« first day (4125 days earlier)      last day (657 days later) » 

kirma
kirma
14:05
@b3m2a1 How large molecules one can model at the quantum level these days without computational requirements completely blowing up?
Meta
Meta
14:29
0
Q: Question regarding user etiquette in answers and comments

ubpdqnPlease see this question In particular, refer to my deleted answer. Firstly, I regret my triggered response. I state extreme overwork and illness as a mitigating factor. This is not an excuse but a context. Secondly, I merely converted a complex analysis problem to a surface integral. The surface...

discussion etiquette users
 
3 hours later…
b3m2a1
b3m2a1
17:45
@kirma you pick your battles, but even my little laptop can do some DFT calculations on molecules with up to say 5-10 atoms from the first and second rows of the periodic table (and maybe a Cl or Br if I'm feeling wild) and from those there are any number of clever ways to model portions of the reactivity or get IR spectra for characterization
If you have access to something with say 24 cores and maybe 70 Gb of ram (not really that huge honestly) you can get up to like 20 atoms using more accurate techniques
The beauty of a lot of this is that chemistry is largely a local process and so in the majority of cases going much larger than that has diminishing returns in terms of what you can really learn and model, which is really why organic chemistry is so successful
For things like metal-containing compounds there are cleverly constructed core-shell potentials that allow you to keep your calculations from blowing up, but I am not an expert in those
There have also been a number of attempts to get GPU-accelerated algorithms for these kinds of calculations, but unfortunately the types of integrals people have to solve and the iterative nature of the algorithms don't provide easy solutions and so instead people are spending more effort on training ML-models to learn the expensive portions of the calculations in a semi-transferable manner, which could also be implemented in Mathematica to provide a relatively fast approach
In my experience, these ML-based approaches allow you to get highly-accurate energies for systems with up to like 10 atoms in a matter of seconds as opposed to half an hour or more if you're using the methods they are based off of

« first day (4125 days earlier)      last day (657 days later) »