Machine learning helps improve the accuracy and efficiency of small molecule calculations

When the experiment is unrealistic, density functional theory (DFT) calculations provide researchers with accurate approximations of chemical properties. The mathematical equations that underpin the calculation are carefully adjusted to the molecules and materials to which they are applied. However, each equation comes with a trade-off between accuracy and computing time demand.

Microsoft researchers believe they have found a way to use machine learning to drive the limitations of small molecules (Arxiv 2025, doi: 10.48550/arxiv.2506.14665). Preprint publications are not peer-reviewed.

DFT is based on the electron density of a molecule and can theoretically be used to accurately determine its properties. In reality, it is impossible to calculate subtle interactions between many electrons. Instead, computational chemists came up with different approximations of exchange correlation (XC) functions. This is the term for the DFT equation that captures their interactions.

Instead of using these handmade features, Microsoft's Paola Gori-Giorgi, Jan Hermann, Rianne Van den Berg, and colleagues have built a deep learning model that infers XC features from a database created with approximately 150,000 reaction energies for molecules with five or fewer non-carbon atoms.

They are not the first group to apply machine learning to the challenge of devising ideal XC features. However, the Microsoft team used more complex algorithms than other teams that incorporate modern tools borrowed from large language models. And their training data is about two orders of magnitude larger than the datasets used by others.

Groups invoke functional scalars from Greek words ladder. This nods to the increasingly complex mathematical “langs” that computational scientists add to their models and sometimes referred to as Jacob's ladder approach.

Researchers report that the prediction error of functions in the calculation of small molecule energy is half that of ωB97M-V, which is considered one of the better functions available today. For calculations in which Skala XC contains untrained metal atoms, it was in the middle of the pack. They say that ωB97M-V calculates molecular properties at computer times of the same or less than other functions.

She has not tested Scala herself, but Marivi Fernandez Serra, a computational scientist at Stony Brook University, says, “I have the impression that this will be a very good feature.” Fernández-Serra, who is also working on a machine learning approach to DFT capabilities, says the way Microsoft incorporates a variety of deep learning tools makes Skala XC efficient at inferring from a large amount of data. She and others also say the group has the advantages of Microsoft's vast resources. Generate training data. This is something many academic scientists don't have, and the current US policy is shrinking even further.

For other DFT researchers, the benefits of Skala are less clear. “For people working on metals, this doesn't work,” says AJ Medford, a chemical engineer at Georgia Tech, who used machine learning for DFT. Features suitable for metals and solids are of particular value in material science, as they can accelerate the exploration of chemical spaces.

Medford is skeptical that Microsoft teams can generate equally high quality training data for atoms with more electrons. And he says researchers using DFT with small molecules may not recognize the need for new features if existing features are fully functional.

Gori-Giorgi, senior research manager for the Skala XC team, disagrees. She says, working with external experts, her group has already identified computational techniques the team believes can be used to expand its training database and include larger atoms.

Nagai, another researcher in the field, from Priority Network, said in an email to C&EN that he was impressed by how Skala XC can learn and would help advance his research. But like Medford, he is skeptical of how it freights in metals and solids. He questioned the group's test of Skala's performance, saying it would like it to challenge with more complicated calculations.

Microsoft's efforts include several echoes of Google DeepMind's foray into DFT, releasing the MachineLearning feature (Science 2021, doi: 10.1126/science.abj6511)Science 2021, doi: 10.1126/science.abm2445). But in the end, the computers were too concentrated to be useful.

James Kirkpatrick of Deepmind, who led the 2021 survey, did not comment on Microsoft's efforts, and emailed C&en that he was “working on a related approach.”