AgriScienceFM — Novelcore

Our role

Work Package Leader — WP3

We lead WP3 and design the modular architectural framework for the three AgriFMs, then build the AgriScienceFM agentic framework: an LLM-based orchestrator that queries the models and knowledge base and draws on our graph-RAG work. We also lead the project's data and risk management, quality assurance and exploitation strategy.

Impact and outcomes

Targeting at least 3 openly released agricultural foundation models, with an open benchmarking suite and documentation
Targeting at least 12 benchmarks across four use cases, at least 3 per use case
Targeting at least 20% improvement over baseline on at least 3 critical yield-related benchmarks
Targeting reuse of the models and tools by at least 10 research teams or institutions
Agentic framework, with user interfaces and APIs for research and farming communities, planned for delivery by month 36

The challenge

Foundation models have transformed weather prediction and protein folding, yet agricultural science has barely benefited. Agriculture is shaped by interacting genetic, environmental and management factors, and its data is small, fragmented and local: even the largest annotated crop-breeding datasets hold under 140K labelled plot cycles. Domain knowledge also stays siloed across disciplines, from genetics to agroecology.

Our approach

AgriScienceFM develops three complementary, self-supervised foundation models: AgriFM-E for environmental drivers, AgriFM-M for management documents and AgriFM-G for plant imaging. They are evaluated on an open benchmarking suite spanning four use cases, from crop and water monitoring from space to pest and disease management. Our architecture defines how the models connect to data, benchmarks and applications, and our agentic framework orchestrates them as callable experts for complex research questions.