Nearly a third of US supercomputer time is spent on molecular modeling, a way to test-drive new drugs and materials on the computer before making them in the lab. The most accurate approach is the quantum many-body (QMB) equation, which calculates where every electron in the material is and how they interact. But those equations are computationally expensive and not practical to use.

A computational shortcut, density functional theory (DFT), uses electron density, a probability map of where electrons are likely to be, plus the exchange-correlation (XC) functional, which sums up how electrons interact. Although DFT simplifies quantum calculations, no one knows the universal form of the XC functional. So, scientists use approximations based on specific conditions of materials. This approach works for spotting trends, but is too unreliable for precise, quantitative predictions about molecules and materials.

“We want to bring the accuracy of QMB methods together with the simplicity of DFT,” said Vikram Gavini, a computational scientist at the University of Michigan, whose team recently published a study detailing their strategy for boosting the accuracy of DFT (Sci. Adv. 2025, DOI: 10.1126/sciadv.ady8962).

In the new study, Gavini’s team shows how machine learning (ML) can be trained on QMB data to discover more universal XC functionals, creating a bridge between the two methods. Earlier attempts to train ML models to improve XC functionals typically used just the interaction energies of electrons as their training data. In contrast, Gavini’s team also included the potentials that describe how that energy changes at each point in space.

“Potentials make a stronger foundation for training because they highlight small differences in systems more clearly than energies do,” Gavini says. This allows the model to capture subtle changes more effectively for better modeling.

To test the idea, the researchers used exact energies and potentials of five atoms and two simple molecules obtained through QMB calculations. They trained the ML model to create new approximations of the XC functional with this compact dataset. When used in DFT calculations, the new functionals delivered striking accuracy. According to Gavini, their models either outperformed or matched widely used XC approximations while keeping computational costs in check.

“This model went beyond the small set of atoms it was trained on and still gave accurate results for very different systems,” Gavini says. It was also relatively inexpensive to train, since it needed data from only a handful of atoms and simple molecules. And because it was built using potentials tied directly to electron behavior, it avoided producing unphysical or meaningless results. In contrast, many earlier ML models worked only within narrow datasets and often strayed from the rules of DFT, limiting their usefulness.

“The trials performed so far are preliminary, but promising,” says Donald Truhlar, a computational chemist at the University of Minnesota who was not involved in the research.

“The model works well for light atoms,” Gavini says. “We would like to expand to solids next.” The team also hopes to pursue higher accuracy by going even further and using the potential gradients, along with potentials and energies, to train future models.

Chemical & Engineering News

ISSN 0009-2347

Copyright ©
2025 American Chemical Society