COSMO Lab

nature.comPET-MAD as a lightweight universal interatomic potential for advanced materials modeling - Nature CommunicationsPET-MAD is a fast and lightweight universal machine-learning potential, trained on a small but diverse dataset, that delivers near-quantum accuracy in atomistic simulations for both organic and inorga...

28 Nov 2025

📢 PET-MAD is here! 📢 It has been for a while for those who read the #arXiv, but now you get it preciously 💸 typeset by @natcomms.nature.com Take home: unconstrained architecture + good train set choices give you fast, accurate and stable universal MLIP that just works™️ www.nature.com/articles/s41...

COSMO Lab reposted

marcel ⊙

@marceldotsci.bsky.social

18 Sept

✨ new work: "Truncated automatic sparse differentiation for MLIPs" with @adrianhill.de & @micheleceriotti.bsky.social has arrived on the arXiv. In a 🥥: we exploit that MLIP Hessians are sparse and decay with distance with a "truncated" variant of automatic sparse differentiation. See figure. 👇

COSMO Lab reposted

Matthias Kellner

@matthiaskellner.bsky.social

18 Sept

We have released ShiftML4! ShiftML4 comes with built-in molecular corrections, elevating the chemical shielding predictions to hybrid-DFT level quality. Try it yourself on shiftml.org, or from the PyPi package.

30 July

A little holiday reading from 🧑‍🚀: arxiv.org/html/2607.26... , or how Michelangelo used #AI to break equivariant atom-centered descriptors all the way to 7 neighbors (and arbitrary many neighbors when considering a finite, practical level of discretization). 🤯

2 July

What will O2, benzene and ozone do on top of a Mendeleev cluster 🤔 ? Thanks to the improved mlip.js webtool from Peter Spackman you can figure it out running #PET-MAD-XS in your browser using webgpu 🚀 . Try it out at @crystalexplorer.net 's amazing MLIP.js peterspackman.github.io/mlip.cpp/

10 June

Time to learn the action 🏃 ! Today I'm sharing a paper, just published on @physrevlett.bsky.social, showing how to construct a symplectic #machinelearning predictor of classical mechanics by learning the Hamilton-Jacobi action.

10 June

TL;DR - this allows running molecular dynamics with much longer time steps, ensuring long-time stability of the trajectories. The method is a true beauty, extremely elegant, although not super-practical - we see it more as a correction to make sure that flashmd.org ⚡ trajectories are working well.

flashmd.orgflashmd.org

journals.aps.orgjournals.aps.org

10 June

You can read more about this little gem from 🧑‍🚀 @filippobigi.bsky.social and Johannes Spies here journals.aps.org/prl/abstract..., and a popular summary on phys.org, physics.aps.org/articles/v19....

9 June

Afternoon action at the @nccr-marvel.bsky.social / @ictp.bsky.social summer school, with the 🧑‍🚀 team introducing MD with #machinelearning potentials using PET, metatomic and the atomistic-cookbook.org 🧑‍🍳📖

chemiscope.orgChemiscopeInteractive data visualization for materials and molecular databases. Correlate atomic structures and their properties, either online, in a jupyter notebook, or with a portable app.

2 June

📢 Chemiscope 1.0 paper is out on #JOSS. If you haven't used ⚗️ 🔭 recently, head to joss.theoj.org/papers/10.21... to read a summary of the new features, to chemiscope.org to try it out, and type `pip install chemiscope` to use it locally.

2 June

On the web, in your jupyter notebooks, in online documentations and in streamlit applications, every time you have to look at atomic-scale data, chemiscope is there to help.

2 June

Kudos to Sofiia Chorna who has taken over as lead developer, @luthaf.bsky.social, @cersonsky.bsky.social , @jakublala.bsky.social and Qianjun Xu to help turn a weekend project into a super-useful tool.

So let us show you just how *universal* #PET-MAD-1.5 can be. This is a movie of a parallel tempering simulation, with replicas from 300K to 3000K, of what we call a "Mendeleev cluster" - one atom each of every element from 1 to 102.

The simulation is stable for a combined 1.6 ns, and the final structures make a lot of physical sense: noble gases, and at higher temperature more volatile stuff, leave the particle. Accuracy is quantitative: the MAE error on forces is ~150 meV/Å - pretty exceptional for something so outlandish!

atomistic-cookbook.orgMendeleev’s nano-clusters - The Atomistic Cookbook

If this sounds too good to be true, don't believe me, see for yourself: there is already a 🧑‍🍳 📖 recipe waiting for you try atomistic-cookbook.org/examples/men...

