Hacker News Reader

Comment by gxt | Hacker News Reader

Since I presume steel is a more or less uniform substance compared to biological processes, why is metallurgical research still more or less experiment first?

Shouldn't it possible nowadays to bruteforce a search for an alloy of any given properties using computer simulations of the atomic or molecular structures?

View on HN

It's a good question; unusually, that doesn't mean I don't have a good answer.

Steel is so far from being a more or less uniform substance that it's not even funny. There are four major phases that play roles even in the commonest carbon steel (ferrite, cementite, austenite, and martensite), plus others that can form at times like graphite, which plays an important role in cast irons. Ferrite and cementite can form nanolaminated microstructures called pearlite and bainite which have a major influence on the properties of the steel, and there are other microstructures that form depending on cooling speed, heat treatment, and cold working. So even the simplest steel is a nanostructured composite of metal and ceramic whose properties are hard to model computationally, though great strides have been made in recent decades.

Then, once you add other alloying elements besides those two (intentionally or not), steel stops being so simple. You can find phase diagrams for most of the binary systems (vanadium-carbon, for example, or vanadium-iron) but most of the ternary systems probably include compounds that haven't been identified yet. In theory you could find them computationally, I think. Even when you have a phase diagram, though, that doesn't tell you how fast the phase transitions happen, which depends on things like the crystal structures of intermediate unstable phases.

I don't know anything about this stuff, I just read about it. Recommended! Start with https://www.tf.uni-kiel.de/matwis/amat/generalinfo_en/guided...

Naive question: Has anyone tried to whack these questions with the machine learning hammer? I figure if we can do protein folding [0], we should be able to do knives.

[0] https://deepmind.com/blog/article/alphafold-a-solution-to-a-...

Folks are certainly working on it from many dimensions, but it’s a pretty hard problem since getting ground truth data involves making and testing materials, which is a very different problem to automate than training machine learning models. You need a fairly cross disciplinary team to make progress. As an example of folks doing good work: https://a3md.utoronto.ca/

I don’t believe we have a good way to compute various macro properties of something like steel. We can compute density and what not, but how much it holds its sharpness or something is something I haven’t seen.

So I am not sure how to get the training data needed for ML.

(Computational chemist, but not computational materials scientist. So could be wrong!)

Kind of, deepmind actually recently published results in metal oxides:

https://ai.googleblog.com/2021/10/finding-complex-metal-oxid...

Steel has similar complexity since the number of combinations is so vast.

That's still only, what, twenty or thirty dimensions? I guess it's hard enough to gather data that it's not as simple as feeding it a big black-box optimizer, something like SageMaker or Vizier that's designed to tune ML models with week-long training times and dozens of hyperparameters, but that'd still be quite a bit more powerful than the manual search that the author talks about.

Questions about prime numbers are asked in one dimension.

The manual search is guided by a lot of very rigorous theory-of-experiment. It's not just trial and error, it's quite a bit more.

And do what? ML requires a way to know if it’s making correct predictions.

This is actually a great question and doesn't deserve to be downvoted. Indeed this is one of the considerations that led me to leave materials science research field after a couple of undergraduate projects with PhD candidates.

It turns out that the domain between Angstroms (where we can computationally model atomic interactions accounting for quantum effects) and Milli (where standard Newton's laws and therefore mechanical engineering tools can be used) is a vast computational desert.

Most properties that affect bulk material properties happen to be developed in the micro-domain (note the photographs in the article) and almost 20 years after I've left the field, I don't believe there's still any rigorous "first-principles" based computational approach yet. In other words, materials are not uniform in the micro domain and this is where materials properties develop.

So materials research process becomes hypothize, create material batch, test it 20 ways, rinse and repeat for a slightly different composition or process

Even the software mentioned in the article (thermo-calc) is primarily empirical with some very smart extrapolations and modeling added (note the first step is experimental data capture [1]. It definitely is a massive step forward from when I was in the field but definitely not first principles based modeling.

[1] https://thermocalc.com/about-us/methodology/the-calphad-meth...

appreciate your use of "first principles based modeling". in my program, that's what we meant by "model based", but usage in the AI community is quite different

your verbiage concisely captures what's so important about the concept

Former computational materials scientist here. There are groups that are using ML for finding materials to simulate, and some beginning to use it to speed up simulations. Still that said, simulating cutting, abrasion, sharpening (I suppose it would be called to some extent tribology) is still in the infancy of simulation. Steel is extra difficult compared to other materials, and it has such a history of innovation that all exists at a sort of mesoscale out of reach of contemporary atomistic simulations. Still some have attempted it: https://www.dierk-raabe.com/icme/ or more recently: https://www.sciencedirect.com/science/article/abs/pii/S09270... Still the from Simulation/Search -> Experiment pipeline is working generally at much smaller scales that steel structures for now. ie micro instead of nano

The search space is much, much more vast than you'd imagine, and there are so many ways that things get non-linear that we have absolutely no idea how (way) more than 99.9999% of the possible alloys that we could make would actually perform in reality. The way most of the alloys we are using were found was that we started with something that we already knew, and then tweaked from there to optimize some property.

Per TFA, it is to an extent? The author briefly discusses a computational search of the design space, and uses that data to encourage the partner company to make a batch of the steel.

That said, simulating material properties from atomic scale principles seems nontrivial compared to predicting them given observed parameters and properties of other alloys. I’d be interested in more informed comments on that possibility!

The important qualities of steel emerge out of microstructure: nanometer to millimeter sized features (several different crystal structures in the same material interacting through the boundaries between them and the bulk properties in them) which requires quantum interactions to be tracked through many orders of magnitude of scale. This includes how these different structures are formed though many stages of melting, tempering, work hardening, etc. In other words it is hideously computationally complex.

One of many fields where yes there is a lot of simulation and yes it is developing but still quite far from having anything close to a complete model which can escape the need for extensive experimentation.

There is a sort of prevalent idea among people outside these fields that simulations exist which can just handle anything. This is very wrong and quite far away.

AFAICT that is, to a large degree, what the author did. Much of the initial "exploration" seems to have been done in Thermo-Calc, with physical experiments following. IMO the novelty in process is more exciting than the novelty of the result.

There's a startup out of U of Toronto working on exactly this:

https://www.thephaseshift.com/

No. This is knives, not science.

Reply on news.ycombinator.com