Data-efficient scientific AI: a new paradigm for low-data discovery across scientific disciplines

Scientific progress increasingly depends on our ability to make sense of complex systems using limited and often uncertain data. While recent advances in artificial intelligence (AI) have shown great promise, most models require vast amounts of data to be effective -- an unrealistic expectation in many scientific fields. From climate modelling to nuclear fusion and materials science, researchers routinely face the challenge of generating insights from sparse, noisy, and heterogeneous data sources. This project tackles these challenges directly. We will develop a new data-efficient AI platform tailored to the needs of scientific discovery for multi-scale physical systems. Inspired by recent breakthroughs in AI-based weather forecasting, where models like Aardvark have replaced traditional numerical simulations with fast, data-scarce alternatives, we aim to extend this capability across multiple scientific domains. These models are not just powerful; they are efficient, generalisable, and able to learn from small datasets. Our project will build on these foundations to deliver a versatile, open-source AI toolkit for science. Over the six-month project, our team at The Alan Turing Institute will develop a paradigm and associated open-source platform for multi-scale physical systems that: (i) supports multiple data modalities; (ii) provides generalisation in data-poor regimes; (iii) provides few-shot learning; and (iv) is computational efficient. We will develop a modular open-source code platform that is: (i) straightforward to switch underlying AI model architectures; (ii) flexible so it can support different data modalities; and (iii) easy to adapt and apply to different physical problems. The platform will be accessible to scientists without deep AI expertise, significantly lowering the barrier to entry for using advanced machine learning methods in their research. We will demonstrate the framework through real-world case studies in priority areas such as materials science, nuclear fusion energy generation, and climate science. We will show how the platform can operate in data-poor regimes, generate scientifically meaningful predictions, and quantify uncertainties. These case studies will serve as templates for applying the toolkit to other areas of science. To foster adoption and community engagement, we will organise two workshops: one at the project launch to present plans and gather input, and another at the end to showcase outcomes and spark future collaborations. The final workshop will also serve as a more general AI4Science symposium, discussing he latest advancements in AI for scientific applications, cross-disciplinary collaborations, and future research directions. All code will be open source and fully documented, accompanied by tutorials and user guides to support a wide range of users. Ultimately, this project will deliver a reusable, production-ready AI toolkit that enables scientists to build models in minutes rather than months. It will promote open, reproducible, and inclusive science, while helping position the UK as a leader in AI for scientific discovery. By reimagining how we learn from limited data, our approach offers a powerful new tool for researchers tackling some of today’s most complex and urgent scientific challenges.