CiderPress

Tools for training and evaluating CIDER functionals for use in Density Functional Theory calculations. Interfaces to the GPAW and PySCF codes are included.

What is the CIDER formalism?

Machine Learning (ML) has recently gained attention as a means to fit more accurate Exchange-Correlation (XC) functionals for use in Density Functional Theory (DFT). We have developed CIDER, a set of features, models, and training techniques for efficiently learning the exchange and correlation functionals. CIDER stands for Compressed scale-Invariant DEnsity Representation, which refers to the fact that the descriptors are invariant under squishing or expanding of the density while maintaining its shape. This property makes it efficient for learning the XC functional, especially the exchange energy.

WARNING: The CiderPress Code Base is Experimental

We want to make clear that both the code and the functionals themselves are experimental. The code base will likely change significantly in the next few years. Therefore, please read the installation guidance, usage instructions, examples, and known issues thoroughly before using CiderPress.

Installation

CiderPress can be installed via PyPI or from source. However, both methods require a few prerequisites (i.e., things you need to set up before running pip install). These dependencies are the following:

• Python 3.9-3.12
• Intel Math Kernel Library (MKL)
• Libxc
• C and C++ compilers with OpenMP support. Note that while mkl and can be installed via pip, this installation is not done automatically by pip install ciderpress because we assume that some users will want to use an external (non-pip) MKL installation.

If you have conda, you can install the MKL and libxc dependencies using

conda install mkl"<=2024.0" mkl-devel"<=2024.0" mkl_fft libxc

If you are using pip, you can install MKL using

pip install mkl"<=2024.0" mkl-devel"<=2024.0" mkl_fft

libxc is not available through pip, so it must be installed separately. It is available on some Linux package managers, e.g.

apt install libxc-dev

Otherwise, you can get the source from https://libxc.gitlab.io/.

For the MKL dependencies, the <=2024.0 is to fix a compatibility issue with PyTorch and MKL, so you can remove it if you don't need PyTorch (i.e. if you don't want to use CIDER24X models). getting the libraries to link. Please make sure MKL and Libxc can be found by cmake before building.

Once you have those dependencies set up, you can install CiderPress via the usual

pip install ciderpress

Note that currently we do not have wheels available for installation, so this command might take a while because it needs to compile C extensions and other dependencies.

You can also install from source if you have the code in the directory <CiderPress>:

cd <CiderPress>
pip install .

The pip command will install all the other Python dependencies, including the quantum chemistry package PySCF.

To use previously developed functionals, you can download them using

cd <CiderPress>
python scripts/download_functionals.py

See the examples directory for details on how to load and use the functionals.

If you want to run plane-wave DFT calculations, you must also install GPAW with LibXC and FFTW. GPAW uses a siteconfig.py file to customize the libraries it links to. This repo's .github/workflows/gpaw_siteconfig.py could be useful for compiling GPAW with MKL, LibXC, and FFTW.

NOTE: GPAW uses MPI for parallelization, and the CiderPress extensions must also link to MPI to run parallel GPAW calculations. If you want to run parallel GPAW calculations with CIDER, make sure cmake can find OpenMPI or an equivalent installation and that you have a working mpicc compiler before building CiderPress and GPAW together. In principle, the MPI version should work when installing from PyPI or source as long as MPI is found by cmake, but so far we have only tested it using installation from source.

NOTE: The CiderPress C extensions must use the same OpenMP as PySCF and GPAW. If they use different OpenMP installations, you will run into parallelization issues and code crashes. The provided gpaw_siteconfig.py assumes Intel's iomp5 as the OpenMP library by default. You can use GNU instead by changing iomp5 to gomp in gpaw_siteconfig.py.

The cmake configuration file (ciderpress/lib/CMakeLists.txt) automatically selects OpenMP using the find_package(OpenMP) command. If iomp5 or gomp is found, that version is used as the MKL threading layer, and calls to MKL made by CiderPress will run in parallel. Otherwise, the MKL threading layer is set to sequential, and MKL calls made by CiderPress will not run in parallel.

NOTE: To run the CIDER24X functionals, you also need to install pytorch.

More Detailed Instructions for Conda Installation.

This section covers how to install CiderPress and its dependencies from a fresh conda environment. Micromamba is also supported; you will just need to replace the 'conda' commands with 'micromamba' below.

0. Make sure you have a C compiler installed. (Or you can install one through conda after creating your environment in step 1.)

1. Create a new conda environment.

conda create -n <my_env> python=3.11
conda activate <my_env>

Python 3.9-3.12 are supported.

