- Horizon Europe topic
- HORIZON-CL4-INDUSTRY-2025-01-DIGITAL-61
- Call
- HORIZON-CL4-2025-01
Horizon Europe • GenAI4EU (RIA)
AI Foundation models in science
Scientific foundation models are an emerging paradigm that extends generative AI toward domain-grounded research infrastructure. The GenAI4EU strand under Horizon Europe aims to unlock that potential and advance AI capabilities shaped by the needs of scientific discovery — not only text generation, but models that researchers can adapt, trust, and reuse.
SciFM
What is a scientific foundation model?
A shared base model that researchers adapt — not a single end-user application.
A foundation model ingests information across multiple data modalities and can be adapted to many specialised downstream tasks. It is typically fine-tuned with additional training and task-specific examples. The model itself is not a finished product: it is a shared starting point from which task-specific variants are built through adaptation.
In scientific fields, such models can be pre-trained on domain corpora — literature, structures, simulations, sensor streams — and then fine-tuned for concrete research problems. When released openly, they widen access to advanced AI tooling across a discipline.
GenAI4EU
Four scientific domains
Each funded project addresses one domain; proposals develop, adapt, and demonstrate foundation models within it.
Domain A · SimuLingua
Materials science
Innovative materials underpin EU economic security and a competitive, sustainable industry — from energy and mobility to construction, health, and electronics. AI foundation models can accelerate design, characterisation, and discovery at scale.
Consortium objectives & work packages →Domain B
Climate change science
Stronger climate research supports EU neutrality and resilience goals. Foundation models can improve insight into climate dynamics, extreme-weather prediction, regional impacts, and tipping-point behaviour.
Domain C
Environmental pollution sciences
Environmental AI can help detect pollution sources, trace pathways, and assess distribution and impacts on ecosystems and human health — including emerging and poorly characterised pollutants.
Domain D
Agricultural sciences
Agricultural foundation models can strengthen crop, livestock, soil, and water management — supporting a competitive, resilient, and sustainable food system.
Topic requirements
What proposals should deliver
Three interconnected goals from the Horizon Europe topic text.
Focus 1
Build the foundation model
Develop domain-specific foundation models for science — not limited to generative AI — using architectures and learning methods suited to the chosen field.
Focus 2
Prove usefulness through adaptation
Demonstrate value by adapting the model to subtasks and real scientific problems within the domain, with clear evaluation and benchmarking where appropriate.
Focus 3
Show broader applicability
Illustrate additional areas where the model and its adaptations could serve researchers, including at least four use cases and scientific challenges.
Governance
Open, trustworthy, multidisciplinary
Cross-cutting expectations for data, compute, governance, and community access.
- 1.Release open models to the scientific community — source code and, where possible, training data and assets for full reusability.
- 2.Document use, limitations, and responsible deployment; include case studies across tasks in the domain.
- 3.Combine AI expertise with domain scientists; provide interfaces usable by researchers without a computer-science background.
- 4.Curate high-quality, preferably multimodal, FAIR datasets with clear quality-control and provenance procedures.
- 5.Contribute to common standards for formats, metadata, taxonomies, and ontologies.
- 6.Plan transparent architectures, computational access for training and inference, and strategies to integrate domain knowledge (e.g. knowledge graphs, machine-readable representations).
- 7.Assess misuse risks and propose maintenance, evolution, and community promotion of the model over time.
- 8.Involve Social Sciences and Humanities expertise where legal, ethical, and privacy questions arise.
Materials science
How SimuLingua fits
Project positioning within the GenAI4EU materials-science strand.
SimuLingua is funded under this GenAI4EU topic in materials science (domain A). The project delivers an open, multimodal Scientific Foundation Model (SciFM) for materials that unifies text, structures, images and simulation data in a closed design–simulate–validate loop. A physics-in-the-loop generative engine and AI-accelerated multi-scale simulation pipeline (DFT → phase-field → CFD/FEM) verify candidates; results flow back through a FAIR knowledge graph with ontologies, provenance and leakage-safe splits. Natural-language interfaces let domain experts query the KG, launch virtual experiments and steer inverse design — progressing from TRL 1 to TRL 4 in 48 months across six use cases.
Further reading
Wider landscape & references
Initiatives and references related to scientific foundation models — external sources linked for context.
CORDIS programme page
Official Horizon Europe topic description for AI Foundation models in science (GenAI4EU).
Open resource →Stanford CRFM — foundation models
Foundational survey on opportunities and risks of foundation models (term origin).
Open resource →Atomistic materials chemistry FM
Example foundation-model direction in materials science (arxiv preprint).
Open resource →Helmholtz Foundation Model Initiative
Large-scale German research initiative for trustworthy scientific foundation models.
Open resource →Trillion Parameter Consortium
International cooperation on reliable generative AI models for science and engineering.
Open resource →Scientific Foundation Models (SciFM)
University of Michigan initiative on scientific foundation models across domains.
Open resource →DOME — ML evaluation in science
Community framework for documenting and benchmarking models in scientific use.
Open resource →