Chemistry and Artificial Intelligence

July 19, 2018 1:45 PM

Chemistry and Artificial Intelligence can give you all possible outcomes Chemistry coupled with Artificial Intelligence builds a robot that can give all possible outcomes/results from self analysis.
posted by Yellow (3 comments total)

Lee Cronin is a self-promoting hypebeast who loves the sound of his own voice, but this is neat work.

The automated analysis and feedback is cool, the ML to my eyes looks reasonable....but

I fear the popsci writeups will focus on the hype like below..

1)"The result is a system that can figure out every reaction that's possible from a given set of starting materials."

bullshit. the paper says nothing of the sort. what it says is we can train a ML network on reactions that work and don't work, and use ML classifiers to predict how reactive things will be in new combos and with related compounds (note: NOT what kind of reactions will happen, just some vague "reactivity" classifier) when mixed in new combos.

this is still very imprecise, but the algorithms are getting better. useful for fishing expeditions in molecular space (“where should I look, not what will I find”).

2) "by realizing only 10 percent of the total number of reactions, we can predict the outcomes of the remaining 90 percent without needing to carry out the experiments." This currently ONLY applies to a very specific subset of the type of chemical space his group and robot explore.

because.... there are two kinds of chemical reactions in the paper

the first type is the one in the paragraph above. 10/90 ain't gonna happen on that kind of problem yet.

the second type explores very tightly a series of VERY LIMITED related reactions that all occur via one proposed mechanism. this is sticking together two very specific and similar "aromatic" molecules through a very specific bond forming reaction.

There are 5800 potential combos of reagents, solvents, catalysts and conditions, and what cronin does is train a neural net on 580 reactions, then get the network to predict what will happen in the next 580 and so on.....

the 10/90 conclusion DOES apply, and is impressive. however, the paper itself notes that other ML chemistry groups have done this successfully though....


In summary: neat, but this is a small proof of principle step, mostly because i) machine learning is hot ii) robotic stuff is hot iii) lee cronin knows how to sell.

---

When it gets combined with this kind of chemistry software, then we're in for some fun....
posted by lalochezia at 3:03 PM on July 19, 2018 [8 favorites]

And this is AI how?
posted by GallonOfAlan at 12:07 AM on July 20, 2018

And this is AI how?

They build an AI model (NN I think) based on reaction inputs (starting materials, temperature, etc.) to predict the reaction result. Each reaction gets fed back into the model so it gets better as it gets used on a given set.

It's AI the same as pretty much everything called AI is. Just instead of learning to recognize interesting faces it's recognizing interesting reactions (or creating them, if that's your preferred metaphor.)
posted by mark k at 7:13 AM on July 20, 2018