Frugal AI for Science

AI in science is moving from scale-only approaches toward a mix of frontier models and efficient, task-specific systems. Foundation models, AI agents, digital twins and autonomous workflows expand capabilities, but must operate under real budget, hardware, energy and data constraints. Frugal AI complements frontier AI with small and domain-specific models, model compression, data-efficient learning, transfer learning, edge AI and open or self-hosted deployment, enabling lower latency, more predictable costs and faster deployment for well-defined scientific tasks. Frugal AI asks whether the same scientific value can be achieved with less data, compute, energy and cost, while preserving reproducibility, control and responsible practice.

• Model compression, quantization, pruning
• Knowledge distillation to compact models
• Light-weight architectures, task-specific AI
• Open-source, self-hosted and sovereign AI
• Data parsimony via curation, active learning, transfer learning and targeted synthetic data
• Foundation and domain-specific scientific models
• Generative AI for discovery and design
• Efficient AI agents and autonomous scientific systems
• Resource-efficient and frugal AI approaches
• Edge, sovereign and distributed AI systems

• Innovation system rewards scale over efficiency, which has high compute and infrastructure requirements
• Sustainability concerns: poorly measured life-cycle compute, energy, storage and data costs
• Dependence on proprietary platforms
• Interoperability gap between data, instruments, models and infrastructures
• Limited access to specialized domain datasets
• Uncertainty of accuracy and bias under new settings
• Skills shortages across disciplines and AI engineering

For scientific workflows, Frugal AI is not just cheaper AI but a design strategy to match model and infrastructure to the research question. Many tasks (classification, scoring, monitoring, image analysis, structured extraction, prediction) can be handled more reliably and reproducibly by curated data and lightweight models than by large general-purpose systems.

Europe has strong scientific communities, research infrastructures, open-science networks and computing facilities. The opportunity is to link these assets into an ecosystem that supports both frontier and efficient, interoperable, sovereign AI. Frugal AI aligns with European priorities on sustainability, digital sovereignty and access for SMEs and public research organisations. Frugal AI can help build useful AI without exclusive dependence on foreign hyperscale platforms.

• Scientific foundation models, physical AI
• Move from scale-only approaches to a mix of Frontier and efficient / task specific systems
• Frugal multimodal and domain-specific models
• Edge, TinyML and offline AI for sensors and instruments
• Federated, privacy-preserving and sovereign AI
• AI agents coordinating efficient modular workflows
• Frugal and sovereign AI
• Data-efficient and sustainable AI, life-cycle based evaluation

The next generation of AI-enabled science will be shaped by the convergence of frontier models, scientific workflows and innovation ecosystems. Europe has an opportunity to develop an approach that combines scientific excellence, industrial competitiveness, openness, sustainability and technological sovereignty, while ensuring that advanced AI capabilities remain sustainable and accessible across the research landscape.