Parallel global optimization of gas phase and surface systems.
Copyright (C) 2018 Huanchen Zhai
Zhai, Huanchen, and Anastassia N. Alexandrova. "Ensemble-average representation of Pt clusters in conditions of catalysis accessed through GPU accelerated deep neural network fitting global optimization." Journal of chemical theory and computation 12 (2016): 6213-6226.
Zhai, Huanchen, and Anastassia N. Alexandrova. "Local Fluxionality of Surface-Deposited Cluster Catalysts: The Case of Pt7 on Al2O3." The journal of physical chemistry letters 9 (2018): 1696-1702.
This code contains three sub-packages. ACNN contains all core algorithms. PGOPT includes settings for supercomputer environment. STMOLE defines the interface to VASP and TURBOMOLE.
If you want to test the code in a non-supercomputer environment, PGOPT and STMOLE are not needed.
Here we will explain installation steps using a docker (https://docs.docker.com/) image of anaconda2 (https://hub.docker.com/r/continuumio/anaconda). Please first install the docker application in your system, then open a terminal to continue. (You can also use anaconda with python2 installed in your local environment without docker.) The following commands will download a image of anaconda2:
docker pull continuumio/anaconda
docker run -it --rm continuumio/anaconda /bin/bash
Now you are inside the docker container. We need to add the following python packages and several packages needed for compiling fortran code:
pip install theano reportlab dill
conda install pygpu
apt-get update
apt-get -y install gfortran g++ make vim
Now get the copy of PGOPT:
cd ~
git clone https://github.com/hczhai/PGOPT.git PGOPT-PROGRAMS
Compile the fortran code. This will generate some warnings but the process should not produce any error.
cd ~/PGOPT-PROGRAMS/ACNN/formod
make
Now add the following to ~/.bashrc (you may need vim ~/.bashrc first):
BASE=~/PGOPT-PROGRAMS
export STMOLE_HOME=$BASE/STMOLE
export ACNNHOME=$BASE/ACNN
export PGOPTHOME=$BASE/PGOPT
export PATH=$STMOLE_HOME:$PATH
export PATH=$PGOPTHOME:$ACNNHOME:$PATH
Then apply the these environment changes:
source ~/.bashrc
The main program is called acnnmain which can be found under $ACNNHOME defined previously. You need to prepare an input file for generating structures. There are some example input files under ACNN/tests. Here as an example, we will try to generate some gas phase Pt7 structures using S-BLDA.
cd $ACNNHOME/tests/structure-generation
acnnmain pt7-gas.json
Note that the output directory is indicated in the input file. Here we can find the results in ./OUT-pt7-gas/fil_structs.xyz.0. The structures are written in XYZ format. A single file will contains more than one structures. Visualization software such as jmol is useful for examine all structures within only one file. For example, if jmol is installed, you can type:
jmol ./OUT-pt7-gas/fil_structs.xyz.0
to look at the structures.
The PGOPT code also has its own visualization implementation. It will generate a PDF file containing images of all structures in the given input file. To generate the PDF, use the following input file (this only works after you run acnnmain pt7-gas.json):
acnnmain pt7-gas-draw.json
Then you will find the PDF in ./OUT-pt7-gas/report.pdf.
The following command will generate Pt7 structures on alpha-Al2O3 surface. The surface is described by a XYZ file. There are some example surface files under $ACNNHOME/tests/surfaces. The computational cell information is written in the comment line of the XYZ file. It can be either 3 numbers or 5 numbers. The program always assume the Z direction is normal to the surface plane. If the cell size is described by 3 numbers n1 n2 n3, then the XYZ components of cell axes are a = (n1, 0, 0), b = (0, n2, 0), c = (0, 0, n3). If the cell size is described by 5 numbers n1 n2 n3 n4 n5, then the XYZ components of cell axes are a = (n1, n2, 0), b = (n3, n4, 0), c = (0, 0, n5). Usually n5 is larger than the actual height of the surface because of the added vacuum gap. So in the end of comment line there is an additional number in parenthesis, indicating the unit cell height along Z without vacuum gap.
The surface group and unit cell (the minimal unit cell, not the computational unit cell) information, indicated in the input file may help determining structure duplicates. If these information is unavailable, use the computational cell for unit cell and "P 1" for space group, and [0.0, 0.0] for space group transformation reference point.
acnnmain pt7-alpha.json
Then we can find the results in ./OUT-pt7-alpha/fil_structs.xyz.0.
The following command will try to find unique structures from a given example XYZ file containing some local minima ($ACNNHOME/tests/data/pt4b4-local.xyz).
cd $ACNNHOME/tests/filtering
acnnmain pt4b4-local.json
The unique structures will be in ./OUT-pt4b4-filter/fil_structs.xyz.0. The additional file ./OUT-pt4b4-filter/fil_list.txt.0 shows how many duplicates of each unique structure appear in the original input XYZ file (the multi column). The other additional file ./OUT-pt4b4-filter/fil_corr.txt.0 lists the structural difference data. In this file, if one line ends with *, then the structure is selected as unique structure, because its structural difference to all previous unique structures is higher than the threshold. The second last column mindm shows the minimal value over structural difference to all previous structures.
If the structure filtering should be performed on surface support clusters, the creation-surface section should be given in input file, which contains the same information as that in the input file for creation.
Note that Neural Network Fitting is only implemented for gas phase clusters containing only one type of element. Other research groups has published more general codes on this topic. For example, see JCTC, 14(7), 2018, 3933-3942.
The following command will try to fit a neural network based on an example Pt9 data ($ACNNHOME/tests/data/pt9-structs.xyz). Note that for realistic results, we need to set sample_number parameter to [200000, 20000, 20000] and epochs to 2000. This calculation normally requires large memory or GPU. If there is no gpu to use, it will automatically switch to cpu.
cd $ACNNHOME/tests/nn_fitting
export OMP_NUM_THREADS=40
acnnmain pt9-fit.json
After the fitting is finished, the fitted network will be stored in ./OUT-pt9-nn/fit_network.dill.0. We have a reference result from executing acnnmain pt9-fit.json stored under $ACNNHOME/tests/nn_fitting_ref.
Next we need to create some new structures, then use the network to optimize them. The file ./OUT-pt9-nn/fit_network.dill.0 is required for optimization.
acnnmain pt9-create.json
acnnmain pt9-opt.json
The optimized structures and energies will be in ./OUT-pt9-nn/opt_structs.xyz.0.