P-047

Justinas Šlepavičius

justinas.slepavicius@ftmc.lt

Andrius Merkys, Antanas Vaitkus, Saulius Gražulis, Linas Vilčiauskas

Center for Physical Sciences and Technology (FTMC), Lithuania

Using Universal Machine Learning Force Fields for the Validation of the Crystallography Open Database (COD)

The Crystallography Open Database (COD) is the second-largest crystallography database globally, which hosts over 500,000 experimental crystal structures, including metals, organics, metal-organics, and inorganics. For novel crystal structures to be included in the COD, they must be verified under proposed conditions, minimizing erroneous entries. One approach is to use computational minimization to ensure structures are physically reasonable. Density functional theory (DFT) provides high accuracy, but at a significant computational cost. Recently developed universal machine learning force fields (MLFF), such as M3G-Net, can model interatomic interactions very efficiently. M3G-Net is a graph neural network trained on DFT data that incorporates three-body interactions describing elements up to Z=89 (Actinium), achieving DFT-level accuracy at a fraction of the computational cost. Using this approach, we were able to validate over 67% of the COD, with the most significant exception being disordered structures, which cannot be modeled using the current methodology. The validation not only showed that the majority of the COD entries are physically correct, but also helped discover some improper entries. The validation increases confidence in the entries of the COD, while aiding in rectifying inconsistencies.