressive benchmarks. This advancement offers a faster, more accurate tool for material science research, particularly for studying dynamic properties like ionic transport. The associated code has been made publicly available to facilitate further research and application in the field.