Jan Hermann

Simulating molecules and materials accurately is one thing, knowing which molecules and materials to look at is another. Look at these new roles for the latter!

I benefited massively from www.ipam.ucla.edu/programs/lon.... I got into ML for science through that program. Now IPAM may be gone mathstodon.xyz/@tao/1149568...

Interested in our mission to make DFT more accurate and push what’s possible in quantum chemistry? Do you want to directly contribute? We're hiring a senior software engineer and a senior researcher:

jobs.careers.microsoft.com/global/en/jo...

jobs.careers.microsoft.com/global/en/jo...

@chrislhayes.bsky.social you achieved what I would have thought impossible. In just the first three chapters of your book you made my phone seem so disgusting that I’ve barely touched it in the last few days

Was it painful?

The OALD for example says a lie is “a statement made by somebody knowing that it is not true”. Ie it implies intent. I don’t think an LLM knows that it says an untruth. So it cannot lie

I mean, when Kepler figured out the laws of planetary motion, he also used old Babylonian astronomical data

Feynman Lectures!

GitHub - microsoft/oneqmc: Pretrained model for molecular wavefunctions Pretrained model for molecular wavefunctions. Contribute to microsoft/oneqmc development by creating an account on GitHub.

Code and pretrained model are available at github.com/microsoft/on...

Future versions of our Skala functional, bsky.app/profile/jan...., will be trained on increasingly diverse yet steadfastly accurate data, and for multireference systems we'll need every possible tool from the quantum chemistry toolbox, and then some more. With Orbformer, we're making our own tools

Orbformer does this for the first time at scale, having been pretrained on 22k equilibrium and dissociating structures. The resulting model rivals the cost–accuracy ratio of traditional multireference methods and can be systematically converged to chemical accuracy

Traditional ab initio methods run always from scratch—no taking advantage of shared electronic structure patterns between molecules. Deep QMC changes this by first pretraining a large wavefunction model that is then cheaply fine-tuned—amortizing the pretraining cost

Why care? Strong correlation appears whenever bonds snap, radicals roam, or near-degeneracy sets in—combustion, catalysis, photochemistry. Take nitrogenase, an enzyme that can break N₂ and whose active site is a poster child for strong correlation. With Orbformer we focused on bond breaking

🚀 Strong correlation is the Everest of quantum chemistry. Next to the coupled cluster highway, the multireference molecular terrain is underserved—gravel roads and promenades. With Orbformer, we're building a new infrastructure by marrying neural network wavefunctions with cost amortization at scale

Cool work! Is the distillation protocol cheap enough that you could use it with DFT directly as the teacher, skipping the foundation FF entirely?

We’ll definitely release Skala as part of some DFT library! Exact plans being finalized. We’ll get in touch when we’re ready to share details. We’d love Skala to be available in ORCA

YouTube video by Microsoft Research Deep learning for DFT

..., @marwinsegler.bsky.social, Victor Garcia Satorras, @riannevdberg.bsky.social, @paolagorigiorgi.bsky.social

www.youtube.com/watch?v=Zzt3...

..., @lab-initio.bsky.social, Deniz Gunceler, @megstanley.bsky.social, @wessel.ai, Lin Huang, Xinran Wei, Jose Garrido Torres, Abylay Katbashev, @balintmate.bsky.social, @oumarkaba.bsky.social, Roberto Sordillo, Yingrong Chen, @dbwy-science.bsky.social, Christopher Bishop, Kenji Takeda, ...

This is a highly collaborative team effort across deep learning, quantum chemistry & physics
⚡🧪 #DFT #ChemTwitter #CompChem #AI4Science

👥 The dream team: @chinweih.bsky.social, @giulia-lu.bsky.social, @derkkooi.bsky.social, Thijs Vogels, Sebastian Ehlert, Stephanie Lanius, Klaas Giesbertz, ...

To test Skala’s practical utility, we show it reliably predicts equilibrium geometries and dipole moments. Though only minimal constraints are built into its neural network design, more exact physical constraints emerge naturally as training data grows!

Which data? Trained on ~150k high-accuracy reaction energies, incl. 80k atomization energies, Skala hits an unprecedented 1.06 kcal/mol on atomization energies on W4-17. On GMTKN55 it reaches 3.89 WTMAD-2, matching SOTA hybrid functionals at the cost of semi-local DFT

What makes Skala different? Skala is a deep-learning based XC functional that bypasses expensive hand-designed nonlocal features typically used to achieve higher accuracy, by learning nonlocal representations directly from an unprecedented amount of high-accuracy data

How is DFT done today? Existing XC functionals rely on hand-crafted features from Jacob’s ladder 🪜 that trade accuracy for efficiency. Yet none achieve the chemical accuracy and generality needed for reliable predictions of the outcome of laboratory experiments

Enter Density Functional Theory (DFT), the backbone 𖠣 of computational chemistry. Although DFT can, in principle, calculate the electronic energy exactly, practical applications rely on approximations to the unknown 🔍 exchange-correlation (XC) energy functional

Why this matters? ⚛️ Electrons act as the glue holding atoms together in molecules and materials. Accurately computing their energy is key to predicting chemical and physical properties relevant for drug 💊 and material design, batteries 🔋 and sustainable fertilizers 🌱

Using deep learning to increase the accuracy of computational chemistry and density functional theory Microsoft researchers achieved a breakthrough in the accuracy of DFT, a method for predicting the properties of molecules and materials, by using deep learning. This work can lead to better batteries,...

Blog: microsoft.com/en-us/resear...
Skala paper:📜 arxiv.org/abs/2506.14665
Dataset paper:💾 arxiv.org/abs/2506.14492

🚀 After two+ years of intense research, we’re thrilled to introduce Skala — a scalable deep learning density functional that hits chemical accuracy on atomization energies and matches hybrid-level accuracy on main group chemistry — all at the cost of semi-local DFT ⚛️🔥🧪🧬

But that’s more semantics, isn’t it? The electron is described by a stationary wavefunction. Is it static or orbiting? Both words from classical mechanics that just don’t apply

If you learn the atomization energy of methane and benzene, and how to break off hydrogen of methane, can you break it off of benzene?

“typically trained on equilibrium structures and rarely include data from highly distorted geometries far from equilibrium”—Is this actually true for UMA? OMol25 contains reaction paths for at least some portion of the chemical space