BUILDING FRONTIER INTELLIGENCE USING FOUNDATION MODELS FOR SCIENCE
Machine learning and data science are being transformed by the advent of the foundation model for physical phenomena. These base models capture core data and compress it into a ‘packaged’ format called the embedding space. This streamlines data science pipelines immensely, allowing months of work to be done in weeks or days.
However, foundation models also offer four additional benefits:
Combinability: Models can be combined across the embedding space, unlocking new opportunities for insight.
Queryability: Models can be combined with text to create queryable systems in natural language.
Agency: This allows "agentic" systems to run on top of them, unlocking situational geospatial intelligence for crisis response, autonomy, and new science.
Efficiency: Models are far more efficient computationally; some can run on a modern laptop.
FDL’s legacy in foundation models
FDL has been a leader in developing foundation models for the physical sciences. Our contributions include:
The first demonstration of generalization for SAR (in Earth Science)
The first Foundation Model for the Sun
One of the first queryable Earth tools: GeoQuery*
An 'AI scientist' that orchestrates* foundation models and classical methods to test hypotheses, working as a trusted colleague.
*The latter three models are ‘agentic’ in that science data can be queried in natural language.
We are proposing FDL Foundations, a new FDL format working in partnership with NASA, ESA, and other FDL partners to support the application of AI foundation models to use-cases in Africa.