Skala is a neural network-based exchange-correlation functional for density functional theory (DFT), developed by Microsoft Research AI for Science. It uses deep learning to predict exchange-correlation energies from electron density features, surpasses state-of-the-art hybrid functionals in accuracy for main group thermochemistry, kinetics and non-covalent interactions, all at a computational cost similar to semi-local DFT.
Trained on a large, diverse dataset — including coupled-cluster atomization energies and public benchmarks — Skala uses scalable message passing and local layers to learn both local and non-local effects. The model has about 385,000 parameters and matches the accuracy of leading hybrid functionals.
The recommended neural functional is skala-1.1, which uses per-atom packed grids, multiple non-local layers, and symmetric contraction. The legacy skala-1.0 traced model is still loadable via load_functional("skala-1.0").
Learn more about Skala in our ArXiv paper.
This repository contains two main components:
- The Python package
skala, distributed on PyPI and on conda-forge. It contains a PyTorch implementation of the Skala model and its bindings to the quantum-chemistry packages PySCF, GPU4PySCF, and ASE. - Examples of using Skala from compiled code through LibTorch and GauXC:
GauXC is a CPU/GPU C++ library for XC functionals.
A development version with an add-on supporting PyTorch-based functionals like Skala is available in the skala branch of the GauXC repository.
GauXC can be used to integrate Skala into other third-party DFT codes.
For detailed documentation on using GauXC visit the Skala integration guide.
All information below relates to the Python package skala.
Skala supports Linux and macOS on Apple Silicon with Python 3.11 through 3.13, the latest
PySCF release (2.14), and the two latest PyTorch release lines (2.12 and 2.13).
pip install skala works out of the box and pulls every dependency from PyPI.
If you don't already have PyTorch installed, install the CPU-only wheel first
to avoid pulling a large CUDA build:
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install skalaFor a reproducible source environment, use the default environment from the committed Pixi lockfile. It uses Python 3.12, PySCF 2.14, and CPU-only PyTorch 2.13:
pixi install --locked -e default
pixi run -e default python your_script.pyRun an SCF calculation with Skala for a hydrogen molecule:
from pyscf import gto
from skala.pyscf import SkalaKS
mol = gto.M(
atom="""H 0 0 0; H 0 0 1.4""",
basis="def2-tzvp",
)
ks = SkalaKS(mol, xc="skala-1.1")
ks.kernel()The GPU install is more involved because gpu4pyscf ships CUDA-version-specific
wheels that must match your CUDA toolkit. GPU environments and helper packages
use the latest tested GPU4PySCF release, 1.8.1.
To install all dependencies from PyPI, use the GPU specific package with the matching CUDA version, e.g., for CUDA 12:
pip install skala-cuda12xThe skala-cuda13x package is available for CUDA 13.
For a reproducible source environment, choose one of the locked GPU environments:
| Environment | CUDA | PyTorch |
|---|---|---|
gpu-cuda12-torch212 |
12 | 2.12 |
gpu-cuda12-torch213 |
12 | 2.13 |
gpu-cuda13-torch213 |
13 | 2.13 |
For example:
pixi install --locked -e gpu-cuda12-torch213
pixi run -e gpu-cuda12-torch213 python your_script.pyThe workspace records CUDA 12 and CUDA 13 as explicit platforms, so the lock can
be installed while building a container without an attached GPU. Check your
driver's maximum supported CUDA version with nvidia-smi.
Run an SCF calculation with Skala for a hydrogen molecule on GPU:
from pyscf import gto
from skala.gpu4pyscf import SkalaKS
mol = gto.M(
atom="""H 0 0 0; H 0 0 1.4""",
basis="def2-tzvp",
)
ks = SkalaKS(mol, xc="skala-1.1")
ks.kernel()Skala uses a single GPU, but importing gpu4pyscf allocates memory on every
visible CUDA device. This can conflict with PyTorch and with other processes
sharing those GPUs (e.g. in MPI-parallel workloads).
Restrict CUDA to one device before launching Python:
CUDA_VISIBLE_DEVICES=0 python my_script.pyFor MPI-parallel runs, assign one GPU per local rank:
mpirun -np 4 bash -c 'CUDA_VISIBLE_DEVICES=$OMPI_COMM_WORLD_LOCAL_RANK python my_script.py'Tracked upstream at pyscf/gpu4pyscf#435.
Skala also provides an ASE calculator for energy, force, and geometry optimization workflows:
from ase.build import molecule
from ase.optimize import LBFGSLineSearch
from skala.ase import Skala
atoms = molecule("H2O")
atoms.calc = Skala(xc="skala-1.1", basis="def2-tzvp")
# Single-point energy (eV)
print(atoms.get_potential_energy())
# Geometry optimization
opt = LBFGSLineSearch(atoms)
opt.run(fmax=0.01)See microsoft.github.io/skala for a more detailed installation guide and further examples of how to use the Skala functional with PySCF, GPU4PySCF and ASE.
Skala model files (.fun) use TorchScript serialization, which can execute arbitrary code when loaded. Never load .fun files from untrusted sources.
When loading the official Skala models via load_functional("skala-1.1") or load_functional("skala-1.0"), file integrity is automatically verified against pinned SHA-256 hashes before deserialization. If you load .fun files directly with TracedFunctional.load(), pass the expected_hash parameter to enable verification:
TracedFunctional.load("model.fun", expected_hash="<sha256-hex-digest>")See the following files for more information about contributing, reporting issues, and the code of conduct:
This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.