Adsorb-Agent

Autonomous Identification of Stable Adsorption Configurations via LLM Agent

---

Overview

Adsorb-Agent is an LLM-powered tool designed to identify the most stable adsorption configurations on catalytic surfaces. By leveraging built-in knowledge and emergent reasoning capabilities of Large Language Models, Adsorb-Agent efficiently reduces the computational cost associated with traditional exhaustive search methods while maintaining accuracy.

This repository also includes a baseline algorithmic approach (ocp-demo) for direct comparison.

⚠️ Note: This project is currently under construction for perfect public usage. Some features may change or be updated, and improvements are ongoing.

---

Getting Started

Prerequisites

Ensure you have the following installed:

• Python 3.8+
• Required libraries (install via requirements.txt):

  pip install -r requirements.txt

• Git clone the fairchem-forked repo. This includes the adsorbate placement strategy used in this study.

  git clone https://github.com/hoon-ock/fairchem-forked.git
  

Running Adsorb-Agent

To execute Adsorb-Agent, use the following command:

python adsorb_agent.py --path adsorb_agent_config_file

Replace adsorb_agent_config_file with the path to your configuration file.

Running OCP-Demo (Baseline Algorithmic Approaches)

For comparison purposes, you can run the baseline algorithmic approach using the following command:

python ocp_demo.py

Postprocessing

After running Adsorb-Agent or OCP-Demo, postprocessing is required to filter out anomalies and identify the most stable adsorption configuration.

• Postprocessing Adsorb-Agent results:

  python postprocess.py --dir result_save_path

  Replace result_save_path with the directory where the Adsorb-Agent results are saved.
• Postprocessing OCP-Demo results:

  python postprocess_ocpdemo.py --path result_save_path

  Replace result_save_path with the directory where the OCP-Demo results are saved.

Citation

If you use Adsorb-Agent in your work, please cite the following:

BibTeX:

@misc{ock2024adsorbagent,
      title={Adsorb-Agent: Autonomous Identification of Stable Adsorption Configurations via Large Language Model Agent}, 
      author={Janghoon Ock and Tirtha Vinchurkar and Yayati Jadhav and Amir Barati Farimani},
      year={2024},
      eprint={2410.16658},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.16658}, 
}

Contact

For questions, feedback, or further information, please contact:

Janghoon Ock
Email: [jock@andrew.cmu.edu]