Accurate and scalable exchange-correlation with deep learning
Summary
Microsoft Research releases Skala, a deep-learning exchange-correlation functional for DFT that achieves 2.8 kcal/mol accuracy on GMTKN55 at semi-local cost, outperforming traditional functionals across broad chemistry benchmarks.
View Cached Full Text
Cached at: 04/22/26, 10:35 AM
Paper page - Accurate and scalable exchange-correlation with deep learning
Source: https://huggingface.co/papers/2506.14665 Today we’re sharing a major Skala update: new paper and model release.
Skala is a deep-learning exchange-correlation functional for DFT developed at Microsoft Research that currently reaches 2.8 kcal/mol on GMTKN55, wins 32 of its 55 subsets, and leads on accuracy short of double hybrids at semi-local DFT cost.
Why does this matter? Electrons are the glue holding atoms together in molecules and materials. Better XC functionals mean better predictions for reaction energies, barriers, structures, and properties that matter for chemistry, materials, and catalysis.
DFT is the workhorse of computational chemistry, but its predictive power is bottlenecked by the unknown exchange-correlation functional. Climbing Jacob’s ladder usually buys accuracy by paying more compute. The result is a functional zoo—one specialized tool per domain.
Our long-run aim is to make that zoo obsolete. Skala is our bet that one deep-learning functional will be the right choice across any system or property. Today’s results are a major step toward that goal, not the destination.
The core idea: instead of leaning on increasingly expensive hand-designed non-local ingredients, Skala learns non-local electronic representations directly from the electron density with a scalable neural architecture, at semi-local DFT cost.
The model is trained on about 400k accurate energy differences spanning atomization energies, conformers, affinities, reaction pathways, and non-covalent interactions. That’s why Skala performs strongly across main-group chemistry rather than only on one narrow benchmark.
It’s not just about energies, either. We now show accurate dipoles, hybrid-level equilibrium geometries, XC integration costs that stay in the semi-local DFT regime, and practical paths into production codes—directions we partly owe to the feedback of the DFT community.
The model is available now on GitHub, PyPI and conda-forge. We’re already working to bring Skala to major DFT codes, includingPsi4andCP2K, throughGauXCand direct code integrations.
If you try Skala on a system you care about, let us know what you find.
Similar Articles
Derivative Informed Learning of Exchange-Correlation Functionals
This ICML 2026 paper introduces Derivative Informed XC-Loss (DI-Loss), a training approach for machine-learned exchange-correlation functionals that incorporates first and second derivative supervision on the Grassmannian of density matrices. Across four architectures, DI-Loss reduces total-energy MAE by 66% compared to energy and density supervision alone, and improves excited-state predictions in TDDFT calculations.
Agentic Discovery of Exchange-Correlation Density Functionals
This paper presents an agentic system using Large Language Models to automate the discovery of exchange-correlation functionals in Density Functional Theory, achieving improvements over human-designed baselines while highlighting challenges with benchmark overfitting.
@AnimaAnandkumar: Tackling a 60-year-old challenge in quantum chemistry: making density functional theory scale nearly linearly with syst…
A novel AI model using a Fourier neural operator variant enables density functional theory to scale nearly linearly with system size, allowing efficient simulations of large quantum systems like a magnesium dislocation with 80k electrons on a single GPU.
@SilkyDogfish: Super excited to release our paper from a collaboration between @Angstrom_ai and @AstraZeneca evaluating our new model …
A collaboration between Ångström AI and AstraZeneca introduces CSP-MACE-Å, a machine learning interatomic potential that aims to replace DFT in crystal structure prediction, achieving comparable accuracy at much lower computational cost.
Stein Kernelized Molecular Dynamics for Active Learning of Interatomic Potentials
Researchers from MIT, University of Warwick, and NVIDIA introduce Stein Kernelized Molecular Dynamics (SKMD), an enhanced sampling method that uses interacting particle dynamics to acquire informative training configurations for active learning and fine-tuning of machine learning interatomic potentials (MLIPs). SKMD is a stochastic variant of Stein variational gradient descent adapted for molecular dynamics, preserving the Boltzmann distribution while achieving higher model accuracy in fewer training iterations compared to baselines.