iiinova — where your digital self keeps evolving

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

marcel ⊙ @marceldotsci.bsky.social · 04/08/2026

is jax 0.11.0 producing rather weird behaviour for anyone else? we're seeing slowdowns and nans under jit.

021

marcel ⊙ @marceldotsci.bsky.social · 07/07/2026

asked fable to "orient yourself in this project so we can continue discussing what comes next"

Meme with Arnold and the text Made 1 scratchpad edit +25, committed d7bcfd9, pushed to megabatch, searched for 1 pattern, read 4 files, listed 1 directory, ran 13 shell commands

010

marcel ⊙ @marceldotsci.bsky.social · 23/04/2026

little footgun in jax.nn.dot_product_attention for the double precision enjoyers out there: github.com/jax-ml/jax/b...

020

marcel ⊙ @marceldotsci.bsky.social · 20/03/2026

🏃🏻‍♂️ largely because the name was too tempting, partially to support the upcoming LOREM release & make CI less annoying, my hackable personal jax training infra is now on PyPI pypi.org/project/mara... — featuring a beautiful logo drawn by a dear friend

020

marcel ⊙ @marceldotsci.bsky.social · 27/01/2026

my long bet to not bother to learn vim commands has finally paid off #sloperator

140

marcel ⊙ @marceldotsci.bsky.social · 30/11/2025

i may be biased but this is a very nice example of a universal MLIP paper

010

marcel ⊙ @marceldotsci.bsky.social · 28/11/2025

reality is proving @adrianhill.de more right by the minute -- now all reviews have been reverted to their pre-discussion state. someone remind me why i spent a week+ preparing rebuttals and reading author comments?

130

marcel ⊙ @marceldotsci.bsky.social · 28/11/2025

ICLR2026 is truly the gift that keeps on giving

030

Reposted by marcel ⊙

Adrian Hill @adrianhill.de · 28/11/2025

When it comes to drama, ICLR 2026 is truly the gift that keeps on giving.

141

Reposted by marcel ⊙

Michele Ceriotti @micheleceriotti.bsky.social · 26/11/2025

made by a friend of a colleague. iykyk

a meme with scilla and cariddi and ulysses ship trying to sail through, making fun of the peer review process at ICLR

052

marcel ⊙ @marceldotsci.bsky.social · 12/11/2025

seems like i was the only ICLR reviewer that bothered to declare LLM usage

020

Reposted by marcel ⊙

Adrian Hill @adrianhill.de · 19/09/2025

This also shows that @typst.app is already a viable alternative to LaTeX for top conferences in ML!

3385

Reposted by marcel ⊙

Adrian Hill @adrianhill.de · 19/09/2025

Our paper on XAI + Automatic Differentiation got accepted at #NeurIPS2025. Can't wait to share the camera-ready soon! 🥳

0111

Reposted by marcel ⊙

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

If you are excited about 30x longer time steps in molecular dynamics using FlashMD, but are worried about it not being symplectic, Filippo has something new cooking that should make you even more excited. Head to the #arxiv for a preview arxiv.org/html/2508.01...

Orbits for a periodic 3-body system, showing the stability of a ML long-time integrator

1132

marcel ⊙ @marceldotsci.bsky.social · 08/08/2025

for the jax kids, i just made my personal training infra repo public: github.com/sirmarcel/ma... -- happy for some feedback/ideas/takes!

000

Reposted by marcel ⊙

Volker Blum @aimsduke.bsky.social · 06/08/2025

This, if true, is incredible. A long program in 2011 at IPAM that I attended practically wrote the book on machine learning in materials science, which is now AI. Not funding IPAM would be deeply destructive to US leadership in critical future technologies! Hard to fathom.

021

marcel ⊙ @marceldotsci.bsky.social · 05/08/2025

✨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 ...

130

Reposted by marcel ⊙

COSMO Lab @labcosmo.bsky.social · 20/06/2025

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

openreview.net

The dark side of the forces: assessing non-conservative force...

