parametrization-clean-cdaksha 1.0.2.2

ReaxFF parameter optimization scheme using generational genetic algorithm and neural networks.

ReaxFF Parametrization with Clean Architecture

Contains Python files and Bash scripts as a basis for automated ReaxFF parametrization. Uses the genetic algorithm (GA) algorithm as well as an artificial neural network (ANN) to optimize a ReaxFF parameter set. Note that this version is refactored to attempt to comply with Robert Martin's Clean Architecture guidelines. Using the generational genetic algorithm and neural net (the latter only if enabled), runs one generation, awaiting submission of standalone ReaxFF optimizations.

Getting Started

As with all work flows, it is a good idea to work within a Python virtual environment to create an isolated environment for this Python project. This source contains a good tutorial on how to setup a virtual environment. Another source for help with Python virtual environments is here.

The project files are available as a GitHub repository here. The project can also be accessed through PyPi here; the corresponding pip installation command is

$pip install parametrization-clean-cdaksha

If you don't have pip and/or Python installed, then this guide may prove helpful in performing a basic setup. If, for whatever reason, a distribution manager such as pip or conda is not available, then the required packages for running the application are shown in requirements/prod.txt.

After installing the package with pip, the current implementation supports a command-line interface with usage

$cli --g GENERATION_NUMBER --t TRAINING_PATH --p POPULATION_PATH --c CONFIG_PATH

where GENERATION_NUMBER is the current generation number in the generational genetic algorithm, TRAINING_PATH is the file path location of the reference ReaxFF training set files, POPULATION_PATH is the location at which the user wishes the generational genetic algorithm files to be outputted, and CONFIG_PATH is the location at which a user-defined JSON configuration file can be entered. The last field is not required, as defaults are provided for each algorithm and genetic algorithm setting. If the last field is specified, a valid user configuration JSON file is assumed to exist at the CONFIG_PATH location.

All options used in the default configuration are shown in the example folder here. The user can tune one (or many) of these parameters by defining a config.json file at the CONFIG_PATH location, such as the following:

{
  "strategy_settings": {
    "mutation": "nakata"
  },
  "ga_settings": {
    "population_size": 50,
    "use_neural_network": true
  }
}

Note that, at the very least, the user should define the population_size parameter. This parameter controls the number of individuals in the genetic algorithm's population.

If GENERATION_NUMBER = 1, then the first population is initialized, whose ReaxFF optimizations can then be submitted for evaluation of the parameters. If GENERATION_NUMBER > 1, the previous generation's data is read from POPULATION_PATH, and classic genetic operators are applied to generate the next generation and output to the POPULATION_PATH once again, after which the corresponding ReaxFF optimizations may be submitted.

To automate the generational genetic algorithm, an example bash script is provided in the example directory here. This allows concurrent submission of ReaxFF optimizations and continuation of the generational genetic algorithm until a threshold, defined by a maximum generation number, is reached. In practice, this example simply uses this Python application to propagate the genetic algorithm from one generation to the next, then submits ReaxFF optimizations for those created individuals, monitoring their completion. After the generation is completed (based on job completion status), the cycle repeats.

In practice, this application lends itself to usage with supercomputing. The corresponding supercomputing job for a SLURM-based environment is also available here. This wrapper SLURM script merely calls the bash script, but makes it so that the user does not need to keep the bash script job running on their own computer; instead, the bash script will be running on a node in the supercomputer.

Again, note that several options are provided for potential mutation, crossover, etc., algorithms that the user may use. Reasonable defaults based on the literature are provided, but they are easy to override by defining the custom config.json file and providing the location to the command line interface, as suggested earlier.

Thus, the typical setup for usage of this application is as follows:

1. Make sure that the application and requirements are properly installed on the machine, as well as standalone ReaxFF.
2. Define the user configuration file with any changes to parameters of interest, or at least the population size.
3. Create a wrapper bash script that automates propagation of generations back-to-back based on the example bash script provided, with proper submission details for the standalone ReaxFF jobs.
4. If using supercomputing resources, create a wrapper job script (SLURM, PBS, etc.) that submits the wrapper bash script, configured appropriately based on the environment.

The workflow for running the optimizations is then as follows:

1. Make sure the user configuration is set as desired and reconfigure beforehand if necessary.
2. Set the reference/training set directory with the required model files: ffield, geo, params, control, trainset.in, and fort.99.
3. Set the population output directory, where you wish the generational GA files to be outputted.
4. Set the generation number. If this is the very beginning of the run, then the generation number should equal one. If a previous run is being continued, then this generation number should be the last successfully completed generation's number incremented by one.
5. Submit the supercomputer job script (SLURM, PBS, etc.).
6. Monitor the total error/objective function as the generations go on, as well as any other statistics of interest.

Dependencies

Through PyPI, the installation should already come with NumPy, Pandas, and Click. TensorFlow 2.0 is used for building, training, and using the feed forward neural network, but is NOT automatically installed. This is because the application can run without TensorFlow, as long as the option to "use_neural_network" is not true, allowing for compatibility with systems that cannot use TF 2.0. However, those who wish to utilize the neural network can run

$pip install -r requirements/prod.txt

which will check and install all required modules as listed in the requirements/prod.txt file, including TensorFlow.

Prerequisites

Project relies on usage of pip for installing required dependencies. Additionally, standalone ReaxFF is required to run the optimizations for the files that are created. Note that reference ReaxFF training files for the system at hand are required. At the very least, the training set directory must contain

training_files
|---ffield
|---geo
|---params
|---control
|---trainset.in
|---fort.99

Note that iopt files are dynamically created with a single line entry, 0, to instruct ReaxFF not to use the "manual" ReaxFF parameter optimization scheme: successful one parameter parabolic extrapolation (SOPPE).

Currently, fort.99 is required in the training set directory to retrieve DFT energies and weights in the beginning.

Note that more information on the ReaxFF files and required setup for optimization can be found in the ReaxFF User Manual from the lead developer, Dr. Duin. More information about ReaxFF itself can be found by first looking at Dr. Duin's first paper, ReaxFF: A Reactive Force Field for Hydrocarbons, and going from there.

Running the tests

The project can easily be tested by running the test suite through

$tox

in the project root after installation. Note that TensorFlow is used in building the neural network. If TensorFlow is unavailable for installation, then the tests corresponding to the neural network will not run. Code coverage can be checked by running

$py.test --cov-report term-missing --cov=parametrization_clean

To check for conformation to PEP standards, one can use

flake8

in the project root directory.

Authors

• Chad Daksha - Initial work - cdaksha

License

This project is licensed under the MIT License - see the LICENSE.md file for details