The method is trained on the data that were available, but it is meant to be re-trainable as soon as new data are published. It would be great to be really sure that even someone else will be able to do it. In case we receive any feedback, we would be really happy to improve our Github repository so as to make the reproduction easier!
This paper presents a fine example of high-throughput computational materials screening studies, mainly focusing on the carbon nanoclusters of different sizes. In the paper, a set of diverse empirical and machine-learned interatomic potentials, which are commonly used to simulate carbonaceous materials, is benchmarked against the higher-level density functional theory (DFT) data, using a range of diverse structural features as the comparison criteria. Trying to reproduce the data presented here (even if you only consider a subset of the interaction potentials) will help you devise an understanding as to how you could approach a high-throughput structure prediction problem. Even though we concentrate here on isolated/finite nanoclusters, AIRSS (and other similar approaches like USPEX, CALYPSO, GMIN, etc.,) can also be used to predict crystal structures of different class of materials with applications in energy storage, catalysis, hydrogen storage, and so on.
Systematically improvable machine learning potentials could have a significant impact on the range of properties that can be modelled, but the toolchain associated with using them presents a barrier to entry for new users. Attempting to reproduce some of our results will help us improve the accessibility of the approach.
Popular descriptors for machine learning potentials such as the Behler-Parinello atom centred symmetry functions (ACSF) or the Smooth Overlap of Interatomic Potentials (SOAP) are widely used but so far not much attention has been paid to optimising how many descriptor components need to be included to give good results.
Even though the approach in the paper focuses on a specific measurement (clumped isotopes) and how to optimize which and how many standards we use, I hope that the problem is general enough that insight can translate to any kind of measurement that relies on machine calibration. I've committed to writing a literate program (plain text interspersed with code chunks) to explain what is going on and to make the simulations one step at a time. I really hope that this is understandable to future collaborators and scientists in my field, but I have not had any code review internally and I also didn't receive any feedback on it from the reviewers. I would love to see if what in my mind represents "reproducible code" is actually reproducible, and to learn what I can improve for future projects!
I tried hard to make this paper as reproducible as possible, but as techniques and dependencies become more complex, it is hard to make it 100% clear. Any form of feedback is more than welcome.