Impact and Analysis of Optimizers on the Performance of Neural Network Force Fields

Molecular dynamics(MD) simulation is widely used in various fields,such as materials science and computational chemistry.In recent years,with the improvement in computational power,the development of neural network models,and the accumulation in first-principle data,neural network force field(NNFF) models have demonstrated high predictive accuracy.Curren-tly,there are multiple training algorithms available for NNFF models,and these models are undergoing rapid iteration.However,there remains a lack of guidance on NNFF models and their compatible optimizers.This paper selects three representative NNFF models and the three most commonly used optimization algorithms for these models,testing and evaluating them on four real-world datasets to analyze factors affecting their convergence.We have designed numerous experiments for a comprehensive evaluation,including the impact of model parameter size on the optimizer,the influence of model depth and width on convergence,and the relationship between model training time and the optimizer.Our work provides recommendations for optimizer algorithms specific to NNFF models.