Multi-Modal Foundation · microsoft
Skala 1.1
A deep-learning exchange-correlation functional for DFT combining semi-local speed with hybrid accuracy.
What Skala 1.1 is
Skala 1.1 is an exchange-correlation functional for density functional theory (DFT) developed by the OneDFT team at Microsoft Research AI for Science. In traditional DFT, climbing 'Jacob's ladder' to higher accuracy typically demands computationally demanding, hand-designed non-local mathematical forms such as exact exchange. Skala 1.1 bypasses this trade-off by using a deep neural network that learns non-local electronic structure representations directly from inexpensive density features.
The functional was trained against a substantial corpus of high-level reference calculations, predominantly coupled cluster theory with single, double, and perturbative triple excitations (CCSD(T)). Its largest training subset draws from the Microsoft Research Accurate Chemistry Collection (MSR-ACC), generated using composite protocols. Supplementary datasets cover diverse chemical phenomena, including reaction pathways, noncovalent interaction profiles, conformers, and ionisation or affinity states.
In self-consistent field (SCF) calculations, Skala 1.1 achieves an accuracy exceeding that of state-of-the-art hybrid functionals on broad chemical benchmarks while maintaining the runtime characteristics and asymptotic computational scaling of semi-local functionals such as meta-GGAs.
How it works
Skala 1.1 takes electron density and derived meta-generalized-gradient approximation (meta-GGA) properties evaluated on a 3D atom-centred integration grid as its primary inputs. These input features are derived from converged B3LYP SCF density matrices without density fitting. To learn both local and non-local representations, the model applies shared local neural network layers combined with scalable non-local message passing performed across the integration grid via an auxiliary, coarser grid.
Key facts
- Model type
- Neural network exchange-correlation functional for DFT
- Input features
- Electron density and derived meta-GGA grid features
- Supported elements in training
- Elements H through Xe
- Tested elements
- H through Xe, Pb, and Bi
- Primary training dataset
- MSR-ACC/TAE25 and extended MSR-ACC subsets
- Computational scaling
- Meta-GGA asymptotic scaling (~3x prefactor relative to r2SCAN)
What you can use it for
- Main-group thermochemistry and kinetics
- Evaluating reaction energies and barrier heights using self-consistent field (SCF) calculations, yielding benchmark performance such as 2.8 kcal/mol WTMAD-2 on GMTKN55.
- Molecular geometry optimisation
- Relaxing molecular coordinates to determine equilibrium bond lengths and angles, matching the structural accuracy of range-separated hybrid functionals.
- Dipole moment evaluation
- Predicting molecular dipole vectors and magnitudes as a reliable proxy for assessing self-consistent electron density distribution.
- Noncovalent interaction modelling
- Calculating interaction energies and dissociation profiles across van der Waals complexes, hydrogen-bonded dimers, and sigma-hole interactions.
Run it here
We list this model but cannot run it yet: no pipeline_tag — needs a human.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Predicts atomization energies on the W4-17 benchmark within chemical accuracy (~1 kcal/mol mean absolute error).
- Surpasses state-of-the-art range-separated hybrid functionals on the GMTKN55 main-group benchmark with a WTMAD-2 of 2.8 kcal/mol.
- Maintains the asymptotic computational scaling of semi-local meta-GGA functionals, running with an approximate prefactor of 3 relative to r2SCAN.
- Produces self-consistent electronic densities that yield lower dipole moment errors than range-separated hybrids.
Watch out for
- Intended strictly for iterative self-consistent field (SCF) calculations and should not be run as a single-pass energy evaluation on a fixed density.
- Trained exclusively on elements H–Xe and tested up to Bi, limiting applicability to heavier transuranic or actinide systems.
- PySCF implementation memory demands can limit scale, with test limits noted at roughly 180 atoms and 5,000 orbitals (def2-TZVP) on evaluated cloud machines.
- Not classified as a production-ready model, requiring careful validation and quantum chemistry expertise to interpret results.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · mit
The mit licence carries terms worth reading before you depend on it.
Questions
Can I use Skala 1.1 commercially?
Check the licence before commercial use. The mit licence carries terms worth reading before you depend on it.
Can Skala 1.1 evaluate energy directly from an arbitrary electron density in one step?
No. The functional is specifically trained and formulated to operate within an iterative self-consistent field (SCF) loop. A single-pass pass over a static density is explicitly out of scope.
Which SCF software packages are compatible with Skala 1.1?
The functional includes reference implementations for PySCF and GPU4PySCF on CPU and GPU. A traced model is also provided for integration into other codes, including a compatible fork of GauXC, and via Azure AI Foundry's Accelerated DFT.
Which elements are supported by the model?
Skala 1.1 was trained on chemical spaces containing elements from hydrogen to xenon (H–Xe). It has also been evaluated on systems incorporating lead (Pb) and bismuth (Bi).
Does Skala 1.1 compute absolute molecular energies reliably?
As is standard for density functional theory, relative energy differences between compounds and reaction states are calculated far more reliably than absolute total energies of isolated molecules.
How to cite it
Luise, G., Huang, C.-W., Vogels, T., Kooi, D. P., Ehlert, S., Lanius, S., Giesbertz, K. J. H., Karton, A., Gunceler, D., Battaglia, S., Simm, G. N. C., Szabó, P. B., Stanley, M., Bruinsma, W. P., Huang, L., Wei, X., Garrido Torres, J., Katbashev, A., Chavez Zavaleta, R., Máté, B., Kaba, S.-O., Sordillo, R., Chen, Y., Williams-Young, D. B., Bishop, C. M., Hermann, J., van den Berg, R., Gori-Giorgi, P. (2025). Accurate and scalable exchange-correlation with deep learning. arXiv:2506.14665 [physics.chem-ph].
Written from Skala 1.1's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.