Machine learning can help robots perform chemistry experiments faster than fleshy boffins, according to research published in Nature. Researchers have been exploring how algorithms can predict the outcome of chemical reactions for a while, but this project goes one step further and actually uses a real robot to carry out some …
“By realizing only 10 per cent of the total number of reactions, we can predict the outcomes of the remaining 90 per cent without needing to carry out the experiments,” the paper said.
That sounds like a great way to never find any new type of reaction or interesting compounds.
Over-confidence and smugness built into the AI.
That is a real risk. As has been said before: Scientific discoveries usually start with "hey, that's odd" rather than "eureka!"
More importantly, will we be able to stand the intolerable air of smugness these AI lab assistants generate whenever they have performed an experiment well?
(Doffs hat to the late, great Douglas Adams)
That was my thought. Automated processes can conclude did/didn't work.
But it takes proper intelligence (and imagination) to conclude that something unexpected and interesting has happened and find out what or why.
AI would have discarded a Petri dish, not discovered penicillin.
In a way it's even worse than that. If you had chemists work out which reactions would occur, you are left with a set of rules bringing you more understanding into those kinds of reactions.
If you just train an AI to predict something, you gain no understanding at all. You just get a black-box which may or may not predict the outcome correctly without actually giving you reasons for it.
This is the first article I read about statistical analysis machines (ie pseudo-"AI") that does not crow to high heavens the incredible benefits that AI will bring to the proceedings as if all issues are already solved, as is usually the case.
The article actually states "It all hinges on human expertise to tweak the algorithms behind it and the AI is only as good as its trained to be." - that is a first.
Fantastic blog from an experienced chemist:
What if the AI is stupid enough to try brewing up FOOF for instance?
In the blog Derek describes compounds so reactive that if they go on fire and you throw the sand bucket on them.. the sand catches fire...
I'm no chemist, but I have seen notices in chemistry departments about compunds which smell so bad you have to notify people. What if the AI fancied a little batch of one of them?
It's very easy and quick to compare chemical similarity, they could set filters for known toxic chemicals, explosives, illegal chemical classes etc. In drug discovery we're also looking for new chemicals so we'd also filter out our own or competitor molecules.
What the "fun filter" would hopefully prevent -->
From that setup, it looks like it's going to be pretty limited in terms of reaction conditions.
I read about this on the Beeb and was singularly unimpressed. Part of that was the claim that this will "allow chemists to discover new wonder drugs" or similar, as if chemists actually just throw stuff together in the hope that it'll a) react and b) produce something useful. Nobody with half a brain does high-throughput screening anymore.
First note the emphasis on organic chemistry, not the insane stuff of TIWWW.
But this obsession with neural networks seems kind of dumb.
AI has been applied to crystallography since the mid 80's using blackboard techniques and network analysis methods (the "synthon" seems to a key idea). IOW there's lots of information about reaction pathways, what works and what does not, already. There's also quite a lot of software that applies thermodynamics (not quantum simulation) to decide if a molecule is viable, but AIUI it's got a pretty bad UI. Now can this thing work out that a molecule with 15 N atoms (mostly in its backbone) is even possible, let alone how (very carefully) to synthesize it?
TBH while an automated synthesis rig is hardly AI the fact it is (from the article) about 4x faster than a human (presumably a chemist in training, who needs the practice) sounds good if you want to "brute force" all the variations of a substance (like a high temperature superconductor) you've just found has excellent properties in one area but is poor in other ways.
But that doesn't help you discover the core compound in the first place.
Frankly I'm amazed that in the 2nd decade of the 21st century automated chemical discovery labs are not more common (if not SOP).
Biting the hand that feeds IT © 1998–2019