Leveraging generative AI to reconstruct unknown positions of hydrogen sites

by Timo Reents (PSI/LMS)

Wednesday 4 Jun 2025, 15:00 → 16:00 Europe/Zurich

OVGA/200

Generative AI methods are rapidly evolving to speed up and improve materials discovery. Diffusion based models can not only be adopted to generate new materials with desired properties but also to reconstruct crystal structures for which structural information is only partially available [2]. In this talk, I’ll present how we use Microsoft’s mattergen [1], a diffusion based model originally designed to generate new stable crystal structures, and apply it to reconstruct missing hydrogen sites in crystal structures reported in experimental databases. This is particularly useful as the experimental measurement of hydrogen sites with standard XRD is typically challenging due to weak scattering of hydrogen. We show how to leverage approaches known from image inpainting in the field of computer vision, combined with universal machine learning interatomic potentials, to improve the success rate of correctly identifying the missing sites while significantly lowering the computational cost with respect to a direct DFT approach. The successful application of this approach will enable the extension of computational databases, such as the MC3D.

[1] Zeni, C. et al., Nature 639, 624–632 (2025)

[2] Zhong, P. et al., arXiv:2504.16893v1 (2025)