Machine learning helps solve a central problem of quantum chemistry

"By applying new methods of machine learning to quantum chemistry research, Heidelberg University scientists have made significant strides in computational chemistry. They have achieved a major breakthrough toward solving a decades-old dilemma in quantum chemistry: the precise and stable calculation of molecular energies and electron densities with a so-called orbital-free approach, which uses considerably less computational power and therefore permits calculations for very large molecules. [...]

How electrons are distributed in a molecule determines its chemical properties—from its stability and reactivity to its biological effect. Reliably calculating this electron distribution and the resulting energy is one of the central functions of quantum chemistry. These calculations form the basis of many applications in which molecules must be specifically understood and designed, such as for new drugs, better batteries, materials for energy conversion, or more efficient catalysts.

Yet such calculations are computationally intensive and quickly become very elaborate. The larger the molecule becomes or the more variants that need checking, the sooner established computing processes reach their limits. The "Quantum Chemistry without Orbitals" project is positioned here at the interface of chemistry, physics, and AI research.

In quantum chemistry, molecules are frequently described using density functional theory, which allows for the fundamental prediction of chemical molecular properties without having to calculate the quantum mechanical wave function. The electron density is used as the main quantity instead, a simplification that finally makes computations practicable. This orbital-free approach promises especially efficient calculations but until now was considered barely useful, since small deviations in the electron density led to unstable or "non-physical" results.

With the aid of machine learning, the Heidelberg method finally solves this precision and stability problem for many different organic molecules.

The new process called STRUCTURES25 is based on a specifically developed neural network that learns the relationship between electron density and energy directly from precise reference calculations, capturing the chemical environment of each individual atom in a mathematically detailed representation. A unique training concept was pivotal: The model was trained not only with converged electron densities, but also with many variants surrounding the correct solution, generated by targeted, controlled changes in the underlying reference calculations.

This computing process is therefore able to reliably find a physically meaningful solution for molecular energies and electron densities even in the case of small deviations. It remains stable without "getting lost" in the calculation, the Heidelberg researchers emphasize.

In tests on a large and diverse collection of organic molecules, STRUCTURES25 achieved a precision that can compete with established reference calculations, for the first time demonstrating a stable convergence using an orbital-free approach. The performance of the method was demonstrated not only on small examples, but on considerably larger "drug-like" molecules as well.

Initial runtime comparisons prove that the computing process can scale better with growing molecule size and hence increase the speed of the calculation. Calculations formerly considered too elaborate are now within reach."