5 ms·

If this could be simulated, can you help me understand why we couldn't have used simulation to find promising SC materials to investigate further earlier? Are t

If this could be simulated, can you help me understand why we couldn't have used simulation to find promising SC materials to investigate further earlier? Are there just too many permutations to investigate? It seems to my own naive self that if LK99 is the real deal, we mostly just got lucky finding it.

ivalm 3y ago
You have to put in the structure and then it's expensive to do the calculation. The space of possible structures is extremely large. If you have candidates then you can run through them, but you can't just random search through trillions of trillions of candidates.
- m463 3y ago
  yet.
  - jacquesm 3y ago
    And then you have to synthesize your candidates to ensure that what you think will happen really happens.
- KRAKRISMOTT 3y ago
  What are all the material science ML companies doing? This would be a perfect technology demo if they can find the target
  - jacquesm 3y ago
    https://www.mpg.de/20096180/artificial-intelligence-in-material-design https://www.mpg.de/20096180/artificial-intelligence-in-mater... 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.
  - throwawaymaths 3y ago
    The problem is that this stuff is severely nonlinear, and for the raw formula there are not that many degrees of freedom to try. If you get a structure, there is no guarantee it's stable, or accessible by our synthetic techniques. Obtaining the training data is also likely to be tricky.
  - Hermitian909 3y ago
    I've worked on the CS side of new material discovery. 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.
    - doingtheiroming 3y ago
      The positive view here is that we are 5-10 years from being able to search for materials computationally using AI. That would potentially be more valuable than a single RTS.
tired_and_awake 3y ago
DFT scales horribly so it's phenomenally expensive to run. You have to have some other mechanism for knowing the general atomic layout before entering the DFT realm. Once you know the atomic positions you can then do little perturbation simulations to model phonon dispersions or ask electron density questions.
- rpep 3y ago
  There are linear scaling DFT codes but they’re not available under open source licensing, only a strange license: https://onetep.org/ https://onetep.org/
  - marcosdumay 3y ago
    What the fuck? 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.
    - tired_and_awake 3y ago
      I'm not intimately familiar with it, just guessing this is some orbital free method that theyve refined. Nothing new, just another approximation that cant model some effects.
    - justinclift 3y ago
      Poking around the website, it seems to be owned by Dassault Systems and available for commercial licensing as part of their "Biovia Material Studio" product. PDF flyer about it here: https://www.3ds.com/fileadmin/PRODUCTS-SERVICES/BIOVIA/PDF/biovia-material-studio-onetep.pdf https://www.3ds.com/fileadmin/PRODUCTS-SERVICES/BIOVIA/PDF/b...
akasakahakada 3y ago
Not an expert but it just happen that my lab is full of DFT folks so I heard a lot about those everyweek. As people above already answered the questions, I gonna talk some extras. 1. Computation cost is large. 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.
- fpordeig 3y ago
  This looks similar to the protein folding problem. Maybe an AlphaFold-like approach could work?
  - fock 3y ago
    not soo similar really, but yes, alphafold-style generative models could help find realistic structures for a specific composition. However a) data is much worse (I would say so at least...), b) the clever tricks of alphafold centered around strings of aminoacid don't really apply to particles in a box... and c) search space might be even larger if you go to interestingly sized systems. 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.
  - akasakahakada 3y ago
    We exactly considering this since this year. But there are some major problems that cannot be solved in short term. Inorganic crystal structure database (and there is one database literally 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.
Sharlin 3y ago
It's like NP problems. It's much harder to find a solution than to check if a candidate solution is valid.
marcosdumay 3y ago
Yes, the permutations are endless, and verifying each one is a bit expensive (up to extremely expensive, depending on the complexity of the material).
foobiekr 3y ago
Because it's too hard. People vastly over-estimate what we can simulate at any level of fidelity with any scale below purpose-built stuff running on supercomputers..
currymj 3y ago
cheaper alternatives to DFT (using ML especially) for this purpose is an active research area

ivalm 3y ago

You have to put in the structure and then it's expensive to do the calculation. The space of possible structures is extremely large. If you have candidates then you can run through them, but you can't just random search through trillions of trillions of candidates.

m463 3y ago
yet.
- jacquesm 3y ago
  And then you have to synthesize your candidates to ensure that what you think will happen really happens.
KRAKRISMOTT 3y ago
What are all the material science ML companies doing? This would be a perfect technology demo if they can find the target
- jacquesm 3y ago
  https://www.mpg.de/20096180/artificial-intelligence-in-material-design https://www.mpg.de/20096180/artificial-intelligence-in-mater... 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.
- throwawaymaths 3y ago
  The problem is that this stuff is severely nonlinear, and for the raw formula there are not that many degrees of freedom to try. If you get a structure, there is no guarantee it's stable, or accessible by our synthetic techniques. Obtaining the training data is also likely to be tricky.
- Hermitian909 3y ago
  I've worked on the CS side of new material discovery. 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.
  - doingtheiroming 3y ago
    The positive view here is that we are 5-10 years from being able to search for materials computationally using AI. That would potentially be more valuable than a single RTS.

tired_and_awake 3y ago

DFT scales horribly so it's phenomenally expensive to run. You have to have some other mechanism for knowing the general atomic layout before entering the DFT realm. Once you know the atomic positions you can then do little perturbation simulations to model phonon dispersions or ask electron density questions.

rpep 3y ago
There are linear scaling DFT codes but they’re not available under open source licensing, only a strange license: https://onetep.org/ https://onetep.org/
- marcosdumay 3y ago
  What the fuck? 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.
  - tired_and_awake 3y ago
    I'm not intimately familiar with it, just guessing this is some orbital free method that theyve refined. Nothing new, just another approximation that cant model some effects.
  - justinclift 3y ago
    Poking around the website, it seems to be owned by Dassault Systems and available for commercial licensing as part of their "Biovia Material Studio" product. PDF flyer about it here: https://www.3ds.com/fileadmin/PRODUCTS-SERVICES/BIOVIA/PDF/biovia-material-studio-onetep.pdf https://www.3ds.com/fileadmin/PRODUCTS-SERVICES/BIOVIA/PDF/b...

akasakahakada 3y ago

Not an expert but it just happen that my lab is full of DFT folks so I heard a lot about those everyweek. As people above already answered the questions, I gonna talk some extras. 1. Computation cost is large. 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.

fpordeig 3y ago
This looks similar to the protein folding problem. Maybe an AlphaFold-like approach could work?
- fock 3y ago
  not soo similar really, but yes, alphafold-style generative models could help find realistic structures for a specific composition. However a) data is much worse (I would say so at least...), b) the clever tricks of alphafold centered around strings of aminoacid don't really apply to particles in a box... and c) search space might be even larger if you go to interestingly sized systems. 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.
- akasakahakada 3y ago
  We exactly considering this since this year. But there are some major problems that cannot be solved in short term. Inorganic crystal structure database (and there is one database literally 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.

Sharlin 3y ago

It's like NP problems. It's much harder to find a solution than to check if a candidate solution is valid.

marcosdumay 3y ago

Yes, the permutations are endless, and verifying each one is a bit expensive (up to extremely expensive, depending on the complexity of the material).

foobiekr 3y ago

Because it's too hard. People vastly over-estimate what we can simulate at any level of fidelity with any scale below purpose-built stuff running on supercomputers..

currymj 3y ago

cheaper alternatives to DFT (using ML especially) for this purpose is an active research area