Solving some of the major challenges of the 21st Century, such as producing clean electricity or developing high temperature superconductors, requires designing new materials with specific properties. To do this on a computer, we need to simulate electrons, which are subatomic particles that govern how atoms bond to form molecules and conduct electricity in solids. However, accurately modeling the quantum mechanical behavior of electrons has been a difficult task for decades.
In a recent paper published in Science, a neural network called DM21 is proposed. This neural network achieves state-of-the-art accuracy in large parts of chemistry. It is designed to accelerate scientific progress, and the code is open sourced for anyone to use.
Nearly a century ago, Erwin Schrödinger proposed his famous equation that governs the behavior of quantum mechanical particles. Applying this equation to electrons in molecules is challenging because electrons repel each other. However, Pierre Hohenberg and Walter Kohn made a major breakthrough in the 1960s by realizing that it is not necessary to track each electron individually. Instead, knowing the probability for any electron to be at each position (i.e., the electron density) is sufficient to compute all interactions accurately. Kohn received a Nobel Prize in Chemistry for proving this and founding Density Functional Theory (DFT).
DFT involves approximating the mapping between electron density and interaction energy, known as the density functional. Over the years, researchers have proposed many approximations to the exact functional, but they all suffer from systematic errors due to the failure to capture certain mathematical properties. This has limited their accuracy.
To overcome these challenges, the researchers expressed the functional as a neural network and incorporated the exact properties into the training data. This approach resulted in functionals that are free from important systematic errors, leading to a better description of a broad class of chemical reactions.
The researchers specifically addressed two long-standing problems with traditional functionals: the delocalization error and spin symmetry breaking. The delocalization error occurs when electron densities are unrealistically spread out over multiple atoms or molecules instead of being correctly localized. Spin symmetry breaking refers to the preference for configurations that break spin symmetry, which is a fundamental symmetry in physics and chemistry. By using a neural network and training it with the expected fractional electron behavior, the researchers were able to solve these problems.
These challenges are crucial to solve because they affect the accuracy of functionals in describing charge movement and bond breaking, which are core processes in various technological applications. By designing functionals that can accurately describe simple chemistry, like the behavior of hydrogen, we can improve their ability to explain more complex molecular interactions.
The neural network-based approach showed promising results, being highly accurate on broad, large-scale benchmarks. This data-driven approach captures aspects of the exact functional that were previously elusive.
Computer simulations have already played a central role in engineering, providing reliable answers to questions about the stability of bridges or the success of rocket launches. As technology continues to explore the quantum scale for materials, medicines, and catalysts, deep learning offers a promising way to accurately simulate matter at the quantum mechanical level.