Artificial Intelligence on Schrödinger's Tracks

Artificial Intelligence (AI) and algorithms for machine learning are now routinely used to predict our purchasing behavior, suggest travel routes, or recognize images and faces. In research, AI is establishing itself as a crucial tool to support scientific discoveries. For example, AI is increasingly used in chemistry to predict the results of experiments or simulations. To achieve this, AI must be capable of systematically incorporating the fundamental laws of physics. An interdisciplinary team of scientists from the University of Warwick, TU Berlin, and the University of Luxembourg has now developed an AI algorithm that, among other things, allows determining the necessary structure of a substance based on its desired chemical properties. A capability that could play an especially important role in the development of new drugs and materials.

The algorithm developed by chemists, physicists, and computer scientists is capable of calculating the quantum states of a molecule, known as the wave function, which determines all properties of this molecule. To do this, AI had to learn to internalize fundamental laws of physics and solve equations of quantum mechanics — such as the Schrödinger equation. The work “Unifying machine learning and quantum chemistry with a deep neural network for molecular wavefunction” has now been published in Nature Communications.

Solving these and similar equations in the conventional way requires enormous computing capacities and, above all, months of computing time. “This is usually where the bottleneck lies in the computer-aided development of new molecules, especially for medical and industrial applications,” says Prof. Dr. Klaus-Robert Müller, Professor of Machine Learning at TU Berlin. The newly developed algorithm, on the other hand, can deliver accurate predictions within seconds on a laptop or mobile phone.

“The publication is the result of a three-year joint effort and required computer science expertise to develop an algorithm flexible enough to capture the shape and behavior of wave functions, but also chemistry and physics expertise to process and visualize quantum chemical data,” says Dr. Reinhard Maurer from the Department of Chemistry at the University of Warwick.

Klaus-Robert Müller adds: “This interdisciplinary work is an important advance because it shows that AI methods can learn the most challenging aspects of quantum chemical simulation. This also includes the so-called inverse design, which has long been a dream of pharmacology and chemistry for drug development.” Inverse design refers to specifying a particular chemical property of a molecule and designing and optimizing the corresponding molecular structure based on these specifications. The interdisciplinary team believes that AI methods will further establish themselves as an essential component in computational chemistry and molecular physics in the future, enabling sustainable inverse molecular design as well.

“This work enables a new level of drug design, where both electronic and structural properties of a molecule can be combined to achieve the desired application criteria,” says Professor Dr. Alexandre Tkatchenko from the Department of Physics at the University of Luxembourg.

Publication:

“Unifying machine learning and quantum chemistry with a deep neural network for molecular wavefunction” https://www.nature.com/articles/s41467-019-12875-2

Further information gladly provided by:

Prof. Dr. Klaus-Robert Müller
TU Berlin
Department of Machine Learning
Tel.: 030 314-78620
Email: klaus-robert.mueller@tu-berlin.de