The Future We Simulate is the One We Create

“the smartest minds can and must create the best code that has ever been written to actually solve the grand challenge problems”

Well - that is Julia then isn’t it?

The Future We Simulate Is The One We Create - The Next Platform

The article presents various important projects (curing cancer, nuclear fusion) as computational problems. But they aren’t purely — you still need the test tubes, the experiments, the engineering.

I am not an expert, but I am not entirely sure that out of an extra $1000 invested in research about these problems, all should be spent on whateverscale computation.


@Tamas_Papp Good point. Looking at the announcement of new battery materials found using AI, that is fascinating. But the “goodness” parameter of what makes a good battery must be basic science.

1 Like

Also in that article the authors call for sub-kilometre weather forecasts. Any experts here care to comment? The physical geography can be used in models at a small scale - but does the sensor (temperature, wind speed etc) data have to be collected at a smaller scale for the forecast to be good?
As an aside, the Swiss weather forecast is at a small scale - lots of mountains!

I suppose it shouldn’t be surprising that an opinion piece in a publication focused on HPC applications would frame HPC as the cornerstone of all future technological development.

I’m super down for optimistic future-oriented reads, but I don’t think this one makes a very strong case for why we should be investing in HPC. Paring down the hyperbole and focusing purely on the assertions for what the author believes HPC could be used to accomplish:

The most concrete points here are the ones about weather forecasting and climate modeling. The points about curing cancer and fusion power are admirable but a bit vague and seem highly-aspirational: it’s not obvious to me that those fields are being significantly held back by access to computational resources. The last point just seems kind of defensive and thoroughly von Braun-ian, though.


The new material is a solid-state electrolyte that uses lithium, sodium, and some other elements, Microsoft said. In comparison with traditional lithium-ion batteries, it uses up to 70% less lithium. The material is still undergoing research, so the scientific significance lies less with the material itself and more with how fast it was discovered

That is as vague as it could be.

There is no such a thing as “the” goodness parameter. There is a lot of experimental science(s) and engineering about a lot of aspects of batteries, published in at least hundreds of thousands of experimental papers and patents (including a couple of mine BTW). After - and only after - you have it, AI, large scale simulations, and whatever computations sometimes can help you with the next steps. After which you will need to go to you lab and synthesize, measure, analyze…

Either you believe HPC simulation and modeling can change the world – can save the world – or you don’t.

So that is likely about some new religion? Or just about money?

We believe. Do you?

No, me not.


If “AI” or brute-forcing can solve anything on this page List of unsolved problems in physics - Wikipedia
I will give up on higher thought and I will willingly become like the characters from the infamous “Idiocracy” film.

I think the machine learning could be useful in helping scientists solve problems by making some mechanical or tedious tasks easier, without actually solving the problems alone. Ie a productivity aid rather than a replacement.

But in any case, I am calling dibs on Beef Supreme.


Of course, in fact it’s already doing so in many ways.
It’s just funnier to be an unfair reductionist to make fun of hyperbolic claims used to obtain funding.


In the meanwhile, a preprint was published, and I found it interesting to compare it with the “high hype” random article I cited above, and also with a “medium hype” actual press release from MS.

I worked in the field many years ago, so it was also highly interesting for me to check the progress of the theoretical methods.

Here a short summary after a browsing through the paper, most probably an inexact one.

They took all feasible chemical elements (54 of them), all crystal structures from a database, and generated some 32 Mio structures by iso-valent substitutions. In the following ML decided which phases could be thermodynamically stable. That followed by a few more ML-based selection steps, with physics-based and semi-physics-based approaches at the later steps, which narrowed the number of compositions down to a two dozen of compounds.

Of these they have tried to synthesize Li3YCl6, Li5YCl8, and Li7Y2Cl13 compositions, which however apparently produced the Li3YCl6 phase only, which is a well-studied solid-state electrolyte.

Furthermore they synthesized Na2LiYCl6, having the predicted structure, with the conductivity which is lower by 3 to 4 orders of magnitude as compared to Li3YCl6, and thus IMO without any practical value. Furthermore it’s a mixed Li/Na conductor, which is interesting scientifically, but would be a problem in battery application.

On the bottom line, they didn’t find any new battery materials, at least not yet. However it is fascinating how you can now “mix chemical elements” virtually, and get reasonable predictions about the properties of the resulting crystals.

P.S. Especially if the resulting substance has been a part of the training dataset.


I wonder what would happen if you benchmarked the trained AI comparing to a veteran chemist who worked with these materials for the last 20 years.