Summary

Researchers at the Paul Scherrer Institute PSI have developed an AI model called XtalPaint that reconstructs missing positions of hydrogen atoms in crystal structures. The team led by Giovanni Pizzi trained the artificial intelligence based on Microsoft's MatterGen model to solve a central problem in materials science: hydrogen atoms are difficult to detect experimentally and therefore frequently missing from material databases. The new method achieves a success rate of 97 percent and enables more precise computer simulations of material properties for applications such as hydrogen storage and battery development.

People

Topics

  • Artificial Intelligence in Materials Research
  • Crystal Structure Analysis
  • Hydrogen Storage
  • Battery Development
  • Computer Simulations

Clarus Lead

The development of XtalPaint addresses a critical gap in modern materials research: while databases contain thousands of potentially interesting materials, many remain inaccessible for computer simulations because hydrogen positions are missing or imprecise. By transferring established inpainting techniques from machine vision to crystallography, the PSI team enables XtalPaint to perform systematic reconstruction of this critical structural information for the first time – with direct consequences for the development of energy storage technologies and superconductor research.

Detailed Summary

The central problem lies in experimental measurement methodology: traditional X-ray scattering procedures determine atomic positions through beam refraction but capture hydrogen only inadequately. This leads to gaps or inaccuracies in structural databases – an obstacle to computer simulations of material properties such as electrical or thermal conductivity. Giovanni Pizzi estimates that several thousand potentially relevant materials therefore cannot be simulated.

The XtalPaint model transfers an established concept from image processing: diffusion models that supplement missing image information through so-called inpainting (for example, a hidden dog's paw). The innovative step lies in selective noise introduction – only unknown atomic positions are overlaid with random information, while known structures remain unchanged. This significantly increases both the success rate and computational efficiency.

In validation tests, the researchers intentionally removed known hydrogen positions and had XtalPaint reconstruct them: in 87 percent of cases the original positions were correctly identified, in a further 10 percent the model found energetically more stable configurations – altogether a success rate of 97 percent. The team plans to use this to complete database entries and has already discovered errors in existing databases. The method is not limited to hydrogen; it can also be applied to lithium and sodium, elements that are central to battery development.

Key Statements

  • XtalPaint reconstructs missing hydrogen positions in crystal structures with 97 percent success rate
  • The method transfers established AI inpainting techniques from image processing to materials science
  • Selective noise introduction (only unknown atomic positions) increases efficiency and accuracy
  • Application potentials: hydrogen storage, battery development, superconductor research
  • Enables use of thousands of previously inaccessible materials for computer simulations

Critical Questions

  1. Validation Method (Evidence): How representative is the test dataset on which XtalPaint was validated? Were crystal structures from various material classes and complexity levels tested, or did validation focus on specific substance groups?

  2. Generalizability (Data Quality): The 97 percent success rate is based on artificially removed hydrogen positions. How does XtalPaint perform with genuine, experimentally missing or erroneous database entries that may follow different patterns?

  3. Training Data and Bias (Conflicts of Interest): On which datasets was XtalPaint trained? Is there a risk that the model perpetuates systematic errors from training data or favors frequently occurring structure types over rare materials?

  4. Energetic Stability (Causality): In 10 percent of cases, XtalPaint identified energetically more stable configurations than the known positions. How are these cases experimentally validated? Could they be attributed to measurement errors, or do they indicate genuine alternative structures?

  5. Computational Power and Scalability (Feasibility): What computational resources does XtalPaint require for large databases with millions of crystal structures? Is the method practical for real-time database completion or only for batch processing?

  6. Applicability to Other Elements (Risks): The method is applied to lithium and sodium. Are there elements or crystal classes where the method performs poorly? How is this characterized?

Reference List

Primary Source: Score-based diffusion models for accurate crystal-structure inpainting and reconstruction of hydrogen positions – npj Computational Materials, 11.06.2026

Supplementary Sources:

  1. Press Release Paul Scherrer Institute PSI: Tracking Down the Missing Hydrogen Atoms
  2. Official Source: news.admin.ch

Verification Status: ✓ 11.06.2026

This text was created with the support of an AI model. Editorial Responsibility: clarus.news | Fact-Check: 11.06.2026