PySCF with Auto-differentiation

Installation

---

• To install the latest release, use the following commands:

# install cmake
pip install cmake

# install OpenMP runtime used with clang
# On Linux:
sudo apt update
sudo apt install libomp-dev

# On OSX:
brew install libomp

# install pyscf
pip install 'pyscf @ git+https://github.com/fishjojo/pyscf.git@ad#egg=pyscf' 
pip install 'pyscf-properties @ git+https://github.com/fishjojo/properties.git@ad' 

# install pyscfad
pip install pyscfad

---

• To install the development version, use the following command instead:

pip install git+https://github.com/fishjojo/pyscfad.git

• The dependencies can be installed via a predefined conda environment

conda env create -f environment.yml
conda activate pyscfad_env

• Alternatively, the dependencies can be installed from source

pip install numpy scipy h5py
pip install jax jaxlib jaxopt

# install pyscf
cd $HOME; git clone https://github.com/fishjojo/pyscf.git
cd pyscf; git checkout ad 
cd pyscf/lib; mkdir build 
cd build; cmake ..; make

export PYTHONPATH=$HOME/pyscf:$PYTHONPATH

---

• One can also run PySCFAD inside a docker container:

docker pull fishjojo/pyscfad:latest
docker run -rm -t -i fishjojo/pyscfad:latest /bin/bash

Running examples

• In order to perform AD calculations, the following lines need to be added to the PySCF configure file($HOME/.pyscf_conf.py)

pyscfad = True
pyscf_numpy_backend = 'jax'
pyscf_scipy_linalg_backend = 'pyscfad'
pyscf_scipy_backend = 'jax'
# The followings are optional
pyscfad_scf_implicit_diff = True
pyscfad_ccsd_implicit_diff = True