That gap may now be closing. Researchers at Zhejiang University have developed AtomDiT, a machine learning framework that transforms a single high-resolution electron micrograph into a complete digital twin of a dislocation system. The twin does not merely reproduce the picture; it simultaneously records the type and position of each dislocation, the surrounding stress field, and the distribution of light elements such as hydrogen, none of which are directly visible in the original image. The work, published in National Science Review, is demonstrated across pure metals and complex alloys, including dual-phase systems and grain boundaries decorated with segregated hydrogen, and it uncovers how stress and chemistry together govern dislocation motion. In effect, the framework turns one static snapshot of atoms into a multidimensional portrait of the mechanics hidden inside a crystal.

The key insight behind AtomDiT is that the information needed to reconstruct the invisible fields is already encoded in the visible structure. The framework relies on the local atomic environment, meaning the geometric arrangement of atoms in the immediate neighborhood of each atomic site. Dislocations distort the crystal lattice in highly specific and characteristic ways: the planes of atoms bend around the defect core, interplanar spacings stretch and compress, and the symmetry of the local neighborhood changes in patterns that depend on the dislocation’s type and position. Because these distortions are caused by the stress field itself, the local arrangements of atoms contain information not only about the dislocations but also about the mechanical and chemical fields surrounding them. AtomDiT learns these relationships from large sets of atomistic simulations and then applies that learned knowledge to experimental images, without requiring any prior information about loading conditions or the detailed composition of the material being studied.

The technical pipeline works in a strikingly modular way. To identify dislocations, AtomDiT first assigns each atom a descriptor based on its distances to nearby neighbor atoms, a compact numerical fingerprint of the local environment. A set of trained neural networks then predicts, for each atom, how far that atom sits from each possible dislocation type. The design resembles a bank of sensors, each one tuned to detect a different kind of defect signal buried in the lattice geometry. From these per-atom predictions, the framework builds probability maps over the entire imaged region and then locates dislocation cores using unsupervised clustering, letting the data itself reveal where the defects are rather than imposing a predefined template. Crucially, the same local descriptors that identify the dislocations are also used to reconstruct the stress and chemical distributions around them, so all components of the digital twin emerge from a single unified representation of the atomic environment.

Building a model that generalizes across real materials required an unusually broad training database. The researchers trained AtomDiT on atomistic configurations spanning several crystal structures, the common dislocation types found in engineering metals, grain boundaries, cracks, triple junctions, and full polycrystalline networks, all sampled under a wide range of deformation states. This diversity ensures that when the model confronts an experimental image, it has already seen the relevant lattice distortions in simulation. For complex alloys with severe lattice distortion, where the local environments can deviate strongly from those of pure metals, the framework employs transfer learning, allowing the model to adapt to a new material system with only a small amount of additional data. In tests across pure metals and multicomponent alloys, the resulting digital twin located dislocation cores with an accuracy well below a single atomic spacing, meaning the predicted positions of the defects are reliable at the finest scale that atomic imaging can resolve.

The real test of any digital twin is whether it reveals physics that was previously out of reach, and the researchers demonstrated this in two demanding scenarios. The first involved a dual-phase alloy, a material composed of two distinct crystalline phases separated by an interface. Dislocations moving through such a material must eventually confront a phase boundary, and whether they pass through it or pile up against it has enormous consequences for the alloy’s strength and ductility. Using AtomDiT, the digital twin tracked a dislocation as it approached and crossed a phase boundary, simultaneously recording the transformation of its dislocation type and the evolution of the stress field around it. The twin showed that a smooth, broadly distributed stress field allows the dislocation to transmit through the interface, while a sharply fluctuating stress field blocks its passage. This connection between the spatial character of the stress field and the fate of a moving defect provides a direct, atomically resolved explanation of how interfaces control plastic flow in multiphase metals.

The second scenario addressed one of the most consequential problems in materials science: hydrogen embrittlement. Hydrogen atoms are so small that they slip into the spaces between metal atoms and segregate at grain boundaries, the interfaces between individual crystalline grains, where they are notoriously difficult to detect and even harder to correlate with mechanical behavior. AtomDiT’s digital twin mapped hydrogen concentration, dislocation character, and stress together in hydrogen-segregated grain boundaries, producing a combined chemical and mechanical picture of the same nanoscale region. The analysis revealed that hydrogen changes dislocations from highly mobile types into less mobile ones and creates a tensile stress peak at the boundary itself. Atomistic simulations confirmed that the energy barrier for a dislocation to transmit across the boundary increases with hydrogen content, directly connecting the chemical picture reconstructed from the micrograph to the mechanical degradation that hydrogen is known to cause.

What makes this approach conceptually powerful is the direction of the information flow. Traditional multiscale modeling struggles to connect the atomistic scale, where defects live and interact, with the continuum scale, where engineers describe plasticity in terms of stresses and strains. The authors note that the digital twin can be built from real atomic structures obtained from experimental measurements, which opens a path to feed experimental data directly into larger-scale plasticity models. Instead of relying on idealized simulated structures as the starting point for multiscale simulations, researchers can now begin from the actual atomic configuration of a real material, complete with its measured defect content and reconstructed fields. This bridges atomic-scale defect dynamics with the behavior of engineering materials, grounding continuum-level predictions in experimentally verified atomic reality.

The implications extend across the materials economy. Alloys designed for aircraft, pipelines, nuclear reactors, and hydrogen-powered vehicles all fail, when they fail, through the motion and interaction of dislocations, often accelerated by environmental species like hydrogen. A tool that can extract defect identity, stress, and chemistry from a single micrograph turns routine atomic-resolution imaging into a quantitative probe of the mechanics that were previously invisible. By turning a single picture into a multidimensional portrait of structure, stress, and composition, AtomDiT offers a new way to see the invisible machinery inside metals, and with it, a new route to understanding, predicting, and ultimately preventing crystal deformation and damage.

Subject of Research: Machine learning reconstruction of dislocation structure, stress, and chemical fields from atomic-resolution electron micrographs

Article Title: How to see the atomic-scale invisibles via digital twinning?

Article References: How to see the atomic-scale invisibles via digital twinning?. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: dislocations, digital twin, machine learning, electron microscopy, hydrogen embrittlement, stress fields, plastic deformation, grain boundaries, dual-phase alloys, atomistic simulation, Zhejiang University, National Science Review