Hacker News Reader

Comment by raziel2701 | Hacker News Reader

Man, I'm feeling stronger about LK-99 being it. This paper is theoretical and she finds that particular Cu substitutions onto specific Pb atomic sites are key to enabling a band structure that is usually linked to high Tc superconductors.

What this means for the more practical minded is that the synthesis of superconducting LK-99 is not trivial and you need to make the appropriate substitutional alloy for this to work.

This is a DFT paper, and a band structure that is usually seen in high Tc superconductors just naturally came out. She also talks about the strong electron-phonon coupling that naturally arose from the structure, which is always necessary for superconductivity.

I am, by far, the most excited I've ever been about this being a RT, ambient pressure superconductor.

View on HN

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.

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.

Based on your descriptions, if LK99 is real, it sounds like there are closer to infinite than 0 of these materials. If there are so many combinations and somebody randomly stumbled on one, it makes sense that if the number of combinations is asymptotically infinite, that number of superconducting materials is very large.

The other thing is that a lot of the results depend entirely on what you put into them. Sounds lovely and obvious, but really isn't. Which basis set would you use for simulating your molecular object of interest? There are hundreds to choose from, all developed for different purposes and under different assumptions (eg "frozen core", Born Oppenheimer, Slater orbitals, etc). It's very much not trivial to know what you need, because what the system does answers that question, and in order to find out what the system does you need to solve it -- for which you need a basis set!

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...

This looks similar to the protein folding problem. Maybe an AlphaFold-like approach could work?

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.

GNNs and graph transformers are the current state of the art methods for this kind of crystal property prediction task. One drawback is that they don’t seem to capture long range structure all that well, and the current generative models (which are really cool) based on GNNs don’t seem to take advantage of symmetry that well

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.

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.

There are linear scaling DFT codes but they’re not available under open source licensing, only a strange license:

https://onetep.org/

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.

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/b...

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.

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.

What are all the material science ML companies doing? This would be a perfect technology demo if they can find the target

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.

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.

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.

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.

yet.

And then you have to synthesize your candidates to ensure that what you think will happen really happens.

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

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..

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

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

I’m reminded how after every LHC result, a million theoretical papers come out that are able to provide a theory which explains it.

Is the solid state theory space similarly underdefined that you can find a theory for every result? Or is this paper significant?

Disclaimer: this is not my area of expertise in the slightest.

If we have the ability to computationally determine these things without any experimental data needed, and we know we're looking for a specific band structure, wouldn't we just do an automated search of possible chemistries to find everything producing said band structure?

Then just whittle down that list to the easiest to produce and most common materials for the first to test... what am I missing?

The parameter space for such a search even with a limited number candidate materials is immense. You'd need to guide the search somehow, that band structure might be the one, or it may not be... and every candidate that you flag will have to be synthesized which may not be all that easy.

Sounds like a good challenge for (Open)AI

they are doing that likely.

Google is doing XC-functionals and Alphafold. Facebook catalysts.

What makes a language model well-suited to this task?

I meant to say DeepMind

>> If we have the ability to computationally determine these things without any experimental data needed, and we know we're looking for a specific band structure, wouldn't we just do an automated search of possible chemistries to find everything producing said band structure?

Isn't this a plot point in that one Star Trek movie (episode?) where they go back in time and program a current-day computer to do this?

edit: OK, I misremembered. Was thinking of this: https://www.youtube.com/watch?v=LkqiDu1BQXY

Does anyone know if there's a way to ensure those "particular Cu substitutions" happen at the correct atomic sites? Or I guess what's the way forward in terms of synthesizing

If you figure out how to ensure it gets dropped at the higher energy subsitution site, you would get a nobel prize as well.

And another one if you determine whether copper is the optimal bond (as opposed to gold et al)

I’m not an expert in anyway but when I see detailed chemical compositions in an arxiv summary, a patents, and multiple publications, it’s almost like it’s ready to smile at all the scrutiny.

It doesn't reproduce: https://arxiv.org/abs/2307.16802. That doesn't mean there's not something to further investigate, but LK-99, at least as described in the paper, is not it.

It doesn't reproduce in that case, which is a useful data point but may not be the final word. The article linked in this thread suggests why making it may not be all that easy.

I'm just saying that it's not really LK-99, then. It's something else that needs to be specified more precisely.

I'm not following you. The compound is specified, the process is what is poorly specified and the researchers more or less admit that they do not exactly know what the right method is. This makes it a lot harder to replicate and may well cause a lot of attempts to misfire before someone finds something that works or we give up on the search. What TFA here suggests is that there are some potential complications in manufacturing this that would explain why the original researchers had problems trying to make the compound, the number that I recall is that '1 in 10' tries resulted in a working sample.

