Google DeepMind's New AI Tool Helped Create Over 700 New Materials (technologyreview.com) 28
From EV batteries to solar cells to microchips, new materials can supercharge technological breakthroughs. But discovering them usually takes months or even years of trial-and-error research. Google DeepMind hopes to change that with a new tool that uses deep learning to dramatically speed up the process of discovering new materials. From a report: Called graphical networks for material exploration (GNoME), the technology has already been used to predict structures for 2.2 million new materials, of which more than 700 have gone on to be created in the lab and are now being tested. It is described in a paper published in Nature today.
Alongside GNoME, Lawrence Berkeley National Laboratory also announced a new autonomous lab. In partnership with DeepMind, the lab takes GNoME's discoveries and uses machine learning and robotic arms to engineer new materials without the help of humans. Google DeepMind says that together, these advancements show the potential of using AI to scale up the discovery and development of new materials.
GNoME can be described as AlphaFold for materials discovery, according to Ju Li, a materials science and engineering professor at the Massachusetts Institute of Technology. AlphaFold, a DeepMind AI system announced in 2020, predicts the structures of proteins with high accuracy and has since advanced biological research and drug discovery. Thanks to GNoME, the number of known stable materials has grown almost tenfold, to 421,000. "While materials play a very critical role in almost any technology, we as humanity know only a few tens of thousands of stable materials," said Dogus Cubuk, materials discovery lead at Google DeepMind, at a press briefing.
Alongside GNoME, Lawrence Berkeley National Laboratory also announced a new autonomous lab. In partnership with DeepMind, the lab takes GNoME's discoveries and uses machine learning and robotic arms to engineer new materials without the help of humans. Google DeepMind says that together, these advancements show the potential of using AI to scale up the discovery and development of new materials.
GNoME can be described as AlphaFold for materials discovery, according to Ju Li, a materials science and engineering professor at the Massachusetts Institute of Technology. AlphaFold, a DeepMind AI system announced in 2020, predicts the structures of proteins with high accuracy and has since advanced biological research and drug discovery. Thanks to GNoME, the number of known stable materials has grown almost tenfold, to 421,000. "While materials play a very critical role in almost any technology, we as humanity know only a few tens of thousands of stable materials," said Dogus Cubuk, materials discovery lead at Google DeepMind, at a press briefing.
Cool (Score:2)
Re:Cool (Score:5, Insightful)
Sounds very expensive so I will likely never benefit from it.
Umm. I think the whole point is that it's cheaper than the old way.
Why do you think you won't benefit from, say, a better battery cathode?
Re: (Score:2)
lab (Score:2)
Re: lab (Score:2)
This is how we end up with lots of grey goo
Sooo.. (Score:2)
Re: (Score:2)
No clue.
But, what I do know. This is patent law and not copyright. You can't patent natural laws. You can't patent math, but things like RSA were patented so the line isn't super clear. I miss the old days of slashdot where we could have counted on a well known patent attorney chiming in.
Re: (Score:2)
IANAL but I thought patenting materials was absolutely possible, so long as it's not something that would have already existed in nature? For example, if I'm mucking about in a lab and create a material that is stronger than buckminsterfullerene and easier to work with, I might be able to patent it. But then, if I realize that it naturally occurs when object a collides with object b at the same time as being around object c, I would not be able to patent it.
I, too, miss having known attorneys chime in.
Re: (Score:2)
if I realize that it naturally occurs when object a collides with object b at the same time as being around object c, I would not be able to patent it.
Isn't that just chemistry? How else are you going to create it?
From my rough knowledge of patents (IANAL), patents aren't on things. They are on implementations. You may be able to to patent a method of creating PTFE. You could patent the process of using PTFE to make a non-stick pan. But you can't patent the chemical itself.
Re: (Score:3)
Re: (Score:2)
They released their predictions under a license that forbids commercial use. I don't know what would happen if they tried to enforce that. You can't patent a material, only a use of the material. Copyright isn't designed for this sort of thing. They can't claim trade secret, since they published it. So it's not clear they have any legal basis to restrict it.
Re: (Score:1)
If they can't claim trade secret on it, or patent it, neither can anyone else, so at least it can't be hijacked by some other entity, but I wouldn't want to be going up against Google's deep pockets in patent court trying to prove that their patent was invalid.
Re: (Score:1)
Its a computer program. By this logic, anything built with a spreadsheet must be public domain, along with anything written by a wordprocessor, compiled with a compiler, etc etc etc. Of course the people who built the machine to help them calculate the end result have rights to it.
