Service Guide for SAFE-MolGen

The SAFE-MolGen is LLM-driven agentic system, where we augmented large language models (LLMs) using datasets from SAFE, cheminformatics tools and supervised machine learning (ML) models to generate new extractants and separation conditions for f-element solvent extraction separation. The full paper has been published. The SAFE-MolGen website here allows researchers to directly use our workflow. If you're still interested in implementation details, welcome to check SAFE-MolGen GitHub repository.

🚀 How to Use Our Service

  1. Log in to the SAFE website at https://safe.lanl.gov/login . If you don't have an account yet, please register first at https://safe.lanl.gov/signup .
  2. Go to the SAFE-MolGen website: https://safe.lanl.gov/molgen .
  3. On the website, choose your preferred LLM and provide the corresponding API key.
    • If you select a GPT model, use an API key from the OpenAI platform.
    • If you choose a Llama model, the key should come from your NERSC account (know more NERSC).
    We provide video guides for both options on the website. Note that we do not store your API key.
  4. You can enter your parameters in the input fields provided to run the workflow.
  5. Click the "Generate New Extractants" button to start the generation process.
  6. Optionally:
    • Click "Preview Experimental Evaluation Table for Prompt" to view the evaluation data and associated DOIs.
    • Click "Preview LLM's Prompt" to see the LLM prompt that will be sent to the model.
  7. Please wait for ~10 minutes as the SAFE-MolGen will generate and evaluate the new extractants.
  8. Once completed, you will receive your results in your registered email. You can download your results (including New Extractants and corresponding experimental separation conditions) and parameters using the provided download button.

📺 More Resources

To know more of how SAFE-MolGen works:

🖥️ Website Implementation

The SAFE-MolGen website is implemented using NERSC SPIN to host the web UI. The open source llama model weights are stored in global common of Perlmutter HPC and the LLM is executed via Perlmutter HPC GPU-based LLM inference via batch jobs. Maintaining up-to-date documentation of this project is important based on NERSC appropriate use policy documentation and is reponsible for Separation ML team.

📄 About SAFE-MolGen

  • About us: Accelerating f-element separation with machine learning
  • Contact us: Contact Page
  • Cite our work: Zhang, B.; Summers, T. J.; Augustine, L. J.; Taylor, M. G.; Geist, A.; Li, R.; Batista, E. R.; Perez, D.; Yang, P.; Schrier, J. Augmenting Large Language Models for Automated Discovery of f-Element Extractants. J. Am. Chem. Soc. 2026. DOI: https://doi.org/10.1021/jacs.5c19738