Machine learning for the simulation of photochemistry

03.12.2020

The aim of photochemistry simulations is to analyse and predict the interactions of molecules and light - as they are crucial for processes such as photosynthesis, human vision and photovoltaics. Quantum chemical methods that can be used to describe the many different excited states of molecules have recently experienced a major upswing through the use of machine learning. The enormous potential of this AI technology for theoretical calculations is demonstrated by the researchers Philipp Marquetand from the University of Vienna and Julia Westermayr from the University of Warwick in a recently published article in "Chemical Reviews".

The theoretical investigation of the excited states of molecules is important in order to provide, hand in hand with experimental research, an even more precise insight into many fundamental processes of life and nature.

"In many cases, we can describe the ground state of molecules, in which the electrons have not yet been excited by photons or other influences, by simple and fast force field calculations," says study author Philipp Marquetand from the Department of Theoretical Chemistry at the Faculty of Chemistry. The much more complicated theoretical characterization and prediction of excited states is what researchers are trying to solve with the help of quantum mechanics, which in turn is very tedious and computationally expensive.

100,000 times faster calculations

In their survey article, Marquetand and his former doctoral student Westermayr, who is now working as a postdoc in Great Britain, show which possibilities the use of machine learning opens up in this regard and what is already possible today:

"This form of artificial intelligence can significantly accelerate quantum chemical simulations. We are currently achieving an acceleration of 100,000 times," says Marquetand, who is investigating excited-state dynamics by machine learning with his group and, building on this, would like to expand the basic understanding of reaction mechanisms in photophysics and photochemistry: "Our dream is to be able to calculate even faster in the future".

Predicting new molecules

Using machine learning methods, the researchers can, for example, also recognise the colour of a molecule much more precisely, i.e., achieve a higher resolution in the simulation of UV/VIS spectroscopy. This approach, in which the machines learn how electrons are excited with ultraviolet (UV) and visible light, also requires transfer performance instead of stupid memorization.  "Here, for example, we try to calculate absorption spectra theoretically and even predict molecules that the machine learning algorithm has not seen before", says Marquetand. In simple terms: "I show you two molecules, please predict a third one" - an approach that "in parts already delivers very promising results", although research on this application of AI is still in its infancy.

Publication in "Chemical Reviews":

Machine Learning for Electronically Excited States of Molecules, by Julia Westermayr and Philipp Marquetand, in Chemical Reviews 2020, doi.org/10.1021/acs.chemrev.0c00749

Scientific contact

Priv.-Doz. Dr. Philipp Marquetand

Department of Theoretical Chemistry
University of Vienna
1090 - Wien, Währinger Straße 17
+43-1-4277-527 64
philipp.marquetand@univie.ac.at

 Back

Machine learning for the simulation of photochemistry (Copyright: P. Marquetand & J. Westermayr)