Virtual Reality (VR) helps to understand chemical reactions
- D-CHAB
- LPC
- Highlights
In a cover article of The Journal of Physical Chemistry, a research team from the University of Bristol and ETH Zurich describes how advanced virtual reality (VR) frameworks enable researchers to intuitively steer the training of machine learning algorithms and accelerate scientific discovery.
High-quality training data are crucial for machine learning and in particular for the training of artificial neural networks, which can be used to model the formation and breaking of chemical bonds in a chemical reaction.
Combining human intuition and VR
The team around David Glowacki (Bristol) and Markus Reiher (ETH Zurich) uses a newly developed open-source and interactive software to teach quantum chemistry to neural networks. Generating datasets to teach quantum chemistry to machines is a longstanding challenge. The researchers show that human intuition, combined with VR, can generate high-quality training data, and thus improve machine learning models. The VR software framework conducts ‘on-the-fly’ quantum mechanics calculations, allowing scientists to explore sophisticated physics models of complex molecular rearrangements – which involve the making and breaking of chemical bonds. This is the first time that virtual reality has been used in that kind of studies.
According to David Glowacki, “immersive tools like VR provide an efficient means for humans to express high-level scientific and design insight. As far as we know, this work represents the first time that a VR framework has been used to generate data for training a neural network.” Markus Reiher from ETH Zurich adds: “This work shows that advanced visualization and interaction frameworks like VR and AR enable humans to complement automated machine learning approaches, and accelerate scientific discovery. The paper offers an interesting vision for how science may evolve in the near future, where humans focus their efforts on how to effectively train machines.”
The results of this pioneering work have been published on May 23, 2019, as a cover article of The Journal of Physical Chemistry. The article is the most-read paper of the journal this month and as ACS Editor’s Choice, is free to read and redistribute.
Reference
Silvia Amabilino, Lars A. Bratholm,Simon J. Bennie, Alain C. Vaucher, Markus Reiher, and David R. Glowacki: Training Neural Nets to Learn Reactive Potential Energy Surfaces using Interactive Quantum Chemistry in Virtual Reality. The Journal of Physical Chemistry A (2019), 123, 4486-4499. DOI: external page 10.1021/acs.jpca.9b01006