AI Breakthrough Speeds Up Quantum Chemistry
Caltech’s OrbNet deep learning tool outperforms state-of-the-art solutions.
Posted Oct 11, 2020 | Reviewed by Kaja Perina
Artificial intelligence (AI) machine learning is being applied to help accelerate the complex science of quantum mechanics—the branch of physics that studies matter and light on the subatomic scale. Recently a team of scientists at the California Institute of Technology (Caltech) published a breakthrough study in The Journal of Chemical Physics that unveils a new machine learning tool called OrbNet that can perform quantum chemistry computations 1,000 times faster than existing state-of-the-art solutions.
“We demonstrate the performance of the new method for the prediction of molecular properties, including the total and relative conformer energies for molecules in range of datasets of organic and drug-like molecules,” wrote the researchers.
Quantum chemistry is the scientific study that combines chemistry and physics. Also known as molecular quantum dynamics, quantum chemistry is a subset of chemistry that studies the properties and behavior of molecules at the subatomic level through the lens of quantum mechanics.
Chemistry is the scientific study of molecules, and it touches nearly every aspect of modern life—food, clothing, housing, pharmaceuticals, energy, materials, consumer packaged goods, agriculture, health care, high technology, and more industries.
The origin of the term quantum mechanics dates to the 1920’s. Nobel Prize-winning physicist and mathematician Max Born first published the term in his 1924 paper titled Zur Quantenmechanik. Quantum mechanics spans across many disciplines and is used for DVDs, CDs, lasers, GPS systems, atomic clocks, quantum computing, microscopy, nanotechnology, and more technologies.
The aim for calculations in quantum mechanics is to solve the Schrödinger equation for a molecular system. The Schrödinger equation is named after the physicist Erwin Schrödinger who won the Nobel Prize in Physics jointly with Paul Dirac in 1933. It describes the state or wave function of a quantum mechanical system and has been successfully used to solve systems with a single electron.
However, the Schrödinger equation cannot solve larger systems. For those systems, only approximations can be made, and these tend to be resource-intensive to calculate. The density functional theory (DFT) is a widely used method of obtaining an approximate solution to the Schrödinger equation that is based on the electron density versus the wave function. In this study, the researchers found a more economical way to perform the quantum chemistry calculations with a relatively high degree of accuracy.
The new Caltech study is a collaboration between Thomas Miller III, a professor of chemistry, and Anima Anandkumar, the Bren Professor of Computing and Mathematical Sciences, along with their research colleagues Frederick Manby, Matthew Welborn and Zhuoran Qiao. Together, the team discovered a new way to perform quantum chemistry calculations by constructing a deep-learning solution that is based on applying symmetry-adapted atomic orbitals features and a graph neural network architecture.
The researchers showed that OrbNet enabled the prediction of energies for drug-like molecules with accuracy similar to density functional theory, but at significantly reduced computational cost of 1,000 times less. The researchers also reported significant improvement in prediction accuracy.
“In comparison to state-of-the-art GNN methods for the prediction of total molecule energies for the QM9 dataset, it is shown that OrbNet provides a 33% improvement in prediction accuracy with the same amount of data relative to the next-most accurate method (DeepMoleNet),” wrote the researchers.
OrbNet uses quantum operators and symmetry-adapted atomic orbitals that are constructed from a low-cost mean-field electronic structure calculation. It creates a graphical representation using node and edge attributes that correspond to elements of the tensors of the constructed symmetry-adapted atomic orbitals. The graphical representation goes through various layers of processing which produces transformed node and edge attributes which are then further processed to produce the energy predictions.
“The method enables the prediction of molecular potential energy surfaces with full quantum mechanical accuracy while enabling vast reductions in computational cost; moreover, the method outperforms existing methods in terms of its training efficiency and transferable accuracy across diverse molecular systems,” reported the researchers.
The Caltech researchers used artificial intelligence deep learning to create a solution that significantly lowers the cost of performing quantum chemistry calculations without compromising accuracy. It is a discovery that accelerates innovation in chemistry that may help unlock novel insights from the periodic table to improve daily life in the future.
Copyright © 2020 Cami Rosso All rights reserved.