Multiscale Interpretable Deep-Learning Framework for the Identification and Visualization of Deformation Stages in Molecular Dynamics Trajectory

Abstract Microscopic deformation stage recognition from molecular dynamics (MD) trajectories is crucial for understanding the evolution of material damage; however, traditional empirical analysis and black-box single deep-learning models lack both high-throughput spatiotemporal modeling and transparent physical interpretability. This work develops a multiscale interpretable deep-learning framework to automatically classify elastic, plastic, and fracture stages from uniaxial tensile MD trajectories and physically decode model decision logic. First, a 3D atomic trajectory is converted to 2D gray-scale image sequences; a CNN-LSTM hybrid architecture is built to jointly extract spatial atomic textures and long-time deformation dynamics, reaching 98.8% test accuracy and far surpassing spatial-only CNN baselines in both supervised classification and unsupervised clustering. More importantly, a hierarchical multiscale interpretability toolkit, including multilayer feature heatmaps, smoothed gradient saliency maps, gradient-weighted class activation maps, and regularized SHAP attribution, is integrated to quantify positive/negative feature contributions and localize model focus regions. The visualized attention zones perfectly match core physical fields (von Mises strain, atomic displacement, and nonaffine deformation), resolving the explainability gap of conventional MD data mining pipelines. This work establishes a generalizable, trustworthy AI paradigm that connects data-driven prediction to intrinsic microscale mechanical mechanisms for computational material research.