Electron beams can reveal individual atoms and damage the arrangement being studied. Researchers at Peking University are tackling that trade-off with a physics-aware neural-network system that reconstructs three-dimensional atomic positions and chemical identities from noisier, lower-dose information.[1]
Published on 23 October, the work concerns atomic electron tomography: a microscope images a sample from different angles, and software reconstructs its structure. The aim is to retain enough information to place atoms accurately while reducing the exposure needed to see them.[1]
Two stages: locate the atoms, then identify them
Yao Zhang and colleagues at Peking University describe a physics-aware neural-network workflow, abbreviated PANN. It has two stages. First, a three-dimensional ResUNet—a deep-learning architecture built to combine local detail with broader spatial structure—refines a rough reconstructed volume and corrects distortions. Second, local atomic neighbourhoods are represented using mathematical descriptors called three-dimensional Zernike moments. A graph-attention Transformer then works with those descriptors and the estimated coordinates to classify the elemental species.[1]
Position and identity answer different materials questions. A reconstruction may place an atomic site correctly while assigning the wrong element to it. Using local shapes and relationships among neighbouring sites lets the second stage draw on more than image brightness.[1]
The team benchmarked the framework using 42,588 simulated reconstructed volumes spanning different material morphologies, noise conditions and exposure settings. The work reports reduced coordinate error and improved atom recovery when compared under reduced-dose conditions. Those measures are more useful than merely displaying a cleaner-looking picture. Coordinate error asks whether a predicted atom is close to its expected location. Recovery asks whether relevant atomic sites were found at all. Mislabelled elements would matter too, particularly where material properties depend on the location of particular species.[1]
Simulation and measured data play different roles
The team also tested retrospective dose-reduction on experimental tomography data from metallic nanoparticles. Restricting information from a measured dataset offers a controlled comparison with the original reconstruction. It is a distinct test from prospectively imaging a fragile specimen at low dose.[1]
For selected sensitive inorganic materials, including the classes discussed in the paper such as halide perovskites, zeolites and quantum dots, the demonstrations are simulation-led. The outstanding experimental challenge is to preserve the gains when specimen drift, imperfect viewing angles and detector behaviour depart from the training conditions.[1][2]
A sharper map of material defects
An atomic defect can change a material’s conductivity, catalytic activity or durability. Better recovery of coordinates and element identities could therefore improve the scientific measurement itself, rather than only make an image easier to look at.[1]
The study places computation inside the microscope’s measurement problem: how much structure can be recovered from a limited electron dose? Its tests establish a route to better inference, with prospective measurements on sensitive specimens providing the next demanding comparison.
Sources & further reading
- Physics-aware neural networks enable robust and full atomic structure determination via low-dose atomic electron tomography npj Computational Materials · 23 Oct 2026 · Accessed 24 Oct 2026
- Physics-aware neural networks enable robust and full atomic structure determination via low-dose atomic electron tomography arXiv · 20 Mar 2026 · Accessed 24 Oct 2026