How to see the atomic-scale invisibles via digital twinning?
Dislocations are the primary carriers of plastic deformation in metallic materials. They directly determine how metals respond to mechanical loading, giving them the ability to deform, but their behavior also underlies failure mechanisms such as fatigue and hydrogen embrittlement. Yet these line defects remain challenging to study: a microscope image can show atomic positions, but the mechanical stress and local chemistry that determine how dislocations move are invisible.
Recently, researchers at Zhejiang University have developed AtomDiT, a machine learning framework that turns a single high-resolution electron micrograph into a digital twin of the dislocation system. The twin simultaneously records the type and position of dislocations, their surrounding stress field, and the distribution of light elements such as hydrogen. The work is published in National Science Review.
AtomDiT relies on the local atomic environment, the geometric arrangement of atoms in the immediate neighborhood of each site. Because dislocations distort the crystal lattice in specific ways, these local arrangements contain information about the dislocations themselves and the fields around them. AtomDiT learns these relationships from large sets of atomistic simulations and then applies that knowledge to experimental images, without requiring prior information about loading conditions or material details.
To identify dislocations, AtomDiT assigns each atom a descriptor based on distances to its near neighbors. A set of trained neural networks then predicts how far that atom is from each possible dislocation type, much like a bank of sensors tuned to different signals. From these predictions, the framework builds probability maps and locates dislocation cores with unsupervised clustering. At the same time, it reconstructs stress and chemical distributions from the same descriptors, integrating everything into one digital twin model.
The framework is designed for broad applicability. Its training database includes several crystal structures, common dislocation types, grain boundaries, cracks, triple junctions, and polycrystalline networks under a wide range of deformation states. For complex alloys with severe lattice distortion, transfer learning allows the model to adapt with only a small amount of new data. In tests across pure metals and multicomponent alloys, the twin locates dislocation cores with an accuracy well below a single atomic spacing.
In a dual-phase alloy, the digital twin tracked a dislocation as it approached and crossed a phase boundary, simultaneously recording the transformation of its type and the evolution of stress. The twin showed that a smooth, broadly distributed stress field allows transmission, while a sharply fluctuating stress field blocks it. In hydrogen-segregated grain boundaries, the digital twin mapped hydrogen concentration, dislocation character, and stress together. It revealed that hydrogen changes dislocations from highly mobile to less mobile types and creates a tensile stress peak at the boundary. Simulations confirmed that the energy barrier for transmission increases with hydrogen content, connecting the chemical picture to the mechanical behavior.
The authors note that the digital twin can be built from real atomic structures obtained from experimental measurements. This opens a path to feed experimental data directly into larger-scale plasticity models, bridging atomic-scale defect dynamics with the behavior of engineering materials. By turning a single picture into a multidimensional portrait, AtomDiT offers a new way to see the invisible mechanics inside metals.
Reference An atomic-scale digital twin of dislocation systems
Qingkun Zhao, Zhenghao Zhang, Yanqi Zhou, Hailin Deng, Qi Zhu, Haofei Zhou, and Wei Yang
Science China Press
