Distortion states as robust embeddings for quantifying and forecasting atomistic damage from electron microscopy

We present an AI-driven framework that helps describe and predict damage at the atomistic level in irradiated materials. Rather than identifying discrete, visible defects, we define damage as a continuous spectrum of lattice distortions arising from atomic fluctuations. These distortions are encoded into a compact distortion state vector that unifies information from atomistic simulations and synthetic scanning transmission electron microscopy (STEM). Deep-learning models learn to reconstruct this state directly from STEM images, linking experimental contrast to the underlying atomic disorder without annotations or expert intervention. Operating in this representation space, the framework can forecast the evolution of hidden defect populations from only a few dozen initial images. Thus, it opens new perspectives for understanding how atomic-scale fluctuations give rise to macroscopic degradation, laying the foundation for physics-based digital twins of materials.