/sci/ - Science & Math » Thread #11241973

210KiB, 639x426, MIT-Predicting-Reactions-1_0.jpg

View Same Google iqdb SauceNAO

I would like to get into computational chemistry

Anonymous Sat 21 Dec 00:34:22 2019 No.11241973 View Reply Original Report

Quoted By: >>11242228 >>11242289 >>11242321

underfag CS major here, lately ive been interested in the field of computational chemistry. My ultimate vision is to have a computer program/AI that is able to compute/discover new materials with different properties without having to explicitly synthesize them. A kind of a algorithmic way or at the very worst, a brute force way to come up with new materials. Anyone here with actual chemistry background/knowledge tell me why this can or cannot be done? I imagine somewhere in some library there exist tables of materials which have been synthesized an whose properties have been studied. What stops me from say feeding these tables into an AI that will combine them and calculate the properties of these new materials?

Here are some articles for anyone interested which apply AI for chemical problems

http://news.mit.edu/2017/computer-system-predicts-products-chemical-reactions-0627

https://deepmind.com/blog/article/alphafold

Anonymous

Anonymous Sat 21 Dec 2019 02:37:40 No.11242228 Report

Quoted By: >>11242265 >>11242268 >>11242334

>>11241973
I work in quantum computing with a focus on algorithms for electronic structure, so I can shed a little light. I don't work directly with the classical methods but I've used software packages and I know the general principles behind them.

I'll talk only about ab initio methods (sometimes called "Schrodinger methods", because it's basically just diagonalizing the electronic structure Hamiltonian). I don't work at all with DFT, as it's not too popular in QC, though I've seen a handful of papers with that, so it's certainly an option.

The most basic thing that ab initio quantum chemistry does is approximate the ground state energy of the electronic structure Hamiltonian. You take the Born-Oppenheimer approximation (nuclei are classical, fixed charges) and set up a nuclear geometry, e.g. near-equilibirum bond lengths/angles. Then you choose a basis of molecular orbitals to expand your Fock space with, e.g. STO-3G. These are typically contracted Gaussian-type functions for efficient computation, with various heuristic modifications depending on the molecule (restricted/unrestricted spin degeneracy, polarization, closed/open shell, etc.) Then you usually do something like Hartree-Fock (a self-consistent mean-field method) to find the best approximation to the ground state with your orbitals via a single Slater determinant (i.e. fine-tune the parameters in your orbital basis set, restricted to the n-electron manifold of your chosen molecule). This, by definition, gives you a state which is closest to the ground state (within your finite choice of basis set) without accounting for electron-electron correlations.

The name of the game, then, is to calculate the missing correlation energy. Although HF gets up to 99% of the absolute ground state energy, it's entirely useless because correlation energy is really what we care about. It's the hard part of the calculation that really requires n-electron quantum mechanical calculations.

Anonymous

Anonymous Sat 21 Dec 2019 03:01:33 No.11242265 Report

Quoted By: >>11242268 >>11242289

>>11242228
To get that last 1%, we use post-Hartree-Fock methods. The most obvious is full configuration interaction -- basically just diagonalize the Hamiltonian with respect to the Hartree-Fock orbitals. As you might imagine, this is exponentially difficult, since the Hilbert space is exponentially large (in your basis size). Even worse is that if you try to truncate, at say the k-body hopping terms, you get a size-dependent approximation (i.e. you need larger k for larger molecules).

The common approach used is called coupled cluster. This is nice because it's size-independent, meaning that you can truncate at some k-body term in the cluster operator (e.g. double or triple excitations) and your approximation accuracy will be the same regardless of the size of your molecule or basis size. Triple excitations (CCSDT) is called the "gold standard" in ab initio quantum chemistry. But as you might imagine, it's pretty expensive as well, even though technically polynomial (I've read something like O(M^7) where M is the # of atoms in the molecule, though I think this is a simplified view of the complexity parameters).

Importantly, CCSDT doesn't necessarily reach the limit of "chemical accuracy" - 1 kcal/mol, or 1.6 mHa. It obviously is practically intractable for exceedingly large molecules (say basis size of 50-100 and beyond), but also just flat out fails for highly correlated systems (a classic example is the H3 molecule - of course, you could just brute force diagonalize it since it's small enough, but CCSDT itself doesn't capture the strongly correlated effects).

This is where my knowledge of classical methods pretty much ends. Chemical accuracy is a lofty goal, especially for any complex molecule that is of practical interest, and CCSDT doesn't seem to be able to do the job in those cases. I don't know what special, "paradigm shifting" advances have been made in the field in terms of methodology (what I do know of are optimizations like low rank factorization).

