iiinova — where your digital self keeps evolving

Reposted by Matthias Kellner

marcel ⊙ @marceldotsci.bsky.social · 18/09/2026

✨ 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. 👇

Truncated automatic sparse differentiation (ASD). (a) Toy chain of 7 atoms, model reach K = 2 hops. The Hessian H is recovered from Hessian-vector products (HVPs) with the columns of a seed matrix S, giving the compressed product HS. Since H is sparse, star coloring needs 5 colors, hence 5 HVPs instead of 7. Each entry of H is read from one entry of HS (its hue), directly or via its symmetric partner; entries that are sums are never read (crossed). Truncating to one hop needs only 3 HVPs, but the neglected two-hop couplings (gray) now contaminate entries that are read (dots), causing errors. (b) MOF-177 with PET-XS: atoms within K = 5 hops of one Zn atom in the model's graph, with its cutoff sphere and one-hop edges; rest of the cell in gray. (c) Same atoms colored by the force-constant block norm ‖Φᵢⱼ‖ to the marked atom, decaying with distance. (d) Whole Hessian of MOF-177, one cell per atom pair, graph neighbors adjacent: hop count above the diagonal, ‖Φᵢⱼ‖ below, zeros white.

1116

Matthias Kellner @matthiaskellner.bsky.social · 18/09/2026

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.

044

Reposted by Matthias Kellner

COSMO Lab @labcosmo.bsky.social · 03/03/2026

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.

2115

Reposted by Matthias Kellner

COSMO Lab @labcosmo.bsky.social · 18/02/2026

📢 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...

arxiv.org

How to Train a Shallow Ensemble

032

Matthias Kellner @matthiaskellner.bsky.social · 17/02/2026

No Install. No Setup. Just Chemical Shift Predictions. At shiftml.materialscloud.io we host the latest ShiftML3 in the web and you can predict chemical shifts of organic crystals for free in the web!

shiftml.materialscloud.io

ShiftML-3 predictor

032

Matthias Kellner @matthiaskellner.bsky.social · 30/11/2025

What a cool applet - running a universal MLIP directly from your webbrowser!

000

Reposted by Matthias Kellner

COSMO Lab @labcosmo.bsky.social · 28/11/2025

📢 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...

nature.com

PET-MAD as a lightweight universal interatomic potential for advanced materials modeling - Nature Communications

PET-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...

0156

Reposted by Matthias Kellner

COSMO Lab @labcosmo.bsky.social · 28/08/2025

Anticipating 🧑‍🚀 Wei Bin's talk at #psik2025 (noon@roomA), 📢 a new #preprint using PET and the MAD dataset to train a universal #ml model for the density of states, giving band gaps for solids, clusters, surfaces and molecules with MAE ~200meV. Go to the talk, or check out arxiv.org/html/2508.17...!

error plots for the PET-MAD-DOS model on different datasets

042

Matthias Kellner @matthiaskellner.bsky.social · 25/08/2025

We're introducing ShiftML3, a new ShiftML model for chemical shielding predictions in organic solids. * ShiftML3 predicts full chemical shielding tensors * DFT accuracy for 1H, 13C, and 15N * ASE integration * GPU integration Code: github.com/lab-cosmo/Sh... Install from Pypi: pip install shiftml

github.com

GitHub - lab-cosmo/shiftml: A python package for the prediction of chemical shieldings of organic solids and beyond.

A python package for the prediction of chemical shieldings of organic solids and beyond. - lab-cosmo/shiftml

123

Reposted by Matthias Kellner

COSMO Lab @labcosmo.bsky.social · 22/08/2025

🚨 #machinelearning for #compchem goodies from our 🧑‍🚀 team incoming! After years of work it's time to share. Go check arxiv.org/abs/2508.15704 and/or metatensor.org to learn about #metatensor and #metatomic. What they are, what they do, why you should use them for all of your atomistic ML projects 🔍.

1148

Reposted by Matthias Kellner

COSMO Lab @labcosmo.bsky.social · 17/06/2025

🎉 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...

1187

Reposted by Matthias Kellner

COSMO Lab @labcosmo.bsky.social · 03/04/2025

When you combine #machinelearning and #compchem, you need to start worrying at the QM details within your ML architecture. We use our indirect Hamiltonian framework and pySCFAD to explore the enormous design space arxiv.org/abs/2504.01187

A schematic of the functioning of a ML/QM hybrid framework

2144

Reposted by Matthias Kellner

COSMO Lab @labcosmo.bsky.social · 19/03/2025

📢 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 🧑‍🚀

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

1289

Reposted by Matthias Kellner

COSMO Lab @labcosmo.bsky.social · 28/02/2025

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!

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

1125