Materials discovery is critical but tough. New materials enable big innovations like batteries or LEDs. But there are ~infinitely many combinations to try. Testing for them experimentally is slow and expensive.
So scientists and engineers want to simulate and screen materials on computers first. This can check way more candidates before real-world experiments. However, models historically struggled at accurately predicting if materials are stable.
Researchers at DeepMind made a system called GNoME that uses graph neural networks and active learning to push past these limits.
GNoME models materials’ crystal structures as graphs and predicts formation energies. It actively generates and filters candidates, evaluating the most promising with simulations. This expands its knowledge and improves predictions over multiple cycles.
The authors introduced new ways to generate derivative structures that respect symmetries, further diversifying discoveries.
The results:
- GNoME found 2.2 million new stable materials - equivalent to 800 years of normal discovery.
- Of those, 380k were the most stable and candidates for validation.
- 736 were validated in external labs. These include a totally new diamond-like optical material and another that may be a superconductor.
Overall this demonstrates how scaling up deep learning can massively speed up materials innovation. As data and models improve together, it’ll accelerate solutions to big problems needing new engineered materials.
TLDR: DeepMind made an AI system that uses graph neural networks to discover possible new materials. It found 2.2 million candidates, and over 300k are most stable. Over 700 have already been synthesized.
Full summary available here. Paper is here.
Is there a materials scientist who can explain the significance of this to a reformed computational chemist?
What does it actually mean when they say they have “found” 2.2 million new stable materials and they think 380,000 are candidates for validation? They say 736 were experimentally validated, but from looking at the paper those 736 were found independently during the time they were developing GNoME, not picked by them as an actual test of their abilities.
How many of the 380,000 candidates have been experimentally validated by the authors of the paper?
To put it in small molecule drug discovery terms, if I said I had created 2.2 million new drugs in silico, and suggested 380,000 for validation I don’t think anyone would be very impressed – pretty much any virtual screening tool can do that and has done for decades without the need for deep learning. If 736 molecules turned up in the Journal of Medicinal Chemistry as having been synthesized and found to really be biologically active or whatever during my research that’s cool, but also that still leaves 379,264 to go before we know my hit rate.
I assume what they have done is actually more impressive than the analogy I have made above…
I don’t think it’s more impressive than that - it’s simply very very hard to this kind of discovery for natural sciences without a wet lab to verify results and streamline the process. This is gaining traction in medicine now (at least in some industry labs/startups), so this could be the start of more collaboration for material sciences. Definitely a “first step” kind of result, though, that won’t have much real-world impact on its own without such collaboration…
They are all known to be stable, because they have a ground-truth simulator to test with. Stable doesn’t necessarily mean useful, but that wasn’t the point.
The benefit here is that training a neural network on simulator data allows you to generate instead of search. The simulator is very computationally expensive (even compared to a deep neural network) and the search space is large and high-dimensional.