Short bond evaluation method for rapidly assessing the generalization ability of deep neural network potential function models and its effectiveness verification

When training deep neural networks using first-principles calculation data to obtain potential functions for molecular dynamics simulations, extensive model capability evaluation work is required. However, the commonly used validation sets for model evaluation are limited by the high cost of obtaining first-principles data, making it difficult to comprehensively assess the strong generalization ability of deep neural network trained models, which requires coverage of a much larger space than the training set samples. This manuscript proposes a short bond evaluation method and conducts evaluation experiments using this method and the self-consistent field labeling evaluation method on multiple tasks under different structures generalization in two complex reaction systems. It also performs correlation analysis between the results of the two methods to validate and explain the applicability and effectiveness of the proposed method. Although this method has the necessary and insufficient characteristics, the results show that this method can accelerate the assessment of model generalization capabilities while maintaining the reliability of the evaluation results. Moreover, this method can particularly accelerate the high-accuracy filter of poor-performing models, thereby helping to improve the convergence speed during the model training iteration process. At the same time, it achieves a significant reduction in evaluation costs.