Neural networks can be used to build a more accurate map of the density and interaction between electrons than was previously attainable, new research from DeepMind shows. The result is a step towards enabling scientists to better understand the interactions between electrons, which hold molecules together, and it also shows deep learning’s promise for accurately simulating matter at the quantum mechanical level — which may enable improved design in a computer by allowing researchers to explore questions about materials, medicines, and catalysts at the nanoscale level. Density functional theory (DFT), which describes fundamental properties of quantum matter, was first established more than 50 years ago. It has become a dominant method to predict the properties of electron interactions in chemistry, biology, and materials science. However, the exact nature of the mapping between electron density and interaction energy – the so-called density functional – has long eluded scientific understanding. Because of this, even state-of-the-art DFT functionals are plagued by fundamental systematic errors in describing fractional electron charges and spins. To address these limitations, James Kirkpatrick and colleagues used the DeepMind platform to develop a framework to train the neural network on accurate chemical data and fractional electron constraints, resulting in the functional “DM21.” According to the authors, DM21 was able to learn functionals free from two critical systematic errors, the delocalization error and spin symmetry breaking, which resulted in better modeling of a broad class of chemical reactions than previous platforms. “The importance of DM21 developed by Kirkpatrick et al. is not that it yields the ultimate density functional but that an artificial intelligence approach addresses the fractional electron and spin problem that has resisted a direct analytical solution to creating the functional,” writes Jon Perdew in a related Perspective.
Pushing the Frontiers of Density Functionals by Solving the Fractional Electron Problem
Article Publication Date