Obtaining the training data is also likely to be tricky.
They are doing this sort of thing (that's more of a research institute). The problem is that you are not looking for the compound but for the exact way to manufacture it assuming that the original sample really is superconducting.
The tricky thing is that you don't actually have that much data so ML is not even close to plug and play. If you want to get results you end up needing to pair ML with a lot of theory and some tricky algorithms to help narrow the search space and even then that space is huge.
Progress is being made but I think we're still at least 5-10 years from CS providing a real inflection in materials discovery.
Once you know the atomic positions you can then do little perturbation simulations to model phonon dispersions or ask electron density questions.
Linear scaling DFT is something way more impactful than room-temperature superconductors. What's next? "Hey, I've used my FTL spaceship to verify the material at those friendly alien's library"?
The fact that there has been no Nobel prize and we didn't spend a week around the web arguing "yes, it works!", "no, didn't work for me", "yes, I verified it!" highly implies that the site is trying to say something different than what we are understanding.
https://www.3ds.com/fileadmin/PRODUCTS-SERVICES/BIOVIA/PDF/b...
1. Computation cost is larger. 1 compute task for a small scale ~100 atoms last about 3 days to 1 week on supercomputer.
2. Search space is hugh. For each composition you can have different atomic (or crystal) structure. And here we are talking doping which means introduce impurities into the molecule. Chemical characteristics differs depending on which atom you swap for the impurity. Sometimes you may want to try all places.
3. Depends on initial values. Sometimes the initial value is just bad that the result is totally unusable, then you have tweak a little bit and throw back to supercomputer. This cycle might happen few times for 1 specific formula and structure.
4. Not 100% accurate. Often the resulting numbers are off by a few % or more which is hugh, compare to experimental results. Reason is that the simulation is not full scale, approximation is here and there to reduce computational cost.
Also there's been some people arguing about the particles in a box situation for a loooong time and the most promising approach currently is diffusion.
Inorganic crystal structure database (and there is one database literally is this name) is way smaller than what we have for proteins. Also by nature, Transformer is hardly useful for crystals because the crystal is repetitive. You don't throw the same sequence over and over to transformer and hope it will work like magic.
My current understanding is that Graph Neural Network is perfect for this job because graph can exactly describe this kind of repetitive nature of crystal.
It's hard enough to get out the hyperfine interaction to the right order of magnitude of a simple metal complex, let alone something like superconductivity, which fundamentally is a many, many body problem...
It seems to my own naive self that if LK99 is the real deal, we mostly just got lucky finding it.