An artificial intelligence (AI) framework is capable of addressing absent or improperly positioned atoms – like hydrogen – within the crystal frameworks of inorganic substances. Adjusting atomic arrangements in this manner could empower chemists to more accurately model material formations or assist in the creation of new substances, including superconductors.
Establishing the placements of hydrogen and other small atoms in crystalline substances is quite difficult because these elements scatter x-rays weakly, which implies that methods such as x-ray powder diffraction can yield imprecise structures. Although neutron or synchrotron diffraction methods provide greater accuracy, they necessitate access to large facilities, require larger quantities of material, and are costly to operate.
‘I recall a time when I aimed to juxtapose the outcomes of our predictions regarding the crystal framework of cellulose with experimental results,’ mentions Artem Oganov from the Skolkovo Institute of Science and Technology in Russia, who was not part of this new study. Despite being the most prevalent polymer on the planet, he notes that the crystal framework still exhibited absent hydrogen atoms.
‘If we lack knowledge of the atomic positions, we cannot simulate the material,’ asserts Timo Reents, a doctoral student at the PSI Centre for Scientific Computing, Theory and Data in Switzerland, who created the model. Changing atomic locations can influence material attributes, including thermal and electrical conductivity, vibrational spectra, and superconductivity in hydride substances.
The Swiss group has now enhanced Microsoft’s MatterGen – a generative AI framework capable of generating novel inorganic materials – to improve the positioning of atoms in crystal arrangements. Reents compares the team’s efforts to utilizing AI tools to eliminate unwanted elements from images. Such a model can then leverage its training on analogous images to reconstruct the omitted section with what is presumably there.
A comparable approach is employed by the team at PSI to rectify absent or inaccurately placed atoms in crystal frameworks. ‘We know the [positions of] heavy atoms, we understand the unit cell shape,’ Reents explains, ‘and we aim to use this host framework [to] forecast the hydrogen positions, which represent this absent segment in the image.’
The generative AI model operates by introducing ‘noise’ or supplementary data points to the uncertain positions within the crystal and fine-tunes until the model yields the lowest energy structure. The team trained the model by systematically removing the locations of hydrogen atoms from established crystal frameworks identified in an inorganic structural database. This encompassed over 800 DFT-generated structures containing as many as 20 atoms per unit cell.
Widening the dataset to include materials with unit cells holding up to 40 atoms permitted the Swiss team to evaluate the model’s efficacy on thousands of structures. In about 85% of instances, the model managed to accurately predict the corresponding crystal structure, and in an additional 12%, it predicted structures that were more stable.
Reents indicates that the model is now accessible to the public. He also points out that the model is ‘hydrogen agnostic,’ suggesting that it can be employed to foresee the positions of other atoms, such as lithium or sodium.
Pierre-Paul De Breuck, a computational materials scientist at Ruhr-Universität Bochum in Germany, believes that this model will be beneficial for highlighting and correcting inaccuracies in current crystal frameworks. ‘Crystallographers can [additionally] utilize it as a rapid, physically grounded foundation for refining ambiguous x-ray structures, rather than depending on “chemically sensible” assumptions,’ he adds.
Nonetheless, De Breuck remarks that the DFT simulations on which the model is trained are typically conducted at 0K. ‘X-ray experiments are performed at finite temperatures, so lattice parameters and atomic placements can vary somewhat from what the model has perceived.’
While this model isn’t a groundbreaking invention, according to Oganov, he still considers that ‘addressing a longstanding issue that has troubled the entire community’ is priceless. ‘Hydrogen possesses the same rights as other atoms,’ he concludes.