accepted now in Machine Learning: Science and Technology! final manuscript, scripts, and data coming soon. iopscience.iop.org/article/10.108…

Hot topics require hot thermostats 🌡️

Simple, cheap, and surprisingly accurate. What's not to like of uncertainty estimation from repurposed prediction rigidities?

Kind of bittersweet that it's so easy to lean equivariance to the point where it does not matter, but that's the evidence 🤷. Also, if you see similar stuff, there's a good set of benchmarks to stress test your unconstrained model!

It is time for jack-o-molecule 🎃! Happy chemiscoping chemiscope.org/?load=https%3A…

Let's go!

torch-pme has landed. if you're curious, head to the better place to see why we're so excited about it 👀

Happy new year everyone! After more than three years of work, and splitting out a whole separate package, I am extremely happy to announce the first full public release of featomic, a package to compute representations for atomistic machine learning! 1/n

In this #EdSugg, researchers COSMO Lab EPFL introduce an improved machine learning method for predicting the electronic density of states of a material, which uses a self-aligning loss function to provide an adaptive energy reference during training go.aps.org/4aEZfFK

🤫 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 👋

🚨help needed!🚨 Does anyone, ORB-affiliated or not, can share with us the model weights orb-v2-20241011.ckpt? We need urgently to reproduce some results we obtained with that model checkpoint, but it was removed from their huggingface repo because "the model reached EoL" 🤷. 🙏

📢For those still hanging out here, a new #cookbook recipe landed, demonstrating >2x speedup of *conservative* MD for the PET-MAD universal #MLIP using multiple time stepping. Check out the details at the other place, or head to atomistic-cookbook.org/examples/pet-m… #compchem #machinelearning

✨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...

Last week, while we were all busy at #psik2025, my paper with Hanna Türk got published in @PRXenergy. We used COSMO Lab's #PET models to study the dynamic reconstruction of LPS surfaces, and how it impacts their structure, stability, and reactivity. journals.aps.org/prxenergy/abst…