hdrl-fp

This is the source code folder for the research project published on Nature Communications: https://www.nature.com/articles/s41467-024-50531-6/figures/3

structure

1. en_array: energy landscape mesh from DFT
2. run_configs: configs for different environments
3. single_agent_one_atom: src for one atom scenario
4. single_agent_two_atom: src for two atom scenario
5. scripts: example running scripts

installation

1. setup GPU environment and install warpdrive package as instructed
2. under the root directory of rlchemists, run bash setenv.sh to setup the Python path for this project

CUDA kernel for environment

Please contact the authors for the kernel functions. The kernel functions are proprietary and are not yet open sourced for now. The code needs kernel functions to run the GPU mode.

run

We simply choose the environment and type to run a particular training, the supported ones are all included in the run_configs folders, for example, run_configs/single_agent_one_atom_diffusion2d can be run by python example_training_script_numba.py --env single_agent_one_atom --type diffusion2d

cite

If you're using this study in your research or applications, please cite using this BibTeX:

@article{lan2024,
  title  = {Enabling high throughput deep reinforcement learning with first principles to investigate catalytic reaction mechanisms.},
  author = {Lan, Tian and Wang, Huan and An, Qi},
  year   = 2024,
  journal = {Nature Communications},
  volume  = {15},
  number  = {6281},
}