📢 We have been working on a new universal atomistic dataset that combines the principles of MAD with a meta-GGA level of theory, so we can all simulate water that does not freeze at 500K 🧊 , and have all our bases covered, with reference data for every isotope with a half-life above 24 hours ☢️

View full thread

This is the result of a massive effort by @cesaremalosso.bsky.social, @filippobigi.bsky.social, @ppegolo.bsky.social, @jwasci.bsky.social, @mahrossi.bsky.social and Arslan Mazitov, as well as all the 🧑‍🚀 working tirelessly on the conceptual and software infrastructure.

arxiv.orgHigh-quality, high-information datasets for universal atomistic machine learning

Looking forward to see what you do with the dataset, MAD-1.5, and the universal potential, PET-MAD-1.5! 🔗: 📄 arxiv.org/html/2603.02...; 🚀 github.com/lab-cosmo/up...

github.comGitHub - lab-cosmo/upet: Universal interatomic potentials for advanced materials modelingUniversal interatomic potentials for advanced materials modeling - lab-cosmo/upet

Now, this is just a milestone on that path, but it's already something worth sharing, so thanks to arXiv you can read about it arxiv.org/html/2603.02..., and thanks to uPET github.com/lab-cosmo/up... and metatomic, you can try already a universal MLIP trained on it.

23 Feb

New recipe just landed on the atomistic-cookbook.org 🧑‍🍳📖. Thanks to @yairlitman.bsky.social for explaining how to use ipi-code.org to perform ring-polymer instanton calculations of reaction rates that include quantum tunneling effects ⚛️⚡. Check it out 👉 atomistic-cookbook.org/examples/rin...

arxiv.orgHow to Train a Shallow Ensemble

18 Feb

📢 New #preprint is out! Investigating the many flavors of last-layer #UQ, Moritz and 🧑‍🚀Matthias propose a practitioners' guide on "how to train a shallow ensemble". TL;DR? for good calibration use NLL, include force, and optimize the backbone, fine-tuning for speed! 📃🔗➡️ arxiv.org/html/2602.15...

Many #machinelearning potentials are built (or understood) in terms of "atomic cluster expansions" that link directly to a body-ordered energy decomposition that can be computed explicitly with a sequence of electronic structure calculations. But what kind of expansion do they learn in practice? A🧵

View full thread

Our conclusion is that the body ordered decomposition is not a very useful prior for conditioning MLIPs. An architecture such as MACE that forces a fast-converging effective decomposition learns well in the low-data regime ...

pubs.aip.orgResolving the body-order paradox of machine learning interatomic potentialsIn many cases, the predictions of machine learning interatomic potentials (MLIPs) can be interpreted as a sum of body-ordered contributions, which is explicit w

... but it is too inflexible to learn highly-correlated fragment energies, and does not extrapolate well to different densities. If you are curious to learn more, read the whole story here pubs.aip.org/aip/jcp/arti...

View full thread

Third, only architectures with sufficient expressive power, such as MACE or PET, can learn the true expansion, when exposed to fragments in the training set.

Last but not least, of the architectures we tried, only PET 😸 could actually do something good out of learning these fragments, while MACE sees its extrapolative power degrade significantly when forced to learn the slowly-converging vacuum cluster expansion.

A 🧑‍🚀 dream team led by Sanggyu Chong, with some FCI help from Joonho Lee @harvard.edu, tries to answer this question, and to understand what it implies for the accuracy and transferability of MLIPs in this new 📄 fresh off the press at @aip.bsky.social #JCP. pubs.aip.org/aip/jcp/arti...

This is a story with many layers. First, the quantum cluster expansion does not converge, mostly because the fragments have crazy geometries and highly-correlated electronic structure. Second, MLIPs exposed only to compact structures learn their own effective body-ordered decomposition.

pubs.aip.orgmetatensor and metatomic: Foundational libraries for interoperable atomistic machine learningIncorporation of machine learning (ML) techniques into atomic-scale modeling has proven to be an extremely effective strategy to improve the accuracy and reduce

13 Feb

Hot off the press on hashtag @aip.bsky.social #JCP, an introduction to the metatensor ecosystem. High-quality 🧑‍🚀 tools for atomistic hashtag#machinelearning - read on pubs.aip.org/aip/jcp/arti... and check it out at metatensor.org 🧑‍🍳 📖 recipes here atomistic-cookbook.org/software/met...

12 Feb

It is a stormy day, but the COSMO retreat is going strong!

ai.epfl.chA New Reference Model for Machine-Learning–Driven Materials Discovery - EPFL AI CenterResearchers at EPFL’s Laboratory of Computational Science and Modeling (COSMO) have reached a significant milestone in material science, reaching the top position on Matbench Discovery, the leading be...

9 Feb

🧑‍🚀 Filippo, Arslan and Paolo doing some PET talk with our friends at the @epfl-ai-center.bsky.social ai.epfl.ch/a-new-refere...

8 Feb

If you want to learn about materials modeling, from DFT to MD, well marinated in a spicy ML sauce, don't miss out the @ictp.bsky.social / @nccr-marvel.bsky.social college. Details and application instructions here indico.ictp.it/event/11146/. See you in Miramare!

dtu.dkInternational AI competition aims to speed up the development of materials for the green transitionThe Pioneer Center CAPeX at DTU has announced the winners of the first phase (Stage 1) an international competition in partnership with the Novo Nordisk Foundation, and the Danish Centre for AI Innovation (DCAI), on using machine learning models to predict synthesis recipes for novel nanoparticles.

