Farallon
Visual geological understanding with GPT-6 Astra.
Core trick
Has an LLM write an executable geological history that builds a 3D model, then scores it against mapped outcrops and real USGS topography.
Tags
Description
Visual geological understanding with GPT-6 Astra. An AI-written geological history generates a 3D geology model, visually compares it with ground-truth data, and infers what needs to be changed to best explain the observed geology. Speeds up a typical geologists workflow by 100x
README excerpt
# Farallon GPT-6 reads a geological map, writes an executable geological history, and tests the resulting 3D world against the observed outcrops. It starts with undeformed sedimentary layers and can add deformation and young Quaternary deposits. All eight mapped units, including yellow cover, contribute to the fit. The demo uses the northwest nose of Sheep Mountain, Wyoming, from the supplied Fiore Allwardt et al. (2007) map, intersected with real USGS 3DEP topography. <img width="1200" height="729" alt="screen-record-2-10x-slim" src="https://github.com/user-attachments/assets/f52758a8-5cab-4c59-92d1-458f93520ff3" /> ## Run locally Dependencies are already installed on this laptop. From this directory: ```sh ./run.sh ``` Open **http://127.0.0.1:3000**. Ctrl+C stops both services. If the app is already running, just open the link. Ports 3000 and 8000 must otherwise be free. For a fresh installation, use Node 22.13+ and uv: ```sh uv sync npm --prefix web ci ``` The backend reads `OPENAI_API_KEY` from the environment or a root `.env` file. See `.env.example`; do not place a key in the frontend. Without a key, manual simulation and saved replay still work. **Test next hypothesis** makes one paid proposal call. **Start auto** runs up to the chosen number of iterations, with one paid proposal per iteration. No model calls run automatically on page load or when the dropdown changes. The **Model** dropdown offers four fixed presets, ordered from speed to capability: | Pre