I've been working with a bunch of ML and quantum chemistry lately. At this point the hardest part is to actually come up with the problem (i.e. translate chemistry into reasonable mathematical expressions) and then make all the servers and packages work. The code and NN architecture themselves are like 10-20% of the work.
@academicalnerd I mean, if you're working with ML for chemistry the utmost important problem is finding some reliable data to use; then how to represent chemical data in a machine interpretable way and then training ML models.
If you're going into NN I hope you've got a lot of data to throw at it.
@rastinza
I'm not the first one to do this, luckily. Materialsproject contains enough data for my purposes and there are other databases to support it.
@academicalnerd Looks cool, I had a friend who was working on something similar (if I correctly understood what you're doing...).
How do you represent phases in a computer? Do you place the dustances and angles of all the molecules or something like that?
There are few ways to do that, the trick is to have all the structures in the dataset have a representation of the same dimensionality, otherwise the neural network won't work. I use slightly modified SOAP descriptor from dscribe package (smooth overlap of atomic positions). You can skim through the docs if interested: https://singroup.github.io/dscribe/1.0.x/tutorials/descriptors/soap.html
@rastinza
I'm building a neural net that predicts stability of solid phases. The details are not quite settled yet, there's a lot of trial going on. Generally it's a classifier that guesses if energy above the hull is greater than zero or is equal to zero.
Convex hull is a thermodynamic concept: set of compounds that are thermodynamically stable, and their unstable polymorphs are higher in energy (hence E above the hull).