Lightnews — Scholar-powered news

Reposted by John Gardner

Volker Deringer @vlderinger.bsky.social · Jun 23

Read more about our MLIP distillation preprint in John’s thread! 🙂 #compchem #chemsky

John Gardner @jla-gardner.bsky.social · Jun 23

Excited to share the pre-print we’ve been working on for the last ~4 months:

“Distillation of atomistic foundation models across architectures and chemical domains”

Deep dive thread below! 🤿🧵

3 12

John Gardner @jla-gardner.bsky.social · Jun 24

Thanks! 😊 In principle yes - our data generation protocol requires ~5 model calls to generate a new, chemically reasonable structure + is easily parallelised across processes. If you were willing to burn $$$ one could generate a new dataset very quickly (as is often the case with e.g. RSS set ups)

3

John Gardner @jla-gardner.bsky.social · Jun 23

... @juraskova-ver.bsky.social, Louise Rosset, @fjduarte.bsky.social, Fausto Martelli, Chris Pickard and @vlderinger.bsky.social 🤩

1

John Gardner @jla-gardner.bsky.social · Jun 23

It was super fun collaborating with my co-first-author @dft-dutoit.bsky.social, together with the rest of the team across various research groups: @bm-chiheb.bsky.social, Zoé Faure Beaulieu, Bianca Pasça...

1

John Gardner @jla-gardner.bsky.social · Jun 23

We hope that you can start using this method to do cool new science!

1

John Gardner @jla-gardner.bsky.social · Jun 23

Code for our synthetic generation pipeline (compatible with any ase Calculator object) can be found here:
github.com/jla-gardner/...

GitHub - jla-gardner/augment-atoms: dataset augmentation for atomistic machine learning

dataset augmentation for atomistic machine learning - jla-gardner/augment-atoms

github.com

1

John Gardner @jla-gardner.bsky.social · Jun 23

For (many) more results, please see the pre-print! arxiv.org/abs/2506.10956

Distillation of atomistic foundation models across architectures and chemical domains

Machine-learned interatomic potentials have transformed computational research in the physical sciences. Recent atomistic `foundation' models have changed the field yet again: trained on many differen...

arxiv.org

1

John Gardner @jla-gardner.bsky.social · Jun 23

I find our results for the modelling of MAPI (a hybrid perovskite) particularly pleasing: the distribution of cation orientations generated by the teacher and student models during NVT MD are ~identical!

1

John Gardner @jla-gardner.bsky.social · Jun 23

We go on to apply this distillation approach to target other chemical domains by distilling different foundation models (Orb, MatterSim @msftresearch.bsky.social) and MACE-OFF), and find that it works well across the board!

1

John Gardner @jla-gardner.bsky.social · Jun 23

Beyond error metrics, we extensively validate these models to show they model liquid water well.

1

John Gardner @jla-gardner.bsky.social · Jun 23

These student models have relatively few parameters (c. 40k for PaiNN and TensorNet), and so have much lower memory footprint. This lets you scale single GPU experiments very easily!

1

John Gardner @jla-gardner.bsky.social · Jun 23

The resulting student models reach impressive accuracy vs DFT while being orders of magnitude faster than the teacher!

Note that these student models are of a different architecture to MACE, and in fact ACE is not even NN-based.

1

John Gardner @jla-gardner.bsky.social · Jun 23

We start by (i) fine-tuning MACE-MP-0 (@ilyesbatatia.bsky.social) on 25 water structures labelled with an accurate functional, (ii) using this fine-tuned model and structures to generate a large number (10k) new “synthetic” structures, and (iii) training student models on this dataset.

1

John Gardner @jla-gardner.bsky.social · Jun 23

Does this distillation approach work? In short, yes! 🤩

1

John Gardner @jla-gardner.bsky.social · Jun 23

This approach is very cheap, taking c. 5 calls to the teacher model to generate a new, chemically relevant and uncorrelated structure! We can build large datasets within one hour using this protocol.

1

John Gardner @jla-gardner.bsky.social · Jun 23

In this pre-print, we propose a different solution: starting from a (very) small pool of structures, and repeatedly (i) rattling and (ii) crudely relaxing them using the teacher model and a Robbins-Monro procedure.

1

John Gardner @jla-gardner.bsky.social · Jun 23

This works well, but has two drawbacks: (1) MD is still quite expensive, and requires many steps to generate uncorrelated structures, and (2) expert knowledge and lots of fiddling is required to get the MD settings right.

1

John Gardner @jla-gardner.bsky.social · Jun 23

In previous work, we and others (PFD-kit) have proposed using teacher models to generate "synthetic data" by using them to drive MD, and to sample snapshots along these trajectories as training points.

1

John Gardner @jla-gardner.bsky.social · Jun 23

The devil is always in the details however 😈 The main problem we need to solve is how to generate many relevant structures that densely sample the chemical domain we are interested in targeting.

1

John Gardner @jla-gardner.bsky.social · Jun 23

At a high level, this builds upon the approach pioneered by Joe Morrow, now extended to the distillation of impressively capable foundation models, and to a range of downstream architectures and chemical domains,

1

John Gardner @jla-gardner.bsky.social · Jun 23

Concretely, we train a student to predict the energy and force labels generated by the teacher on a large dataset of structures: this requires no alterations to existing training pipelines, and so is completely agnostic to the architecture of both the teacher and student 😎

1

John Gardner @jla-gardner.bsky.social · Jun 23

Both of the above methods try to maximise the amount of information extracted per training structure from the teacher. Our approach is orthogonal to this: we try to maximise the number of structures (that are both sensible and useful) we use to transfer knowledge.

1

John Gardner @jla-gardner.bsky.social · Jun 23

Somewhat similarly,
@ask1729.bsky.social
and others extract additional Hessian information from the teacher. Again, this works well providing you have a training framework that lets you train student models on this data.

1

John Gardner @jla-gardner.bsky.social · Jun 23

@gasteigerjo.bsky.social
and others attempt to align not only the predictions, but also the internal representations of the teacher and the student. This approach works well for models with similar architectures, but is incompatible with e.g. fast linear models like ACE.

1

John Gardner @jla-gardner.bsky.social · Jun 23

At their heart, all model distillation strategies attempt to extract as much information as possible from the teacher model, in a format that is useful for the student.

Various existing methods in the literature do this in different ways.

1