Yes, someone did. It's actually an active field right now, from high-throughput simulations to try to limit the search space, to additive manufacturing of thousands of samples, to semi-automated characterisation and some testing (e.g. corrosion; for things like mechanics it's more difficult). The idea is te be more efficient than semi-random.
ok, so quasi-random. Would it be more useful to have such a machine to feed data back to models to better the simulations where we could do more in less time and prune down to more likely candidates?
That’s the active research area GP mentioned. In startup land there are a few large outfits, Lila Sciences, Period Labs, Radical AI are all doing a mix of simulations, AI, and autonomous laboratory infrastructure specifically for materials science. (Lila does a lot of biotech but the have materials researchers too)
Also lots of interest and activity in this space in the national labs and academic research scene
It’s something we are trying to do. Currently, there is no machine that can do that. The problem is that the composition space is huge (5 elements out of a dozen, plus about twenty possibly useful minor additives, the combinatorial explosion sets in very quickly) without even taking into account processes and things like microstructure.
No model is sufficient: predictive physical models like DFT are impractical at the required scale (in term of simulation size and compositions to consider, as well as computational cost), and all the fancy regression machines are terrible when extrapolating. Which is too bad, because again the search space is huge and the regions we actually know and can use to train our models are tiny, which means that apart from proofs of concept, we are always extrapolating. And so we need experimental data in the unknown regions of that space to validate the models. It’s like trying to describe the Earth with only being able to see 1cm squares from random positions.
We are not just sitting waiting for CS people to solve everything with LLMs, these things are genuinely complex ;)
