Domain expertise?
That's not how we do things in software engineering.
AI on its own may not be as useful for discovering new materials as Google's DeepMind team has suggested. Two materials scientists affiliated with UC Santa Barbara have analyzed a Google paper published in Nature last November and conclude that it promises more than it delivers. Google's DeepMind team, however, disagrees and …
> "While the methods adopted in this work appear to hold promise, there is clearly a great need to incorporate domain expertise in materials synthesis and crystallography."
Which is what you'd expect from a technology in its infancy.
I am sure the reasoning behind Google's paper was as much to publicise their process, as to report loads of potentially usable new materials.
We all know that for scientists, the quantity of publications is often of greater importance than their quality. Since so few papers report world-changing phenomena. On that basis, the Google scientists are engaging in exactly the same behaviour as their detractors. Using the currently trendy keywords (AI and all that flows from it) to bolster their reputations.
Essentially it is an exercise in self-promotion, as so many other publications are.
Knowledge be damned! From now on, I vote for belief, hearsay, truthiness, gossip, religion, upvotes, wishful thinking, mythification, salespersonship, sophistry, scant evidence, and a little closet S&M whip bang on the weekends! It's so much more satisfying ...
《2.2 million, of which nearly 400,000 are believed to be stable》
Reminded me of the old jibe against String Theory "not so much a theory of everything as a theory of anything." :)
Enumerating a range of inorganic crystalline compounds whose physical and chemical properties would be fairly predictable is fairly low hanging fruit, which probably could be just as effectively done without the benefit of LLM, and is probably more a testament to the sheer amount of CPU/GPU cycles the googleers can throw on the fire of AI vanity.
I am a reasonable chap☆ I will give you as many C, N, H and O atoms as you might need - show me the same trick with all the compounds containing ring (cyclic) structures you can enumerate. :)
☆Probably not.
《2.2 million, of which nearly 400,000 are believed to be stable》
and of which a few hundreds, or less than 0.01% have been synthetized, and 0 has proved novel or useful, otherwise Google would have been quick to boast about it. A real achievement indeed.
Perhaps this is the best way to do science? Make so many claims that people can't disprove them all without using the same computer tools you're using.
Even better if you can get them to pay to use your datacentres when doing it.
It's a bit like arguing with a conspiracy theory fan, or Russian propaganda. As soon as you've shot down one claim, as demonstrably false, they've brought up 3 others so you run out of time before they run out of arguments.
Gish Galloping the chemists!
When this was first reported in El Reg, I posted this... <engage smug mode>
I'm not a materials scientist, but it does sound a bit like an over-enthusiastic comp-sci major has bounded up to Paul Hollywood and told him they've generated 380,000 recipes for cake by working out every possible combination of eggs, flour and water...
How useful is this? You've made the finest green(*), it's sitting there sparkling on the lab bench... now what? How do you figure out if it is remotely useful, and under what circumstances? A lot of our currently exciting materials are only useful when you apply specific materials to their surfaces, after you've sliced and diced them in a particular way, then kept them at a specific temperature whilst some combination of electricity, light, radiation, water, gasses, other chemicals are applied in a very precise manner.
Worse still, you've got 379,999 other crystals lined up to test..
I'm no fan of ludicrous AI claims but the 'trifecta of criteria' here seems to really support the continued employment of material scientists.
They are not claiming that the results are fake, merely that experts such as themselves will be required to define whether something is novel or useful.
This might be the case but nevertheless results in a system where the AI makes the discoveries and the materials guys only validate.