A Domain-Adapted Multimodal AI Platform for Substellar Astronomy: Foundation Models, Vision- Language Spectral Analysis, Retrieval-Augmented Generation, and Automated Research Workflows

Obiettivo

SubstellarLLM is the first domain-adapted multimodal AI platform designed to dramatically reduce the “Time-to-Science” for substellar astronomy—from months to minutes. It integrates a fine-tuned text foundation model, a Vision-Language Model (VLM) for automated classification of Euclid’s 2D spectral images and multi-band imaging, Retrieval-Augmented Generation (RAG), real-time public catalogue APIs (Gaia DR3, 2MASS, AllWISE, SIMBAD), and a suite of specialised tools including visual spectral classification, multi-catalogue crossmatching, interactive plotting, and automated manuscript generation from publisher-specific LaTeX templates.

Built on the ERC Substellar project’s breakthrough results—including the first Euclid substellar science paper (Martín et al. 2025, A&A)—SubstellarLLM addresses the critical operational bottleneck facing petabyte-scale survey astronomy: researchers cannot process Euclid’s projected tens of thousands of brown dwarf detections using existing manual workflows. The platform’s key technical moat is its Visual Spectral Reasoning engine, which analyses 16-bit radiometric grism data with instrument-specific artifact awareness—a capability where general-purpose VLMs fail. The underlying scientific principles are proven through the ERC SUBSTELLAR project; this PoC validates the commercial feasibility of deploying them as a scalable product. The 18-month, €150,000 proof-of-concept will harden the inference pipeline, deploy a prototype with eight specialised AI agents, validate time savings of 60–80% with 10–15 consortium researchers, and execute a structured commercialisation plan targeting data archives, space agencies, and research institutions.