Chemistry matters. Join us to get the news you need.

If you have an ACS member number, please enter it here so we can link this account to your membership. (optional)

ACS values your privacy. By submitting your information, you are gaining access to C&EN and subscribing to our weekly newsletter. We use the information you provide to make your reading experience better, and we will never sell your data to third party members.


Computational Chemistry

Quantum layer boosts machine learning predictions

Added algorithms reduce errors in predicting electronic properties

by Sam Lemonick
March 31, 2019

Computational chemistry lets chemists predict molecules’ properties without measuring them in the lab. Some of the most accurate computational chemistry tools use quantum chemistry, but these calculations can be time-consuming.

Today, at the American Chemical Society national meeting in Orlando, David J. Yaron of Carnegie Mellon University described a way of speeding up calculations of molecular properties by combining quantum chemistry with machine learning. By adding a quantum chemistry “layer” to a machine learning algorithm to predict the dipole, atomic charge, and other properties of a molecule, Yaron said, his team was able to reduce calculation errors by as much as two-thirds relative to standard machine learning algorithms.

Speaking during a Computers in Chemistry Division session, Yaron said he wants chemists to be able to model organic molecules more quickly without sacrificing a lot of accuracy. When calculations have to run overnight or over a weekend—a common occurrence for quantum chemistry calculations—it can slow down discovery, he emphasized.

Typically, machine learning algorithms can predict an unknown molecule’s properties after being trained with data sets that contain the properties of thousands or more molecules. Quantum chemistry can calculate those properties exactly but takes much longer.

Adding a quantum chemical algorithmic layer allows machine learning programs to more accurately predict energies of organic molecules, like the one shown.

Quantum information can be incorporated into these training sets. But Adrian Roitberg of the University of Florida, who helped develop the machine learning algorithms that Yaron’s group built on, said that his group and others usually strip quantum calculations from the predictive algorithms because including it would make calculations take too long.

Yaron’s quantum layer uses data predicted by machine learning, like energy gaps between molecules’ electronic orbitals, to rebuild quantum chemical knowledge and feed it back into the predictions. The researchers trained their augmented algorithm on more than 12,000 organic molecules containing up to seven nonhydrogen atoms, then tested its ability to predict properties of molecules with eight of these atoms.

The best version of their model predicted molecular energies with a 67% error reduction and dipole moments with a 59% error reduction relative to a machine learning algorithm not enhanced with the quantum layer. His group published the research in October (J. Chem. Theory Comput. 2018, DOI: 10.1021/acs.jctc.8b00873).

The group is turning its attention to predicting molecules’ electronic orbitals next. Yaron said the team hopes to be able to predict d-orbital behavior in organometallic interactions.



This article has been sent to the following recipient:

Leave A Comment

*Required to comment