CAIG: AI-Enabled Geochemical Modeling for Research and Education

Geochemical modeling has applications throughout the geosciences including fields where its impact has been historically modest. We propose creating the first comprehensive body of human-written geochemical modeling documentation optimized for AI-assisted geochemical modeling through retrieval-augmented generation (RAG). This involves providing comprehensive context about modeling concepts, tools, and techniques written by human experts that allow LLMs to craft responses useful enough to assist with geochemistry research and education. We will also provide a unique open science platform for users to interact with LLMs through the existing NSF-funded Water-Organic-Rock-Microbe (WORM) Portal online geochemical modeling platform via the new “Ask WORM” feature; a novel application of LLM technology to geochemistry and thermodynamics where specialized models do not yet exist. Users will be able to seamlessly interact with LLMs informed by tutorials, code documentation, and decades of practical wisdom to run geochemical models in a safe, self-contained sandbox. The geochemical modeling tools we created for the WORM Portal are custom open-source Python packages that permit LLMs to perform geochemical speciations, mass transfer reactions, calculate microbial energy supplies, estimate properties of aqueous molecules, and more. We will provide safeguards to reduce AI hallucinations and mistruths by ensuring calculations are calibrated against a "ground truth" (e.g., replicating results from published studies), and iteratively updating the geochemistry context informing the LLM as it is tested in classrooms and research projects. All context we provide for RAG will be version controlled and disseminated as an open-source AI framework that anyone can adopt to enhance their own LLMs.