Needs Relativistic Model (Score:2)
The paper says they use VASP to calculate the energies but the VASP site seems to suggest that their relativistic option only accounts for effects near the nucleus and is super compute-intense.
https://www.vasp.at/wiki/index... [www.vasp.at]
Probably not worth it as the interesting effects are in the outer orbitals.
e.g. we use lead acid batteries instead of tin-acid batteries due to outer-shell relativistic effects. They're like six times better instead of a classical 15% better or whatever.
It will be cool when this system
Re: (Score:2)
I wonder how much data they are passing around the properties of base or simpler structures that they are creating. Emergent, interesting properties (conductivity, magnetic, etc.).
Analyzing the "evolution" from simple stable materials to more complex would be very interesting. I know little of material science, but am up on organic evolution.
Fascinating stuff.
It didn't "create" the new materials (Score:5, Informative)
What DeepMind did was help predict new materials. Most of them haven't been made yet. Their properties are predicted by means that are known to be fallible.
This is a list of 700 materials that labs might find it interesting to investigate. Worthwhile, but hardly living up to the headline.
Re: (Score:2)
The paper states "736 structures have been independently experimentally verified, with six examples shown", and refers to these six papers:
Re: (Score:2)
Ah. That's why 700 stuck in my mind, when they hypothesized a much larger number. OK, they've built sample of over 700 of them. That's really just the start of deciding whether something is useful or not.
This is a very useful first stage of investigation. Don't overhype it. Things are moving fast enough without the excess hype.
Re: It didn't "create" the new materials (Score:2)
What's so bad about hype? Can you ignore it or are you addicted?
How it's going (Score:3)
Before AI: "we've used computer simulations to predict 4 materials that may have a certain property in a computer! Our team is now working to design lab processes that will try to produce each of these materials in a lab and see which of them actually works... this time next year we hope to announce something exciting!"
After AI: "we had the computer check out hundreds of millions of possible materials. It thought 2.2 million of them might be stable, so we had it pick 700 of the most promising. Then we had the computer design manufacturing processes for the most promising and, uh, yeah, so we made 700 new materials last week and, uh, yeah, we just have a couple hundred technicians here checking all of those in real life now. So, yeah, I'm hoping to sign a ten-billion-dollar production deal and maybe retire to the Bahamas by Christmas, fingers crossed!"
Almost makes me wanna be an AI engineer, before the AI engineers are replaced by AGIs. And hey, don't worry, the AGIs will never harm us. That would be too sci-fi! Now, I know, technically psychopathy is defined by what is absent, not by what is present, but billions of investment dollars are flowing into building the first AGI. Surely our business executives will produce a perfectly reliable morality module that cannot be bypassed even by an egregious mistake in the config file! To do otherwise would just be inviting lawsuits!
Google should use it to stop fake DMCA reports (Score:2)
Until we see AI used for practical solutions for regular users it will for many look like a toy.
Easy to invent (Score:2)
The future for everything is this (Score:2)
I can see a future where simulations can be so powerful that computers can simulate unknown physical and chemical processes and reactions with a high degree of accuracy. Once that is available, they can brute force billions of combination of materials and techniques to create all kinds of new things including medicines, batteries, fuels, food, and more. Eventually we may even be able to have a digital version of a living being and go even farther. I wouldn't be shocked if someone heard Elon say its only 2 y
Really bad summary (Score:3)
Thanks to GNoME, the number of known stable materials has grown almost tenfold, to 421,000.
Everything about that statement is wrong. They started from a database of 150,000 known materials. They came up with 2.2 million new materials that they predicted could possibly exist, but not necessarily be stable. Of those, they predicted that 380,000 should be stable. Only about 700 of them have actually been verified experimentally. That means the number of known stable materials has grown by about 0.5%.
Wired has a better article [wired.com] about it.
Re: (Score:1)
Called GNoME, the software was trained using data from the Materials Project, a free-to-use database of 150,000 known materials overseen by Persson. Using that information, the AI system came up with designs for 2.2 million new crystals, of which 380,000 were predicted to be stable—not likely to decompose or explode, and thus the most plausible candidates for synthesis in a lab—expanding the range of known stable materials nearly 10-fold
Perhaps the 10-fold number might make sense if many of the 150k known materials are not stable.
My fav material is (Score:1)
...Elevenfingerum.
Finally! It's the Year of AI (Score:2)
The solution turned out to be, we just had to describe all computer programs as being AI. So easy!
Scientists originally predicted that all things computer would be described as "Cyber", although "Crypto" would later turn out to be a strong contender.
mumble mumble CyberCryptoAI Overlords