The use of machine learning to estimate the energy of a group of atoms, and the forces that drive them to more stable configurations, have revolutionized the fields of computational chemistry and...

1105

Reposted by marcel ⊙

COSMO Lab @labcosmo.bsky.social · 20/06/2025

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!

061

marcel ⊙ @marceldotsci.bsky.social · 16/06/2025

obvious in hindsight but news to me: accessing np.load-ed arrays by index is orders of magnitude slower than forcing them to RAM. notes.marcel.science/2025/numpy-l...

notes.marcel.science

numpy.load access performance – marcel's notes

marcel's science notes

000

marcel ⊙ @marceldotsci.bsky.social · 04/06/2025

very annoying when papers don‘t include raw data for plots

010

marcel ⊙ @marceldotsci.bsky.social · 19/03/2025

cosmo has joined the universal FF club!

010

marcel ⊙ @marceldotsci.bsky.social · 16/02/2025

watch out cancer, mihail is coming for you 🔥

010

Reposted by marcel ⊙

Adrian Hill @adrianhill.de · 13/02/2025

Completed my post preprint project: a split keyboard to help with my wrist pain (pictured with the pain-relieving gel it is trying to replace).

Two halves of a 3D printed ergonomic split keyboard surrounding pain relief gel.

3142

marcel ⊙ @marceldotsci.bsky.social · 11/02/2025

anyone at the lausanne applied ML days?

000

Reposted by marcel ⊙

COSMO Lab @labcosmo.bsky.social · 05/02/2025

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

0122

Reposted by marcel ⊙

Adrian Hill @adrianhill.de · 30/01/2025

You think Jacobian and Hessian matrices are prohibitively expensive to compute on your problem? Our latest preprint with @gdalle.bsky.social might change your mind! arxiv.org/abs/2501.17737 🧵1/8

Figure comparing automatic differentiation (AD) and automatic sparse differentiation (ASD).

(a) Given a function f, AD backends return a function computing vector-Jacobian products (VJPs). (b) Standard AD computes Jacobians row-by-row by evaluating VJPs with all standard basis vectors. (c) ASD reduces the number of VJP evaluations by first detecting a sparsity pattern of non-zero values, coloring orthogonal rows in the pattern and simultaneously evaluating VJPs of orthogonal rows. The concepts shown in this figure directly translate to forward-mode, which computes Jacobians column-by-column instead of row-by-row.

314028

marcel ⊙ @marceldotsci.bsky.social · 28/01/2025

the recording of my talk at orbital materials, together with jonathan schmidt, is out: www.youtube.com/watch?v=0xpR...!

youtube.com

Big Data & Symmetry in Machine Learning Models, with Marcel Langer & Jonathan Schmidt

YouTube video by Orbital

020

marcel ⊙ @marceldotsci.bsky.social · 17/01/2025

now on day ~2 of debugging bizarre multiprocessing issues with Google Grain, the alleged hot new data loading framework. has anyone here actually used it by any chance?

100

Reposted by marcel ⊙

marcel @sirmarcel.bsky.social · 12/01/2025

nice find while back-filling links for tabs.marcel.computer: crosstowntrail.org

crosstowntrail.org

SF Crosstown Trail – The San Francisco Crosstown Trail

001

marcel ⊙ @marceldotsci.bsky.social · 06/01/2025

guess i should mention this: as of this month, i’m honoured to receive a walter-benjamin-fellowship from the DFG to fund my stay @labcosmo.bsky.social! i’m very happy this worked out, and look forward to working on the project!

040

marcel ⊙ @marceldotsci.bsky.social · 17/12/2024

we did some further work on the role of physical priors in machine learning potentials. this time on conservative forces (obtained as gradient of the energy). turns out that not doing that, i.e., direct force prediction, is problematic for simulations! we do some tests & discuss solutions. ✨

0132

marcel ⊙ @marceldotsci.bsky.social · 08/12/2024

gave myself a little makeover!

000

marcel ⊙ @marceldotsci.bsky.social · 17/11/2024

guess time to turn this thing on…

030