A Universe To Be Decided: Democratising Science for Everyone (AI LLM LINCC Tech Talk)

This LINCC Tech Talk, originally aired 2024 July 11, features Jo Ciuca of Australia National University. This Tech Talk was hosted by Colin Orion Chandler (DiRAC Institute, University of Washington) and Mi Dai (University of Pittsburgh), both Project Scientists of the LSST Interdisciplinary Network for Collaboration and Computing (LINCC) Frameworks. These Tech Talks are aired the 2nd Thursday of each month at 10:00 AM Pacific time. The Title and Abstract, provided by the speaker, are below:
A Universe To Be Decided: Democratising Science for Everyone
I discuss the application of Foundation Models in Science through the collaborative efforts of the UniverseTBD consortium, with a mission to Democratise Science for Everyone. I first present HypoGen, a novel methodology for generating, evaluating, and refining scientific hypotheses using Large Language Models (LLMs). This project includes creating a shared task and dataset release to foster LLM development for hypothesis formulation in Computer Science, Astronomy and beyond. We further demonstrate the capability of LLMs in informed scientific critique and hypothesis generation, as validated by human experts. Secondly, I introduce AstroPT, a specialised foundation model trained on astronomical imagery from the DESI Legacy Survey. AstroPT demonstrates the potential of scaling laws in astronomical data, similar to those observed in natural language processing. Finally, I discuss our ongoing work on AstroLLaVA, a multi-modal model allowing interaction with astronomical images via natural language. Through UniverseTBD, we aim to inspire and explore how artificial intelligence can assist human intelligence in scientific discovery, particularly in Astronomy, but with broader implications for Science.