Scientists doubt DeepMind’s AI is as good for fractionally loaded systems as it looks

Relationship between BBB test systems and fractional charge atoms of the training set. Credit: Michael Medvedev (Zelinsky Institute of Organic Chemistry at RAS)

In their article published in Science in December 2021, a team from DeepMind showed how neural networks can be used to describe electronic interactions in chemical systems with greater precision than existing methods. A team of researchers from Skoltech, Zelinsky Institute of Organic Chemistry, HSE University, Yandex and Kyungpook National University show in their commentary in Science that the ability of DeepMind AI to generalize the behavior of such systems does not follow from published results and requires review.

Knowing where the electrons are in a molecule can go a long way in explaining its structure, properties, and reactivity. Chemists use density functional theory (DFT) methods, approximations of the Schrödinger equation, to create accurate and computationally efficient models of molecules and materials. But there are well-known circumstances where DFT tools fail. One predicts how atoms share electrons; in a famous example, DFT methods incorrectly predict that even when a chlorine atom and a sodium atom are infinitely far apart, the chlorine atom retains a fraction of one of the sodium atom’s electrons.

Such errors arise because DFT equations are only approximations of physical reality. Researchers from the DeepMind machine learning project say their neural network eliminates this error from the part of an electron and makes more accurate predictions than traditional DFT methods.

“Basically, DFT is a method of solving the Schrödinger equation. Its accuracy is determined by its exchange-correlation part, which is unfortunately unknown. To date, more than 400 separate approximations for this part have been proposed,” explains Petr Zhyliaev. , senior researcher at Skoltech.

“One way to construct a good deal of exchange-correlation is to transfer information about it from numerical methods that are more ‘advanced’ than density-functional theory, which are however orders of magnitude less efficient in terms of computation. In their work, DeepMind used a neural network as a universal interpolator to learn the exchange-correlation part of the functional. Their attempt was not by far the first, but it is one of more ambitious.

DeepMind has built a neural network-based density functional referred to as DM21, trained on fractional electron systems, such as a hydrogen atom with half an electron. To prove its superiority, the authors tested DM21 on a set of stretched dimers (called the BBB set), for example, two far-distance hydrogen atoms with a total of one electron.

As expected, the DM21 functional shows excellent performance on the BBB test set, far outperforming all classical DFT functionals tested and DM21m, trained identically to DM21 but without the fractional electron systems in the training set .

Although it may seem like DM21 has figured out the physics behind fractional electron systems, closer examination shows that all of the dimers in the BBB set become very similar to the systems in the ream. Indeed, due to the locality of electroweak interactions, atomic interactions are only strong at short distances, outside of which the two atoms essentially behave as if they were not interacting (see figure above).

“In a way, neural networks are like humans: they prefer to get the right answer for the wrong reason, then the other way around. Therefore, it is not so difficult to train a neural network as to prove that he has learned physical laws instead of memorizing the correct answers.Testing a neural network on systems he has seen during training is like examining a schoolboy with a task he has seen a teacher solve there barely five minutes ago,” says Michael Medvedev, the head of the Theoretical Chemistry Group at the Zelinsky Institute of Organic Chemistry of the Russian Academy of Sciences.

Thus, the BBB test set is not appropriate: it does not test DM21’s understanding of fractional electron systems: DM21 can easily get away with memorization. A thorough analysis of the other four proofs of the DM21 management of such systems also did not lead to a decisive conclusion: only its good accuracy on the SIE4x4 set can be reliable, although even there a clear trend of growth errors with distance suggest that DM21 is not completely free of problems with fractional electron systems.

The use of fractional electron systems in the training set is not the only novelty in DeepMind’s work. Their idea of ​​introducing the physical constraints into a neural network via the training set, as well as the approach to imposing the physical meaning by training on the correct chemical potential, are likely to be widely used in the construction neural network DFT functionals in the future.

Simulate matter at the nanoscale with AI

More information:
Igor S. Gerasimov et al, Commentary on “Pushing the Frontiers of Density Functionals by Solving the Fractional Electron Problem,” Science (2022). DOI: 10.1126/science.abq3385

Provided by Skolkovo Institute of Science and Technology

Quote: Scientists Doubt DeepMind’s AI is as Good for Fractionally Loaded Systems as It Seems (2022, August 15) Retrieved August 15, 2022, from 08-scientists-deepmind-ai-good-fraction-charge.html

This document is subject to copyright. Except for fair use for purposes of private study or research, no part may be reproduced without written permission. The content is provided for information only.

Leave a Reply