Guillaume Fraux

Very proud to send Filippo Bigi to Vancouver to give an oral presentation at @icmlconf.bsky.social about our investigation of the use of "dark-side forces" in atomistic simulations. The final version is here openreview.net/forum?id=OEl... and it's worth a read even if you already read the #preprint

🎉 DFT-accurate, with built-in uncertainty quantification, providing chemical shielding anisotropy - ShiftML3.0 has it all! Building on a successful @nccr-marvel.bsky.social-funded collaboration with LRM🧲⚛️, it just landed on the arXiv arxiv.org/html/2506.13... and on pypi pypi.org/project/shif...

The link above is missing a password, pleas use epfl.zoom.us/j/6836877674... instead!

We wouldn't be @labcosmo.bsky.social if we didn't want you to go break it, so head to atomistic-cookbook.org/examples/fla... for a crash-course 🧑‍🍳📖 recipe, but not before reading the warnings arxiv.org/html/2505.19.... Have fun! #compchem #machinelearning #md @nccr-marvel.bsky.social @erc.europa.eu

If you do, the rewards can be very impressive: you can run solvated alanine dipeptide and observe superionic behavior in LiPS with 16fs time step, and watch the Al(110) surface pre-melt in strides of 64fs. And all with the same universal model, no fine-tuning needed!

Darth Vader dressed as Santa presents a non-conservative forcefield

Much as for direct force prediction [ arxiv.org/html/2412.11... ] you better know what you are doing: you've no guarantee of energy conservation, or of equipartition, so you should know your thermostats VERY well. Caveat emptor.

Filippo's idea was to use the heavy-duty PET-MAD model [ arxiv.org/html/2503.14... ] to generate a bunch of trajectories of wildly different compounds and use a PET-like architecture to learn (q',p') from (q,p), and and to think A LOT about the many things that could possibly go wrong.

Learning Large-Time-Step Molecular Dynamics with Graph Neural Networks Molecular dynamics (MD) simulation predicts the trajectory of atoms by solving Newton's equation of motion with a numeric integrator. Due to physical constraints, the time step of the integrator need ...

There are reports as early as 2021 [cf. arxiv.org/abs/2111.15176 ] of using neural nets to predict a MD trajectory in large strides, but these were usually limited to a single system in a given thermodynamic state point. Nice, but not life-changing.

Scheme of the GNN architecture of the FlashMD method.

📢 Running molecular dynamics with time steps up to 64fs for any atomistic system, from Al(110) to Ala2? Thanks to 🧑‍🚀 Filippo Bigi and Sanggyu Chong, with some help from Agustinus Kristiadis, this is not as crazy as it sounds. Let us briefly introduce FlashMD⚡ arxiv.org/html/2505.19...

If you don't want to read, but want to cook, guess what? The 🧑‍🍳📖 #atomistic-cookbook has you covered. Head to atomistic-cookbook.org/examples/lea... to see how to use this, as simple as `mtt train mymodel.yaml`.

If you don't have time for the 20-pages appendix, the TL;DR is that approximating tensors is harder than the vector case, but can be made as simple as possible using angular momentum theory. The practical implementation we propose is not as rigorous, but works well and is very fast in practice.

Schematic representation of the lambda-MCoV architecture for predicting tensorial quantities using scalar approximators

First #preprint from recent 🧑‍🚀 Michelangelo Domina (+ @ppegolo.bsky.social and Filippo) provides theoretical foundations and practical architectures to build scalar-function-based approximations of tensors, in the spirit of the "scalars are universals" paper #compchem arxiv.org/html/2505.05...

You can use these safely in MD, by multiple time stepping (a '90s classic: doi.org/10.1063/1.46...) so you get reliable conservative trajectories, for any material, twice as fast! Let us know if it works for you, and even more importantly, if it doesn't! 👉 atomistic-cookbook.org/examples/pet...

Plot of the potential and total energy trends along a MD trajectory, showing drift for non-conservative forces, and how that is fixed using multiple time stepping.

