CiderPress: Machine-Learned Exchange-Correlation Functionals

CiderPress provides tools for training and evaluating CIDER functionals in density functional theory calculations. Interfaces to PySCF and classic GPAW are included to enable full DFT calculations with CIDER functionals.

Installation

CiderPress contains compiled C/C++ numerical kernels. A working C/C++ build toolchain, CMake, BLAS/LAPACK, and an FFT implementation are required. PySCF is installed as a core Python dependency; GPAW is a separate optional host code.

Install a published release with:

pip install ciderpress

From a source checkout, use:

git clone https://github.com/mir-group/CiderPress.git
cd CiderPress
pip install .

Install the optional runtime required by the selected model:

pip install 'ciderpress[cider24]'  # PyTorch-backed CIDER24X models
pip install 'ciderpress[d4]'       # CIDER26XCCHEMD4 post-density energy (pyscf-dispersion + dftd4)

See the installation guide for compiler, MPI, FFT, MKL, GPAW, and source-build details.

Documentation

• Documentation index
• Choosing a packaged model
• PySCF guide
• GPAW guide
• SCF convergence recommendations
• CIDER framework and model lineage

Packaged Functional Families

CiderPress 0.5.0 includes mapped models from three CIDER generations:

• CIDER23X: six semilocal/NLDF exchange models, for PySCF and classic GPAW.
• CIDER24X: CIDER24Xne and CIDER24Xe, SDMX exchange models for PySCF.
• CIDER26XC: CIDER26XCCHEM, CIDER26XCCHEMD4, and CIDER26XCSURFSCI, full XC models for molecular chemistry and combined molecular/solid/surface-science applications. CIDER26XCSURFSCI is supported in both PySCF and GPAW. CIDER26XCCHEM is supported in both codes as well, but only recommended for molecular calculations in PySCF. CIDER26XCCHEMD4 is only supported in PySCF currently, due to the need for the D4 dispersion term.

All of these functionals can be selected and loaded by their short name anywhere an mlfunc path is accepted. The model guide lists every name, feature representation, supported backend, functional composition, and release checksum.

CIDER23X and CIDER24X store exchange and normally use the explicit PBE0/CIDER surrogate-hybrid composition. CIDER26XC stores full XC and uses the semilocal baseline contained in its model file.

Quick starts

For a molecular calculation:

from pyscf import dft
from ciderpress.pyscf.dft import make_cider_calc

mf = make_cider_calc(dft.RKS(mol), "CIDER26XCCHEM")
energy = mf.kernel()

For an exchange-only model:

mf = make_cider_calc(
    dft.RKS(mol),
    "CIDER23X_NL_MGGA_DTR",
    xmix=0.25,
    xkernel="GGA_X_PBE",
    ckernel="GGA_C_PBE",
)

For a classic GPAW calculation, load the periodic model as a full-XC functional:

from ciderpress.gpaw.calculator import get_cider_functional

xc = get_cider_functional(
    "CIDER26XCSURFSCI",
    xmix=1.0,
    xkernel=None,
    ckernel=None,
    pasdw_store_funcs=False,
)

D4 evaluation is supported by PySCF and is added once after the density SCF. The current CIDER molecular gradient contains the electronic contribution. The GPAW interface reports an error when a D4 model is selected. CiderPress 0.5.0 supports classic GPAW with PAW setups; gpaw.new is outside this release interface.

Complete examples and fallback ladders to fix SCF convergence problems are linked in the documentation list above.

Support

Use the GitHub issue tracker for bug reports, documentation problems, and technical questions. Scientific inquiries may also be sent to Kyle Bystrom at kylebystrom@gmail.com.

Citing

For work using CiderPress or CIDER functionals, cite the following article:

@article{PhysRevB.110.075130,
  title = {Nonlocal machine-learned exchange functional for molecules and solids},
  author = {Bystrom, Kyle and Kozinsky, Boris},
  journal = {Phys. Rev. B},
  volume = {110},
  issue = {7},
  pages = {075130},
  numpages = {30},
  year = {2024},
  month = {Aug},
  publisher = {American Physical Society},
  doi = {10.1103/PhysRevB.110.075130},
  url = {https://link.aps.org/doi/10.1103/PhysRevB.110.075130}
}

This article introduces the CIDER23X functionals and much of the numerical framework in CiderPress. Work using CIDER24X should also cite:

@article{doi:10.1021/acs.jctc.4c00999,
  author = {Bystrom, Kyle and Falletta, Stefano and Kozinsky, Boris},
  title = {Training Machine-Learned Density Functionals on Band Gaps},
  journal = {Journal of Chemical Theory and Computation},
  volume = {20},
  number = {17},
  pages = {7516-7532},
  year = {2024},
  doi = {10.1021/acs.jctc.4c00999},
  note ={PMID: 39178337},
  URL = {https://doi.org/10.1021/acs.jctc.4c00999}
}

The CIDER26XC models accompany the forthcoming manuscript Machine-Learned Exchange-Correlation Functionals in the CIDER Framework and Application to Chemistry and Surface Science. See the citation guide for the current citation record.