CHEAT is a suite of modules for inference of adsorption energies and modeling catalytic reactions on high-entropy and solid-solution alloys. This workflow was originally published in High Entropy Alloys & Materials and the original version of this repository can be found in the v1.0-legacy branch.
It is recommended to fetch the latest version of the main branch using:
git clone https://github.com/catalyticmaterials/cheat.git
as the newest implemented features might not be included in the latest release yet.
The required packages can be installed into a conda environment running:
conda env create -f env.yml
Note the different installation procedure if you intend to use inference models from FAIR Chemistry.
After environment creation, activate the environment and navigate to this folder to install cheatools:
conda activate cheat
pip install -e .
The examples folder contains working examples of different applications with further explanation and instructions within each subdirectory. Start by unzipping the files in the gpaw folder. These contain pre-calculated DFT trajectories (GPAW 22.1.0) of *OH and *O on Ag-Ir-Pd-Pt-Ru high-entropy alloy surfaces which will form the basis for these examples.
run_dft demonstrates querying your own DFT calculations used to train the inference algorithms. This aids in sampling multiple binding sites on the same slab to minimize compute per adsorption energy optained. Note that this requires installing GPAW and some additional setup to conform to whatever high-performance cluster you are using.
train_lgnn reduces the optimized geometries from the DFT calculations to graph features and subsequently trains a lean graph neural network (lGNN) to perform adsorption energy inference.
surface_simulation emulates a solid-solution alloy surface via a grid-based approach. This surrogate surface is used in conjunction with the lGNN to infer the distribution of adsorption energies on the surface. Additionally, competitive co-adsorption of different species can be included for certain sites.
bayesian_optimization applies the above step in a Bayesian optimization procedure to maximize a catalytic activity by sampling surfaces within a specified composition space.
fairchem_is2re showcases the finetuning and application of a more advanced adsorption energy inference model from FAIR Chemistry and it's implementation in the surrogate surface simulation.
fairchem_s2ef demonstrates how to finetune a machine learning potential from FAIR Chemistry, providing a much(!) faster alternative to a DFT calculator.
plots contains examples of a few plotting functions that can be handy to visualize results in high-dimensional composition space.