The PET-MAD universal forcefield mingled with the dark side, and got twice as fast 🚀. Read on, or head to the 🧑‍🍳📖 atomistic-cookbook.org/examples/pet..., if you are curious of what this is all about. #atomistic-cookbook #compchem #machinelearning #mlip🧵

Fast and flexible long-range models for atomistic machine learning Most atomistic machine learning (ML) models rely on a locality ansatz and decompose the energy into a sum of short-ranged, atom-centered contributions. This lea

📢 For those who missed the #preprint, torch-pme is now published in @aip.bsky.social #JChemPhys. #compchem classics like Ewald and P3M meet #machine learning. Read all about it here pubs.aip.org/aip/jcp/arti...

Header for the atomistic-cookbook.org recipe for the PET-MAD universal potential

You can read all about it here, arxiv.org/abs/2503.14118, see it in action as a #cookbook recipe 🧑‍🍳 atomistic-cookbook.org/examples/pet... or rush to install it following the instructions here github.com/lab-cosmo/pe...

Polar plot showing the errors of several machine-learning potential of different test sets. Smaller is better here!

Plots showing the evaluation time per atom for several machine-learning potentials as a function of the number of atoms in a simulation. Smaller is better

📢 PET-MAD has just landed! 📢 What if I told you that you can match & improve the accuracy of other "universal" #machinelearning potentials training on fewer than 100k atomic structures? And be *faster* with an unconstrained architecture that is conservative with tiny symmetry breaking? Sounds like 🧑‍🚀

Should be super-easy to adapt to whatever tensor you're trying to learn (and to extend to more sophisticated models than symmetry-adapted regression). Let us know what you cook with it 😋!

In particular, this recipe relies on metatensor.github.io/featomic to compute the features, and the docs.metatensor.org/latest/torch... backend of #metatensor to export a self-contained ASE-compatible calculator. Easy to use, fast, and accurate.

Polarizability of a small molecule, represented as an ellipsoid

🧑‍🍳 A new #cookbook recipe to train, export and use a simple but effective linear model for tensorial properties - specifically molecular polarizabilities
atomistic-cookbook.org/examples/pol...
Thanks @ppegolo.bsky.social for the recipe, and the whole 🧑‍🚀 team for the underlying infrastructure.

Header of the webpage showing the title ("Atomistic Water model for MD") and the authors (Philip Loche, Marcel Langer, Michele Ceriotti)

Happy to share a new #cookbook recipe that shocases several new software developments in the lab, using the good ole' QTIP4P/f water model as an example. atomistic-cookbook.org/examples/wat.... TL;DR - you can now build torch-based interatomic potentials, export them and use them wherever you like!

No, Overleaf, I don't want to use "AI"
No, Outlook, I don't want to use "AI"
No, Slack, I don't want to use "AI"
No, DuckDuckGo, I don't want to use "AI"
No, Samsung, I don't want to use "AI"
No, my *fucking oven*, I don't want to use "AI"

@marceldotsci.bsky.social getting into very dangerous territory.

Interesting to note, packaging was by far harder than the actual implementation. Kudos to 🧑‍🚀 Filippo, @luthaf.bsky.social, and @cscsch.bsky.social Nick Browning for the perseverance!

Maybe the biggest problem with LLMs as expert systems is that they can't say how confident they are. I'd be a lot more OK with LLMs sometimes giving incorrect answers if they didn't deliver those wrong answers as though they were certain.

An isosurface of a solid harmonic, incluing a visualization of its gradient.

Very happy to announce that github.com/lab-cosmo/sp... has reached the 1.0 release milestone. Most notably, you can now `pip install sphericart[torch,jax]` and get it installed without compilation, complete with CUDA support where available. Spherical (& solid) harmonics made fast and easy!

Many thanks to everyone who helped with this development effort, in particular Philip Loche, Filippo Bigi, Joe Abbott, @micheleceriotti.bsky.social and everyone at @labcosmo.bsky.social, their help was invaluable in getting this project out and useful! 10/n, n=10!

We are also open to adding other representations you use for your research, please let us know what you would like to see! 9/n

Documentation is available here (metatensor.github.io/featomic/), in a rough state but complete. The next step for us will be to improve this documentation, and then bring GPU acceleration in the mix! 8/n

GitHub - metatensor/featomic: Computing representations for atomistic machine learning Computing representations for atomistic machine learning - metatensor/featomic

As usual, everything is open source (github.com/metatensor/f...), and you can now use `pip install featomic` or `pip install featomic-torch` to install it! 7/n