Hacker News Reader

Comment by hypercube33 | Hacker News Reader

One would think with all of our crazy AI and supercomputers and quantum computers that a team would give it evolutionary goals of just trying simulations of molecular combinations to reach superconductivity. Sure, it'd be one thing to make it in a computer, and making the materials in the real world is quite another but I'm kind of shocked no one has come forward with something yet. I saw simulations of whole viruses running on a cluster of computers where they test drugs out and how they interact with the virus and simulated human cells so one would think its something with enough effort would be possible?

View on HN

This is done across many disciplines to try and aide in new discovery paths. Typically you’re limited in exactly what you can simulate and often times solution candidates may be found that are impractical, currently impossible, or perhaps actually impossible to produce. Sometimes you can add search constraints to tie simulations together to narrow down such false positive solutions found but not always. Heck in some cases it’s literally cheaper and more accurate to do the bench science no matter how alluring virtualized renditions may be.

Most fields are still left with piles and piles of potential solutions to sort through. They often select candidates that are the cheapest and most practical to approach or they have high suspicion of success and pursue those. At the end of the day though we don’t have full universe simulators at every scale we’d want, we have very specific area simulators within very specific bounds. You have to go out an empirically test these things.

But this is and has already been going on for decades across most disciplines I’ve interacted with, they just weren’t using DNN or LLMs at the time but domains are adopting these as well to leverage where feasible in the search process.

I work with a variety of people interested in leveraging simulation and everyone wants to take the successes they see in LLMs or say RL from AlphaStar or AlphaGo and apply them in their domain. It’s alluring, I get it, the issue is that we often lack enough real understanding in domains and the science isn’t as airtight and people think it is, its too general or narrow, or on some cases we have good suspicion of how to build better more accurate simulations but there’s not enough compute power or energy in the world to make them currently practical, so we need to take some tradeoffs and live with less accurate and detailed simulation which leads to inaccurate representations of reality and ultimately inaccurate solution suggestion candidates.

Supercomputers (the classical kind) typically don't want to run codes that try lots of combinations; they were designed for, excel at, and cost a lot because of, the need to speed up single runs at a time. This is partly due to history, and partly due to definitions of supercomputers, but every time I proposed runs like this (proteins/drugs) to the supercomputer centers, they told me to bug off because my codes "only scaled to 64 processors" (that's 64 servers, mind you; before SMP was common).

We did what you described using idle cycles at Google (search for "Exacycle") and we got great results doing large scale parameter explorations (either randomly sampled, or sampled based on where the previous sims suggested looking next)- although, nobody actually did material simulations like this, we did proteins.

Realistically, almost nobody does this because it's just not cost effective (the search space is too large, the loss functions aren't accurate enough, and it uses TONS of energy), and more importantly, somebody else is just going to find a way to generate 75% of the results with 25% of the energy, and that person will get published faster.

You far overestimate the state of the art, and even our basic understanding of what superconducting is mechanistically. Simulating a single atom, alone in the universe is still a struggle not quite achieved.

Isn't the whole point that it's something we might not predict from what we understand about the materials so far? Why would it be likely that a simulation would do better than theory at predicting unknown experimental results?

You might hit on some interesting interactions between known properties that haven't been investigated but I would assume the real interesting results are from things we just don't know to model, or how to model.

There's plenty of researchers working on material simulation on a molecular level. It's not easy to just search millions of possible combinations and accurately predict their behavior

I believe that the fundamental physical particle interactions are not yet well enough understood to make a precise simulation, even if you had a supercomputer.

Ie. currently it can't be formulated as a search problem entirely on a computer.

Quantum computers are still not useful, fwiw.

Reply on news.ycombinator.com