Density functional theory (DFT) describes matter at the quantum level, but popular approximations suffer from systematic errors that have arisen from the violation of mathematical properties of the exact functional. DeepMind has overcome this fundamental limitation by training a neural network on molecular data and on fictitious systems with fractional charge and spin. The result was the DM21 (DeepMind 21) tool. It correctly describes typical examples of artificial charge delocalization and strong correlation and performs better than traditional functionals on thorough benchmarks for main-group atoms and molecules. The company claims that DM21 accurately models complex systems such as hydrogen chains, charged DNA base pairs, and diradical transition states.
DFT is a crucial technology and is essential to designing functionals that get simple chemistry correct before explaining more complex molecular interactions. To solve some of the major challenges of the 21st century, like producing clean electricity or developing high-temperature superconductors, it is essential to design new materials with specific properties. Doing this on a computer requires simulation of electrons, subatomic particles that govern the way in which atoms bond to form molecules and are also responsible for the flow of electricity in solids. Despite years of effort and several significant advances, it is still a challenge to model the quantum mechanical behaviour of electrons accurately.
The team has addressed two long-standing problems:
- The delocalization error: In a DFT calculation, the functional determines the charge density of a molecule by finding the configuration of electrons which minimizes energy. Thus, errors in the functional can lead to errors in the calculated electron density. Most existing density functional approximations prefer electron densities that are unrealistically spread out over several atoms or molecules rather than being correctly localized around a single molecule or atom.
- Spin symmetry breaking: When describing the breaking of chemical bonds, existing functionals tend to unrealistically prefer configurations in which a fundamental symmetry known as spin symmetry is broken. Since symmetries play a vital role in our understanding of physics and chemistry, this artificial symmetry breaking reveals a major deficiency in existing functionals.
These longstanding challenges are both related to how functionals behave when presented with a system that exhibits “fractional electron character.” DeepMind found that the problems of delocalization and spin symmetry-breaking can be solved by using a neural network to represent the functional and tailoring their training dataset to capture the fractional electron behaviour expected for the exact functional.
Their function showed itself to be highly accurate on large-scale benchmarks, suggesting that the data-driven approach is capable of capturing aspects of the exact functional.