2. Install dependencies. The scripts .github/workflows/mm_install_torch.sh and .github/workflows/mm_install_mpi.h can both be used to set up an environment for running CIDER calculations. mm_install_torch.sh installs MKL, libxc, FFTW, and pytorch, so it is useful if you want to run calculations with CIDER24X functionals, which require pytorch. mm_install_mpi.sh installs MKL, libxc, FFTW, OpenMPI, and mpicc, so it is useful if you want to run GPAW calculations. Note that the conda MPI installation might not work well for multi-node jobs on clusters, so you might want to use your own MPI/mpicc instead if that is your use case. Single-node jobs should work fine with conda's MPI.

3. Build C extensions and install CiderPress.

pip install .

4. (If using GPAW) Install GPAW from source. We recommend using our siteconfig.py to link gpaw to MPI and MKL for simplicity and speed. (You can download GPAW at gitlab.com/gpaw/gpaw.)

cd <place you want to save the GPAW source>
git clone https://gitlab.com/gpaw/gpaw.git
cd gpaw
cp <CiderPress>/.github/workflows/gpaw_siteconfig.py .
python setup.py build install

Note: Currently CiderPress does not support the new GPAW version (gpaw.new), but we plan to support it in the future.

How can I run a CIDER calculation?

CIDER calculations can be run in PySCF (for non-periodic, all-electron calculations) and GPAW (for periodic, plane-wave PAW calculations) using the functional initializers ciderpress.pyscf.dft.make_cider_calc and ciderpress.gpaw.calculator.get_cider_functional, respectively. Periodic PySCF calculations are not yet supported, except for the CIDER24X functionals with uniform grids and pseudopotentials. See examples/pyscf/simple_calc.py and examples/gpaw/simple_calc.py for a demonstration of setting up a typical calculation, and refer to the docstrings of the initializers for a more detailed explanation of all the input options. As explained in the docstrings, the defaults are sufficient for most of the input options. The recommended functional for most applications in which the goal is to reproduce hybrid DFT is the CIDER23X_NL_MGGA_DTR exchange functional, which is a meta-GGA with nonlocal features of the density. The more recent CIDER24Xne and CIDER24Xe functionals use more powerful descriptors and are therefore more accurate than any of the CIDER23X functionals, but they are also more expensive and only available for use in PySCF. CIDER24Xe is fit to molecular HOMO-LUMO gaps, so it might be useful for properties where band gaps are important.

How can I train a CIDER functional?

The basic ML training framework for CiderPress is stored in ciderpress.models. You are free to use these tools if you find them helpful, but they are not yet documented or prepared for widespread use. Also, CiderPress currently only contains the ML model classes themselves, but not the various training tools. If you are interested in training your own CIDER model, we suggest reaching out to us to discuss (email [email protected]).

Known Issues

CiderPress has a few known issues that we are currently investigating. Please be aware of these when attempting calculations with CIDER functionals. We will make a note and publish a new release when we fix these issues. If you run into any other problems, please post an issue on the Github repository.

• For some periodic systems in GPAW within the PAW formalism, significant numerical instability issues arise for the nonlocal functionals. In our experience thus far, these issues are uncommon and seem to be caused by the nonlocal PAW corrections to the CIDER features as opposed to the functionals themselves.
• For the GPAW interface, the memory overhead for the nonlocal features can be fairly high, occasionally causing memory issues. Please be aware that you might need to allocate more memory for a nonlocal CIDER calculation than for, say, a PBE calculation.
• For the PySCF interface, there are (mostly minor) convergence issues for some systems. These issues are much less common and less severe for our most robust functionals (like NL-MGGA-DTR). Even for NL-MGGA-DTR, occasionally a system will not quite converge. Usually the energy convergence is fine, but the orbital gradients are somewhat unstable; it might be necessary to set conv_tol_grad to a higher value than the default. These issues are likely a mix between inherent functional stability and the stability of the fast feature evaluation algorithm.
• The code spits out a lot of divide-by-zero and invalid value warnings from numpy, which occur because (as with many functionals) some terms in CIDER functionals become numerically unstable at very small densities. These issues are corrected by setting the XC energy and potential at very low density to zero, and we will clean up various warnings and unnecessary debug statements as soon as possible.
• The construction of the CIDER PAW corrections within GPAW have a very small numerical stability issue that results in different energies on different runs (with energy differences of roughly $10^{-11}$ eV). The difference is so small that it is insignificant for most applications, but it might affect finite difference calculations with very small perturbations.

Questions and Comments

Find a bug? Areas of code unclearly documented? Other questions? Feel free to contact Kyle Bystrom at [email protected] AND/OR create an issue on the Github page at https://github.com/mir-group/CiderPress.

Citing

If you find CiderPress or CIDER functionals useful in your research, please 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}
}

The above article introduces the CIDER23X functionals and much of the algorithms in CiderPress. If you use the CIDER24X functionals, please 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},
  eprint = {https://doi.org/10.1021/acs.jctc.4c00999}
}