Anonymous

Anonymous Sat 21 Dec 2019 03:04:54 No.11242268 Report

Quoted By: >>11242289 >>11242418

>>11242265
>>11242228
its hilarious to see chemists use big words and think they are hot shit when its very very basic stuff

thanks for confirming, chemists are barely over biologists on the totem

Anonymous

Anonymous Sat 21 Dec 2019 03:20:09 No.11242289 Report

Quoted By:

>>11242265
>>11241973
Ok, now to address your questions. Sure, you can come up with a "new" material by putting together various nuclei and geometries and a number of electrons and specifying some spin degeneracy. It's not necessarily meaningful if you can't actually synthesize it, if it's not stable, etc. (and these are deep questions in the realm of actual chemistry that I know nothing of). But certainly you can do it and you can try to estimate the ground state using the methods I mentioned. At the computational level, it's nothing but choosing a selection elements, arranging them in some way, and running the algorithm.

I think what you really need is some guideline or idea of what you want to achieve. In principle you could "put together" whatever you want, but very likely it won't have any sort of useful property. The field of chemistry is already pretty good at understanding what sorts of combinations bring about what sorts of effects. What they have trouble with is understand specific mechanisms that occur at the electronic interaction level, and that's sort of the idea of electronic structure calculations. Sure you might hit on some "never before seen combination" of seemingly random elements that have some insane feature, but I personally doubt it. Besides, the search space is just too large and the calculations become intractable real fast (at least, to the precision that you would like).

>>11242268
Sorry if you're offended for whatever reason. I'm not putting on airs like I think I'm "hot shit" and I'm not using "big words", I'm using the standard jargon so that OP can easily look them up. The literature is vast and it's nice to know where to start looking. And yes it's basic quantum chem, of course it is, OP says he's an undergrad CS, I doubt he has any experience with the field at all.

Also I said I work in quantum information, and I in no way count as a chemist, both by lack of expertise and in title (physics in undergrad and grad).

Anonymous

Anonymous Sat 21 Dec 2019 03:38:42 No.11242321 Report

Quoted By:

>>11241973
https://www.sciencedaily.com/releases/2016/08/160819114221.htm
There was an article somewhere talking about them using computer models to create a ceramic some time ago. Not AI maybe but they had the molecular structure of this or a similar ceramic made.

Anonymous

Anonymous Sat 21 Dec 2019 03:44:21 No.11242334 Report

Quoted By: >>11242390

>>11242228
Hello. I'm just a dumbass undergrad whose research revolves around the use of computational chemistry (mainly DFT). I have a somewhat mediocre idea of whats going on, and it's embarrassing on my part that I don't have a thorough understanding of whats going on. What did you do/resources did you use for deeper insight? Right now only using Crammers book.

Anonymous

Anonymous Sat 21 Dec 2019 04:27:00 No.11242390 Report

Quoted By:

>>11242334
Szabo and Ostlund is the standard text for ab initio. I think it's decent, dirt cheap since it's a Dover book. Really handy reference to carry around, that's for sure. Molecular Electronic?Structure Theory by Helgaker et al. is the more advanced text (nearly 1000 pages), but I prefer it much more - goes way more into depth with the various types of perturbation theories and special tricks you can do with CC. Plus it doesn't mess around with any 1st quantization nonsense, just starts with the Fock space right out of the gates. Of course, i say that as someone coming from physics.

TMP Chem's channel is pretty nice, if not a bit sparse/short. Still, it's nice as a supplement, but definitely not enough on its own. https://www.youtube.com/channel/UC3dZQdfv67X49cZkoXWYSwQ

I personally used a variety of these sources, mainly Helgaker, but also just sort of... osmosis from reading papers in my field. That's why my knowledge is pretty limited, I'd say - since I don't do computational chem per se, I just know the aspects that I need to know to get by. But yeah, Szabo+Ostlund and Helgaker et al. are great places to start, a bit old but good foundations.

Anonymous

Anonymous Sat 21 Dec 2019 04:43:02 No.11242418 Report

Quoted By:

>>11242268
I am an insecure computer "scientist"
the post

Anonymous

Anonymous Sat 21 Dec 2019 05:11:12 No.11242458 Report

Quoted By:

My job is applying machine learning to chemistry and drug design. Material property prediction has been done. There are a few companies that specialize in it. The difficult part is getting the models to generalize far beyond what they were trained on.

Also the reaction prediction think you linked is dated. Current models are around 90% accuracy (for forward synthesis, not retrosynthesis). It's actually pretty easy. You can frame it as a sequence to sequence text translation problem.

Capcode	All Only User Posts Only Moderator Posts Only Admin Posts Only Developer Posts
Show Posts	All Only With Images Only Without Images
Deleted Posts	All Only Deleted Posts Only Non-Deleted Posts
Ghost Posts	All Only Ghost Posts Only Non-Ghost Posts
Post Type	All Only Sticky Threads Only Opening Posts Only Reply Posts
Results	All Grouped By Threads
Order	Latest Posts First Oldest Posts First

Your latest searches