5 ms·

Describing the issues behind the cost prohibition of material science experiments and discovery, then handwaving it by just saying we can use supercomputing is

Describing the issues behind the cost prohibition of material science experiments and discovery, then handwaving it by just saying we can use supercomputing is incredibly ignorant of how inaccessible and coat prohibitive supercomputers are. Not to mention, even if your supercomputer has an idea you still need a lab that can confirm the idea.

bee_rider 1y ago
They spend a lot of text describing the fact that it wouldn’t be possible without Frontier, so I think they are aware of the cost. The whole ecosystem of little, easy to apply for compute credit grants, and then larger grants (like the one they mention applying for at the end) is part of the solution to that problem. It doesn’t seem like ignorance, but rather a choice to emphasize in the writing that they are part of the solution.
toss1 1y ago
Yes supercomputer time is scarce and expensive. Physical laboratory time, materials, and workers are also scarce and expensive. And physical lab tests can consume a lot of calendar time. So, even with the constraints and costs, using the supercomputer to run 1,000 or 10,000 tests and then physical lab tests on only the few most promising results can still less costly in both funds and especially calendar time than the costs and time to run a similar number of physical lab tests to produce the results. It is literally a balancing of the scarcity and costs of each mode of getting to the results, and those experts in the field, with their own labs, are making the claim that this is the way their research should proceed. Do you have actual information and specific knowledge that in fact proves them wrong (b/c that comment seems a lot more hand-wavey vs theirs)?
- lazide 1y ago
  Only if the simulations are usefully predictive. A lot of the physical lab tests end up showing effects we did not know exist, which is why they are interesting - simulations are historically terrible at predicting things we don’t know can exist. See the pharma adventures in simulating potential drugs in the 90’s-early 2000’s. A lot of heat a noise for little actual effect.
  - toss1 1y ago
    >>Only if the simulations are usefully predictive. Of course; if you cannot make your simulations produce realistic output you are either wasting your time or are clearly not yet done building your simulations. >> A lot of the physical lab tests end up showing effects we did not know exist, which is why they are interesting - simulations are historically terrible at predicting things we don’t know can exist. Of course, there are the lab surprises, discovery of penicillin an early famous one. I'm absolutely not saying supercomputing simulation should replace lab work — In my own shop I do lab testing well beyond where most would go to simulation, precisely because there's no substitute for seeing the physical side-effects. But when someone randomly claims the team publishing very useful work is somehow wrong to use simulation, I think the real researchers in their lab & at their terminals know better than that post.
    - lazide 1y ago
      I guess my point was less if the simulations produce realistic output, and more if they can be used to discover or model novel - but real - phenomena. The closer we get to ‘the truth’ (if such a thing exists!), the more likely this is of course. Ideally we’d have simulations that we could even discover novel physics behaviors in that match real life! Practically, most are simplified to the point that we can match behavior we already well understand, and maybe a few things can validate we don’t - which can help explore some of the gaps we haven’t been able to touch in the labs. it’s not likely we’ll discover new principles that way though. But I’m guessing I’m just preaching to the choir, eh?
      - toss1 1y ago
        Yup, seems exactly right. And I think we might be nearly to the stage you're identifying, where there is enough fidelity in the models to discover new things. I sort of recall reading at least some of that happening in some fields of math?
motorest 1y ago
> Not to mention, even if your supercomputer has an idea you still need a lab that can confirm the idea. I think this is the critical aspect. You don't need to own a supercomputer to run simulations in supercomputers. Sometimes you don't even need supercomputers at all. It might be a challenge but it's not prohibitive. What's clearly a blocker is the verification & validation work to assess if a specific model actually works. You need to run experimental tests to check if your module is any good, and you need to run experimental tests to check if the output of your computational models correlate with reality. That costs both time and money.
white_tiger 1y ago
Supercomputers are becoming more accessible. There's a large gap between research and real-world use, such numerical models need to be thoroughly tested for robustness before they can be deployed. And according to experience, such studies are usually conducted simultaneously with Laboratory experiments, or already have Lab results they can compare with. If not, such study acts as a new source of synthetic MD data which can be picked up by the experimental folks. I do agree that people who write such blogs tends to sell these ideas or studies as finished products which they are far from. The product of such studies are usually a paper or in other cases a patent. An idea others can build upon, like lego bricks.

