A new approach improves the accuracy of quantum chemical simulations using machine learning

New tricks to model molecules with quantum accuracy take a step towards uncovering equations at the heart of common simulation approaches used in basic chemistry and materials science research.

Efforts to understand materials and chemical reactions eat about a third of the supercomputer time in the US national lab. This is key to chemical and material behavior, as electrons are responsible for chemical reactivity and bonding, electrical properties, etc. However, the calculation of quantum many bodies is extremely difficult, so scientists can use them to calculate atoms and molecules with a handful of electrons at a time.

Functional theory of density, or DFT, is simple. Each requires the computational resources needed to scale the number of cubes and the number of cubes, rather than rising exponentially with new electrons. Instead of following individual electrons, this theory calculates electron density. Here, the electrons are most likely to be located in space. This way it can be used to simulate the behavior of hundreds of atoms.

An important issue for DFT users is the exchange correlation function, which explains how electrons interact according to quantum mechanical rules. So far, researchers have had to solve the problem to approximate the XC functionality of a particular application.

“We know there is a universal function. It doesn't matter whether the electrons are in molecular systems, metal fragments, or semiconductors. But we don't know what their form is.” Advances in science.

Due to the importance of DFT for future materials and basic science, the Department of Energy provided funds and supercomputer time to approach the universal XC capabilities of the UM team.

Researchers began studying individual atoms and small molecules in quantum many-body theory, allowing them to turn the problems of DFT. Instead of adding approximate XC functions to give the behavior of atoms and molecule electrons, they use machine learning to give the behavior of electrons where XC functions are calculated through quantum many-body theory.

“Many-body theory gives the right answer for the right reasons, but at irrational computational costs, our team translates many-body results into a simpler, faster form that retains most of its accuracy,” student Jeffrey Hatch.

The Zimmerman group created a training dataset for five atoms and two molecules, specifically lithium, carbon, nitrogen, oxygen, neon, dihydrogen, and lithium hydride. They tried to add fluoride and water, but these additions did not improve XC functionality. The team believes that extracting data on photoatomics and molecules is already a good idea.

However, DFT calculations using its XC feature were already far superior to what we expected in terms of its complexity level. DFT accuracy is described as a set of ladder rungs. In the most basic, first form, electrons are considered to be present in a uniform cloud. In the second version of Gavini's team used, the density of the electron clouds considered gradients changed.

In the third Lang, the researcher adds information about electrons, such as kinetic energy. This usually means bringing in simplified versions of many of the difficult electronic wave functions, which can be better explained what is happening with the electron. However, by calculating better XC features, Gavini's team had achieved accuracy for the third time.

“The use of precise XC features is as diverse as chemistry itself. It's just as relevant to researchers who want to find better battery materials for people who discover new drugs to those who build quantum computers.”

Researchers can use XC features discovered directly by groups or experiment with team approaches. For example, Gavini says they started with light atoms and molecules, and then he wants to explore solid materials.

Again, the XC feature is expected to take a universal form, but the tricky part is to figure out what it is. Does his team at XC features discover that it works well with solids? Will new functional calculated for solids be more successful? And can they build combination features that work well on both sets of materials?

Another improvement that the team wants to pursue is higher accuracy. This means that instead of collectively looking at electrons, we need to include the individual orbitals through which electrons travel as electron density. In that case, the trick to flip the problem to get the XC feature becomes much more difficult. Even using density gradients, this path requires more computing time, as it required computations on one of the largest supercomputers in the United States.