COSMO Lab

Extremely sad news, I really hope IPAM somehow manages to carry on. If this is not seen as a successful program then I don't see what will be.

Logo for the metatrain project, featuring a locomotive coming out from behind a blocky M letter

Thanks to the 🧑‍🚀🧑‍🚀🧑‍🚀 who put this together, Sofiia in particular, and thanks to the #metatrain team as this would not be so easy without their work metatensor.github.io/metatrain/la...

Training curves for "conservative fine tuning" a PET model

Two new recipes landed in the #atomistic-cookbook 🧑‍🍳📖. One explaining how to fine-tune the #PET-MAD universal model on a system-specific dataset, one training a model with conservative fine tuning. Check them out on atomistic-cookbook.org/examples/pet... and atomistic-cookbook.org/examples/pet...

Periodic boundary conditions for bosonic path integral molecular dynamics We develop an algorithm for bosonic path integral molecular dynamics (PIMD) simulations with periodic boundary conditions (PBC) that scales quadratically with t

Very few things make me prouder than having Jacob and @yotamfe.bsky.social's paper published today in JCP in the special issue honoring one of my personal heroes, Abraham Nitzan.

Abe is a giant of theoretical chemistry (in Israel and globally) and an inspiration to us all.
doi.org/10.1063/5.02...

So give it a try and let us know if it works for you or if there are 🐛 to fix. You can build plumed-metatomic from source, or use our conda build, that is also used in the recipe. Have fun!

You can also use anything you can call from torch, including other tools from the #metatensor ecosystem, such as metatensor.github.io/featomic, and of course you can then use this with whatever code uses PLUMED, from LAMMPS to ipi-code.org.

code snippet for a PLUMED input reading ``` histo: METATOMIC ... MODEL=histo-cv.pt EXTENSIONS_DIRECTORY=./extensions/ SPECIES1=1-38 SPECIES_TO_TYPES=18 ... cv1: SELECT_COMPONENTS ARG=histo COMPONENTS=1 cv2: SELECT_COMPONENTS ARG=histo COMPONENTS=2 ```

Basically, you just need to define a `torch.nn.module` with a specific API, and then you can define anything you like as a CV calculator. Export as .pt torchscript model, and it's just on METATOMIC action away from reading it in #plumed

snapshot of a trajectory of a Ar38 cluster undergoing a transition from the fcc to the icosahedral minima

New 🧑‍🍳📖 #recipe landed, doubling up as a @plumed.org tutorial 🐦 atomistic-cookbook.org/examples/met..., and explaining how to use the #metatomic interface in #plumed to define custom collective variables with all the flexibility and speed of torch.

Kudos to 🧑‍🚀 Arslan, Sofiia & @luthaf.bsky.social, Giovanni and Marnik from the #MC3D team, and Sandip from BASF with who'll be pleased to see that sketch-maps of atomistic datasets are back in action! #compchem #machinelearning

As a nice side-effect, we distribute (well, PR still underway 😆) a featurizer based on PET-MAD latent features that you can use together with `chemiscope.explore` to as a universal materials cartography tool - it even works out of the box to follow the melting of Al!

In short, we suggest that rather than building YUUUGE datasets we need to look to maximize diversity and make sure the reference energetics is 100% internally consistent. You can see in the PET-MAD paper how this pans out when used for training (TL;DR: it works really well) arxiv.org/html/2503.14...

Diverse data is good data! Took a while to polish it, but we have finally released the small-but-smart MAD dataset we used to train PET-MAD. You can find more on the #preprint arxiv.org/html/2506.19... or just head to the #materialscloud to fetch MAD archive.materialscloud.org/records/xdsb...

So kudos to 🧑‍🚀 Filippo and @marceldotsci.bsky.social bsky.app/profile/labc..., thanks to funders @erc.europa.eu @snf-fns.ch @cscsch.bsky.social and @nccr-marvel.bsky.social, and go to Filippo's talk if you are at #icml25, to see this nice mix of #compchem and #machinelearning!

Demonstration of multiple time stepping with non-conservative forces

And you can also use with multiple time stepping #moleculardynamics, recovering most of the speed-up of direct forces, and avoiding the sampling artefacts. Pretty sweet deal, and easy to realize with ipi-code.org, see the 🧑‍🍳📖 atomistic-cookbook.org/examples/pet...

A plot showing the speedup achieved by conservative fine tuning

Best still, you can use direct forces to pre-train a ML potential, and then use "conservative fine tuning" with back-propagation - saving time and achieving similar (or better 😆) accuracy

More importantly, we elaborate better how direct forces *can still be used*. First, we re-use an idea from FlashMD arxiv.org/html/2505.19... to enforce energy conservation. Not a fix, but a mitigation strategy.

A scheme explaining why the extended receptive radius afforded for the derivatives of an interatomic potential don't happen when doing direct force prediction.

We have however added quite a bit - from direct stresses, to a discussion of why direct forces are shorter-ranged than those you get from differentiating a potential with the same receptive radius.

A plot of energy conservation for conservative models (that work) and non-conservative models with direct forces (that blow up)

The conclusions haven't changed: using direct force predictions in atomistic simulations is a bad idea - they don't conserve energy, lead to unstable geometry optimization, and break energy equipartition when used with thermostats. Clear from the plot: conservative 👌, non-conservative 💥.

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...

EPFL research is back on the European stage Eight EPFL researchers have been selected by the European Research Council (ERC) as part of the 2024 call for proposals for the Advanced Grant competition, including four in the School of Engineering.

🏆 Congratulations to Prof Francesco Stellacci from Materials Science and Engineering at EPFL for receiving an ERC Advanced Grant for his project Engineering Protein Interactions Using Small Molecules

Find out more: actu.epfl.ch/news/epfl-re...

#metatensor day about to start! Join us on zoom if you're not at #EPFL epfl.zoom.us/j/68368776745 @nccr-marvel.bsky.social

Great job from the #SCITAS team deploying an energy-efficient GPU cluster to support #AI research (and all accelerated flavors of scientific computing) at #EPFL www.lenovo.com/us/en/case-s... @epfl-ai-center.bsky.social

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...

🧑‍🚀Kevin shows off his multilingual skills with a public defense that is very much for the general public 😅