Matthias Kellner (@LinusKellner) header
Matthias Kellner (@LinusKellner) profile picture

PhD student in @lab_COSMO Website: https://teapawt.org/t.co/8qpggZ5rlE

Switzerland
Joined February 2024
Photo by Matthias Kellner (@LinusKellner)
Matthias Kellner retweeted
We are alarmed by reports that Germany is on the verge of a catastrophic about-face, reversing its longstanding and principled opposition to the EU’s Chat Control proposal which, if passed, could spell the end of the right to privacy in Europe. signal.org/blog/pdfs/germany…
686
8,354
1,175
29,204
4,780,305
Matthias Kellner retweeted
✨preprint alert: "learning long-range representations w/ equivariant messages" in which we get into the fray of long-range MLIPs and propose to use equivariant charges w/ classical electrostatics as long-range MP mechanism. we design LOREM, an equivariant MLIP, around this...
2
8
1
37
3,647
AI + physics turns dynamics into a data engine. FeNNix-Bio1: a foundation model that drives QM-level, reactive MD on a single GPU, hydration, folding, binding ΔG within ≈1 kcal/mol, yet pushes 7M-atom boxes at force-field speed
3
19
82
6,232
Matthias Kellner retweeted
#compchem What could we do if we had a fast and accurate foundation #MachineLearning learning model for condensed phase molecular dynamics simulation of biological systems? 👉Check the @ChemRxiv preprint: "A Foundation Model for Accurate Atomistic Simulations in Drug Design": doi.org/10.26434/chemrxiv-20… Amazing work by @Thomas__Ple and the teams at @LCT_UMR7616 and @qubit_pharma introducing the FeNNix-Bio1 foundation model. Trained exclusively on synthetic quantum chemistry data, it provides predictive condensed- phase Molecular Dynamics simulations including quantum nuclear effects. Its full-range of capabilities is demonstrated by modelling: water properties, ions in solution, small molecules hydration free energies, complex folding free-energy landscapes, large-scale protein dynamics, protein-ligand binding, chemical reactions and by coupling it to protein structure prediction foundation models’ outputs for further refinement. FeNNix-Bio1 is accurate and systematically improvable while limiting human parametrization efforts. Big kudos to all other co-authors: @OlivierApp, @QMC_Anouar, @EPosenitskiy, C. Villot, L. Lagardère. Implemented and #GPU-accelerated in Tinker-HP/Deep-HP @TINKERtoolsMD. Funding @ERC_Research (project EMC2). A big thank to @Genci_fr (grand challenge H100 @ IDRIS/Jean Zay), @EuroHPC_JU (Cineca/Leonardo) and to @argonne (#INCITE project/Aurora) for #supercomputing time.
1
17
2
70
11,788
Matthias Kellner retweeted
Last week, we continued @lab_COSMO's study of solid-state electrolytes, focusing on the properties of the LPS surfaces arxiv.org/abs/2504.11553. Many thanks to my office mate Hanna Tuerk who did a lot of the work
1
5
13
473
🤫 you can get a better universal #machinelearning potential by training on fewer than 100k structures. too good to be true? head to arxiv:2503.14118, to the atomistic cookbook, or to the better science social if you want to find out more about PET-MAD. 🧑‍🚀 over and out 👋
1
14
47
3,637
Matthias Kellner retweeted
New: Fast and flexible range-separated models for atomistic machine learning, spearheaded by Kevin and Philip, with many cosmonauts and @MicheleCeriotti ! We present a fast and differentiable torch and JAX implementation of (Particle Mesh) Ewald. ✨ arxiv.org/abs/2412.03281
1
8
33
1,467
torch-pme has landed. if you're curious, head to the better place to see why we're so excited about it 👀
3
24
1,229
With < 2 weeks from @Faraday_D in Oxford, here's a 🧵about the @lab_COSMO contribution (doi.org/10.1039/D4FD00101J) on "Prediction Rigidities". Our paper ended up like a Korean dining table full of banchans, with something for everyone in the atomistic ML community 😋 (1/n)
1
5
1
17
2,558
(1/2) VERY excited to be part of the newest iPI 3.0 release. pubs.aip.org/aip/jcp/article… Thank you @MicheleCeriotti, @lab_COSMO, @venkatkapil24, @dr_marossi @MPSD_SabIA, and @YairLitman for letting us be part of this wonderful open source software!
1
2
19
1,196
🚨 i-PI 3.0 is out! 🚨 From 2x to 10x faster🏃 built-in ensemble propagation for #ML, driven MD, better docs & examples, and ofc all the best path integral algos - now also for bosons🤩 Read the (highlighted) @JChemPhys paper pubs.aip.org/aip/jcp/article… and update pip install -U ipi
13
4
31
10,700
Over the past ~year at @lab_COSMO we have been working on something new, and we want to share a #preview of the #atomisticcookbook: lab-cosmo.github.io/atomisti…. This is a repo/website collecting recipes to solve simulation problems 📖🧑‍🍳
1
15
2
58
5,604
🌶️🌶️🌶️ how I learned to stop worrying and let #machinelearning models learn 🌶️🌶️🌶️ 🧑‍🚀 #preprint just landed on the @arxiv: we try to show that the non-equivariant PET model by @spozdn leads to spectacular errors when used in actual simulation, and fail 🤷arxiv.org/html/2406.17747v1
1
16
8
97
22,861
Matthias Kellner retweeted
well… @spozdn and @MicheleCeriotti and I had a look: arxiv.org/abs/2406.17747! In this, we check what exactly non-invariance breaks in practical terms, for water and a PET model. To my surprise, even though angular momentum is not conserved, observables seem largely unaffected!
has anyone tested the simulation time until failure for non-invariant models versus invariant ones? maybe @spozdn? seems like that would help settle this ongoing debate about what equivariance is really needed for, in practical terms.
4
8
2
20
6,520
🚨 chemiscope.org 0.7 is out 🚨 thanks to brilliant @nccr_marvel #inspirepotentials fellow 🇺🇦 Sofiia Chorna 👏👏👏 you can now include chemiscope viewers in documentation pages, using sphinx and/or sphinx-gallery. The possibilities are endless!
1
11
24
2,246