6 Feb

PET continues its victory round of benchmarks and challenges 🥇🥉. And this one has a (bit far-fetched) end goal that would also make it useful! Congrats to Filippo and Cesare (and @marceldotsci.bsky.social who got a honorable mention and will also try further his LOREM model)🚀 dtu.dk/english/news...

3 Feb

Congratulations to 🧑‍🚀 Sergey Pozdnyakov who very deservedly won the @materials-epfl.bsky.social doctoral distinction award. A good time to go check on his papers, if you haven't read them already!

1 Feb

Release candidate 3 of chemiscope 1.0 is out, with class and range based highlighting of points. Try it, break it, report it on github.com/lab-cosmo/ch...

actu.epfl.chA new national research programme recognizes EPFL's expertiseThe Swiss Confederation launches six new National Centres of Competence in Research (NCCRs). The NCCR “Separations”, which aims to accelerate research in separation sciences - the quest for chemical a...

30 Jan

Fantastic news from the @snf-fns.ch, who despite the budget cuts managed to fund six new NCCRs. Looking forward to doing some cool simulations to advance separation science! actu.epfl.ch/news/a-new-n...

23 Jan

If you got curious by the PET-OAM results a week ago, you can learn more reading up arxiv.org/abs/2601.16195. Including some general considerations on how to train and use safely an unconstrained ML potential.

23 Jan

If you're scared by the 700M parameters (you shouldn't be) there's a whole set of models from 🐁 to 🦣. You can find them all on github.com/lab-cosmo/upet !

14 Jan

Not going to make a big deal out of a benchmark table, but PET just got the top spot on matbench-discovery.materialsproject.org. And don't be fooled by the huge parameters count, it's faster and can handle larger structures than eSEN-30M 🚀. Kudos to 🧑‍🚀 Filippo, Arslan and Paolo!

github.comGitHub - lab-cosmo/upet: Universal interatomic potentials for advanced materials modelingUniversal interatomic potentials for advanced materials modeling - lab-cosmo/upet

14 Jan

You can fetch the model here github.com/lab-cosmo/upet, as easy as `pip install upet`, and then, for the ASE interface, `from upet.calculator import UPETCalculator; calculator = UPETCalculator(model="pet-oam-xl", version="1.0.0", device="cuda")` Have fun and go break it!

5 Jan

📢 chemiscope.org 1.0.0rc1 just dropped on pypi! We are making (a few) breaking changes to the interfaces, fixing a ton of bugs and introducing some exciting features (you can finally load datasets with > 100k points!). We'd be grateful if you test, break and report 🐛 github.com/lab-cosmo/ch...

3 Jan

Hope y'all are getting a great start of 2026. Here we're taking some time to add the 2025 winter card to the archives www.epfl.ch/labs/cosmo/i... 🎅=🧑‍🚀

17 Dec 2025

📢 New chemiscope.org release just landed! To make it even easier to integrate ⚗️🔭 into your workflow, we added a @streamlit.bsky.social component, so you can run analyses and show you atomistic data in a web app by just writing a few lines of python! try it, break it, report it!

10 Dec 2025

Congrats to 🧑‍🚀 Sergey Pozdnyakov who received a distinction (best 8% of theses at @materials-epfl.bsky.social) for his PhD thesis "Advancing understanding and practical performance of machine learning interatomic potentials". Поїхали 🚀! infoscience.epfl.ch/entities/pub...

9 Dec 2025

No day goes by without a new universal #ML potential. But how different they really are? Sanggyu and Sofiia tried to give a quantitative answer by comparing the reconstruction errors between their latent-space features. If you are curious, check out the #preprint arxiv.org/html/2512.05...

nature.comMassive Atomic Diversity: a compact universal dataset for atomistic machine learning - Scientific DataScientific Data - Massive Atomic Diversity: a compact universal dataset for atomistic machine learning

23 Nov 2025

📢 Let us (re)introduce to you our Massive Atomic Diversity dataset for universal MLIPs. MAD includes molecules, clusters, surfaces and plenty of bulk configs, we cover a lot of ground with fewer than 100k structures, using highly consistent DFT settings. Read more 📑 www.nature.com/articles/s41...

archive.materialscloud.orgMassive Atomic Diversity: a compact universal dataset for atomistic machine learningThe development of machine-learning models for atomic-scale simulations has benefitted tremendously from the large databases of materials and molecular properties computed in the past two decades usin...

23 Nov 2025

Developed in collaboration with the THEOS group with support from the @nccr-marvel.bsky.social, and stored on the #materialscloud here archive.materialscloud.org/records/c4en..., you can also take a look at the data as a #chemiscope 🔭⚛️ chemiscope.materialscloud.io?load=https%3...

COSMO Lab reposted

AIhub.org

@aihub.org

15 Oct 2025

‪In this blog post, Filippo Bigi, Marcel Langer (@labcosmo.bsky.social‬) and @micheleceriotti.bsky.social write about the need to balance speed and physical laws when using ML for atomic-scale simulations aihub.org/2025/10/10/m...

aihub.orgMachine learning for atomic-scale simulations: balancing speed and physical laws - ΑΙhub