bee_rider 1y ago

They spend a lot of text describing the fact that it wouldn’t be possible without Frontier, so I think they are aware of the cost. The whole ecosystem of little, easy to apply for compute credit grants, and then larger grants (like the one they mention applying for at the end) is part of the solution to that problem. It doesn’t seem like ignorance, but rather a choice to emphasize in the writing that they are part of the solution.

toss1 1y ago

Yes supercomputer time is scarce and expensive. Physical laboratory time, materials, and workers are also scarce and expensive. And physical lab tests can consume a lot of calendar time. So, even with the constraints and costs, using the supercomputer to run 1,000 or 10,000 tests and then physical lab tests on only the few most promising results can still less costly in both funds and especially calendar time than the costs and time to run a similar number of physical lab tests to produce the results. It is literally a balancing of the scarcity and costs of each mode of getting to the results, and those experts in the field, with their own labs, are making the claim that this is the way their research should proceed. Do you have actual information and specific knowledge that in fact proves them wrong (b/c that comment seems a lot more hand-wavey vs theirs)?

lazide 1y ago
Only if the simulations are usefully predictive. A lot of the physical lab tests end up showing effects we did not know exist, which is why they are interesting - simulations are historically terrible at predicting things we don’t know can exist. See the pharma adventures in simulating potential drugs in the 90’s-early 2000’s. A lot of heat a noise for little actual effect.
- toss1 1y ago
  >>Only if the simulations are usefully predictive. Of course; if you cannot make your simulations produce realistic output you are either wasting your time or are clearly not yet done building your simulations. >> A lot of the physical lab tests end up showing effects we did not know exist, which is why they are interesting - simulations are historically terrible at predicting things we don’t know can exist. Of course, there are the lab surprises, discovery of penicillin an early famous one. I'm absolutely not saying supercomputing simulation should replace lab work — In my own shop I do lab testing well beyond where most would go to simulation, precisely because there's no substitute for seeing the physical side-effects. But when someone randomly claims the team publishing very useful work is somehow wrong to use simulation, I think the real researchers in their lab & at their terminals know better than that post.
  - lazide 1y ago
    I guess my point was less if the simulations produce realistic output, and more if they can be used to discover or model novel - but real - phenomena. The closer we get to ‘the truth’ (if such a thing exists!), the more likely this is of course. Ideally we’d have simulations that we could even discover novel physics behaviors in that match real life! Practically, most are simplified to the point that we can match behavior we already well understand, and maybe a few things can validate we don’t - which can help explore some of the gaps we haven’t been able to touch in the labs. it’s not likely we’ll discover new principles that way though. But I’m guessing I’m just preaching to the choir, eh?
    - toss1 1y ago
      Yup, seems exactly right. And I think we might be nearly to the stage you're identifying, where there is enough fidelity in the models to discover new things. I sort of recall reading at least some of that happening in some fields of math?

motorest 1y ago

> Not to mention, even if your supercomputer has an idea you still need a lab that can confirm the idea. I think this is the critical aspect. You don't need to own a supercomputer to run simulations in supercomputers. Sometimes you don't even need supercomputers at all. It might be a challenge but it's not prohibitive. What's clearly a blocker is the verification & validation work to assess if a specific model actually works. You need to run experimental tests to check if your module is any good, and you need to run experimental tests to check if the output of your computational models correlate with reality. That costs both time and money.

white_tiger 1y ago

Supercomputers are becoming more accessible. There's a large gap between research and real-world use, such numerical models need to be thoroughly tested for robustness before they can be deployed. And according to experience, such studies are usually conducted simultaneously with Laboratory experiments, or already have Lab results they can compare with. If not, such study acts as a new source of synthetic MD data which can be picked up by the experimental folks. I do agree that people who write such blogs tends to sell these ideas or studies as finished products which they are far from. The product of such studies are usually a paper or in other cases a patent. An idea others can build upon, like lego bricks.