Awesome
xcauto
Arbitrary order exchange-correlation functional derivatives using JAX
This library computes arbitrary-order exchange-correlation function(al) derivatives using JAX.
The emphasis of this project is on ease of use and ease of adding functionals in Python. The focus is not (yet) on performance. Our hope is that this project can make it easier to test new implementations of functional derivatives but maybe also used directly to provide functional derivatives in a density functional theory program.
The code is in proof of concept stage. We have ideas for more.
Acknowledgements and recommended citation
JAX does all the heavy lifting by automatically differentiating the exchange-correlation functions. Please acknowledge their authors when using this code: https://github.com/google/jax#citing-jax:
@software{jax2018github,
author = {James Bradbury and Roy Frostig and Peter Hawkins and Matthew James Johnson and Chris Leary and Dougal Maclaurin and Skye Wanderman-Milne},
title = {{JAX}: composable transformations of {P}ython+{N}um{P}y programs},
url = {http://github.com/google/jax},
version = {0.1.55},
year = {2018},
}
- We have used Libxc as reference to double check the computed derivatives.
- The functional definitions for VWN and PBE were ported to Python based on the functional definitions found in XCFun (Copyright Ulf Ekström and contributors, Mozilla Public License v2.0).
Authors
- Radovan Bast
- Roberto Di Remigio
- Ulf Ekström
Installation
You can install this code from PyPI:
$ pip install xcauto
Installing a development version:
$ git clone https://github.com/dftlibs/xcauto
$ cd xcauto
$ python -m venv venv
$ source venv/bin/activate
$ pip install -r requirements.txt
$ flit install --symlink
Using on the Hylleraas Fleet
The machines in the Fleet are properly set up to run JAX on their GPUs. The instructions to install xcauto
are slightly modified, since we'll need a GPU-aware version of JAX:
$ git clone https://github.com/dftlibs/xcauto
$ cd xcauto
$ python -m venv venv
$ source venv/bin/activate
$ pip install -r requirements.txt
$ pip install --upgrade https://storage.googleapis.com/jax-releases/`nvidia-smi | sed -En "s/.* CUDA Version: ([0-9]*)\.([0-9]*).*/cuda\1\2/p"`/jaxlib-0.1.51-`python3 -V | sed -En "s/Python ([0-9]*)\.([0-9]*).*/cp\1\2/p"`-none-manylinux2010_x86_64.whl jax
$ flit install --symlink
To run:
$ env XLA_FLAGS=--xla_gpu_cuda_data_dir=/lib/cuda python -m pytest tests/test.py
The environment variable is important: JAX won't be able to use the GPU otherwise.
Example
In this example we only go up to first-order derivatives but no problem to ask for higher-order derivatives (mixed or not).
First pip install xcauto
, then:
# use double precision floats
from jax.config import config
config.update("jax_enable_x64", True)
from xcauto.functionals import pbex_n_gnn, pbec_n_gnn, pbec_a_b_gaa_gab_gbb
from xcauto.derv import derv
def pbe_unpolarized(n, gnn):
return pbex_n_gnn(n, gnn) + pbec_n_gnn(n, gnn)
print('up to first-order derivatives for spin-unpolarized pbe:')
n = 0.02
gnn = 0.05
d_00 = derv(pbe_unpolarized, [n, gnn], [0, 0])
d_10 = derv(pbe_unpolarized, [n, gnn], [1, 0])
d_01 = derv(pbe_unpolarized, [n, gnn], [0, 1])
print(d_00) # -0.006987994564372291
print(d_10) # -0.43578312569769495
print(d_01) # -0.004509994863217848
print('few derivatives for spin-polarized pbec:')
fun = pbec_a_b_gaa_gab_gbb
a = 0.02
b = 0.05
gaa = 0.02
gab = 0.03
gbb = 0.04
d_00000 = derv(fun, [a, b, gaa, gab, gbb], [0, 0, 0, 0, 0])
d_10000 = derv(fun, [a, b, gaa, gab, gbb], [1, 0, 0, 0, 0])
d_01000 = derv(fun, [a, b, gaa, gab, gbb], [0, 1, 0, 0, 0])
d_00100 = derv(fun, [a, b, gaa, gab, gbb], [0, 0, 1, 0, 0])
d_00010 = derv(fun, [a, b, gaa, gab, gbb], [0, 0, 0, 1, 0])
d_00001 = derv(fun, [a, b, gaa, gab, gbb], [0, 0, 0, 0, 1])
print(d_00000) # -0.0002365056872298918
print(d_10000) # -0.020444840022142356
print(d_01000) # -0.015836702168478496
print(d_00100) # 0.0030212897704793786
print(d_00010) # 0.006042579540958757
print(d_00001) # 0.0030212897704793786
For more examples, see the tests folder.
Functionals
List of implemented functions:
slaterx_n
slaterx_a_b
vwn3_n
vwn3_a_b
vwn5_n
vwn5_a_b
pbex_n_gnn
pbex_a_b_gaa_gbb
pbec_n_gnn
pbec_a_b_gaa_gab_gbb
b88_n_gnn
b88_a_b_gaa_gbb
lyp_n_gnn
lyp_a_b_gaa_gab_gbb
Where are all the other functionals? We will be adding more but our hope is that the community will contribute these also. It is very little work to define a functional so please send pull requests! All the derivatives you get "for free" thanks to JAX.
Ideas
Here we list few ideas that would be good to explore but which we haven't done yet:
- Check performance
- Try how the code offloads to GPU or TPU
- Verify numerical stability for small densities
- Adding more functionals
- Directional derivatives
- Contracting derivatives with perturbed densities