Machine learning helps solve a central problem of quantum chemistry

Scientists solve a decades-old quantum chemistry problem

Researchers at Heidelberg University have used machine learning to solve a long-standing problem in quantum chemistry. Their new method, called STRUCTURES25, allows for the precise and stable calculation of molecular energies using a computationally efficient "orbital-free" approach.

The work was published in the Journal of the American Chemical Society. It represents a major step toward simulating very large molecules, like potential drugs, which was previously impractical.

The promise and problem of orbital-free calculations

Understanding a molecule's properties requires knowing how its electrons are distributed. Quantum chemists often use density functional theory for these calculations, which relies on electron density instead of more complex wave functions.

An "orbital-free" version of this theory promises much faster computations. For decades, however, it was considered unreliable. Small errors in calculating electron density would cause the entire process to become unstable and produce nonsensical, "non-physical" results.

"Orbital-free density functional theory long held the promise of faster calculation—but not at the expense of physics, please," said Prof. Dr. Fred Hamprecht, who leads a research group at the university's Interdisciplinary Center for Scientific Computing.

A neural network trained for stability

The Heidelberg team's breakthrough came from a specially designed neural network. This AI model learns the direct relationship between a molecule's electron density and its energy from highly accurate reference calculations.

A unique training method was key to its success. The model was trained not just on correct solutions, but on many slightly incorrect variants generated from the reference data.

This taught the AI how to correct course. As a result, the STRUCTURES25 process can now find the correct physical solution even when starting with a poor estimate, without spiraling into instability.

Precision that matches gold-standard methods

In tests on a diverse set of organic molecules, the new method achieved a precision competitive with established, more computationally intensive techniques. Critically, it did so stably.

The researchers demonstrated the method on larger, "drug-like" molecules where the computational savings become significant. Initial analyses show the approach scales better with molecule size than traditional methods.

"Optimization is no longer unstable, and hence a major step forward for considerably faster predictions with high precision," said Prof. Dr. Andreas Dreuw, who co-led the research.

What this new computational power enables

The stability and efficiency of STRUCTURES25 opens doors to simulations that were previously out of reach. Researchers can now contemplate studying systems that are too large or require too many configurations for older methods.

This has broad implications for fields that rely on molecular design and simulation. Potential applications include:

• Designing new pharmaceutical drugs
• Developing better battery materials
• Creating more efficient catalysts
• Engineering novel materials for energy conversion

"Now simulations are within reach that classic processes could barely touch," Dreuw stated. The work combines advances in AI, physics, and chemistry to finally deliver on a decades-old computational promise.