That would also be a fantastic way to pull a hoax, because it will result in 10x more effort spent on your hoax. If it turns out that it was a hoax I think the original researchers will find immense gratitude from Pons and Fleischmann for taking over the top spot for the textbook example of bad science. But for now, as far as I can see the jury is still out, and if anything the paper linked here actually improves the chances of it being real a bit more than it is offset in the other direction by a failed replication attempt.

Then why do we have separate names for Graphite and Diamond? Same element, different molecular (crystalline) structures. If CuO25P6Pb9 can exhibit superconducting properties, and reproduction attempts are failing, then there must be a specific isomer that needs to be achieved, the creation of which is not described in the LK-99 paper. I'm just trying to understand how to wade through this stuff, I could be wrong, and I dropped solid state physics because I hated it, but that's how I currently understand the situation.

Edit: s/isomer/polymorph

Same molecular structure entirely different crystal lattice.

It is the formation that is different not the molecules. C = C. They are technically allotropes:

https://en.wikipedia.org/wiki/Allotropy

And just like here that's a function of how the bulk carbon got to be formed, under extreme pressure and temperatures or less pressure and temperature. It's the recipe that makes the difference, not the ingredients.

As far as I understand, we're saying the same thing.

    isomer
    /'aɪsəmər/
    noun
    a compound that exists in forms having different arrangements of atoms but the same molecular weight

I'm using isomer to refer to possible different arrangements of "the LK99 compound". I see you're saying allotropes refers to diamond vs carbon (allotrope referring a single element vs isomer referring to compound).

That raises a question, what's the name for different crystalline structures of the same compound? Is that still an isomer or something different? I'm out of my element here.

An isomer can have different valences but still have the same overall chemical formula.

So just for an example (I have no idea how to put a tetrahedron in a comment :) ) a chain of C-C-C-C and C-C=C-C (which you likely can not synthesize) would be an isomer but C-C-C-C in one crystal lattice versus C-C-C-C in another crystal lattice would be allotropes.

The carbon in graphite sits in sheets (hence the possibility to form graphene), whereas the carbon in a diamond always forms three dimensional lattices.

edit nah: bad example, sorry, I can't find a good way to visualize this in text the 2nd C-C-C-C chain should have the last C dangling from the 2nd down.

edit2: examples are wrong, I should have used a more complex molecule for the isomer and the carbon can have either three or four other carbons hanging of it for the allotrope. tricky...

Just double checking, the definitions I'm finding for allotrope are specific to a single element.

    allotrope
    /'ælə,troʊp/
    noun
    a structurally different form of an element

I'm having trouble finding allotrope used to refer to different forms of a compound.

https://www.differencebetween.com/difference-between-allotro...

Yes, you're right, I should have used a more complex example for an isomer.

Isomers are structurally different. Allotropes refer to variations of the same element.

You're looking at the word "polymorph": "(chemistry, geology) Any substance or mineral that forms different types of crystal."

Thank you.

Pons and Fleischmann, not Jan Hendrik Schön?

Wow that's a bad one, I never even knew about it, thank you for pointing that out. And yes, he's probably worse, with Pons & Fleischmann I never managed to rule out that they themselves were 'true believers' that had deluded themselves rather than fraudsters. But with this guy there is little doubt.

Oh yeah, he is absolutely fucking wild. There's a great documentary on him, here's part one, the other two parts of the documentary are also on YouTube on the same channel: https://youtu.be/nfDoml-Db64

> Pons and Fleischmann ... bad science

They found a new phenomenon, that took years to explain. They didn't do anything bad.

Defining LK-99 as “the thing you get from following the steps in this paper” and not “the allegedly superconducting material these guys have a sample of” is silly.

Why? I've always understood it to be the thing the paper describes.

You say potato... The paper is fucking vague on all counts, from synthesis to result. "LK-99" at this point is understood to be most likely a mixture of compounds. It is not fully defined, it is not one "thing," it's broad proposal that encompasses a family of materials.

Essentially all of my knowledge of materials science comes from reading HN's discussion of this over the last ten days, but I've read repeatedly that these kinds of synthesis processes are not very reliable. Even with well known materials with well known "recipes," someone trained can follow a procedure apparently to the letter and get no result, and then repeat the procedure apparently the same and get a good result.

Semiconductor yield at high end fabs can be lower than 20%. If you tried that process at a new fab and wound up with a complete failure, would you say "well that process just doesn't produce any semiconductors" or would you think you might need to refine the implementation of the process.

What the simulation is saying though is that it requires some luck, not all the arrangements result in the band structure.

I understand. I'm saying that it's not LK-99 as described, then. Rather, it's something else that needs a more precise specification.

Ah no they called the working prototype LK-99 so if it works it is if it does not then it is not. Instead of what you are saying that it works so is not LK-99.

Or it's lk-99 as described, 5% of the time... basically they need to tune the process to improve yields. I'm not sure if that means it wouldn't be LK-99 if that's all that's going on here.

/me skeptical either way

Reply on news.ycombinator.com