TL;DR

The Alan Turing Institute has secured £375,000 from UKRI’s Engineering and Physical Sciences Research Council to develop AI techniques that can deliver accurate predictions from limited or uncertain data. The six-month project aims to create open-source tools that enable scientists to build AI models in minutes rather than months.

Addressing AI’s Data Hunger

Most AI systems require vast quantities of high-quality training data to function effectively. This creates significant barriers for scientific applications where data may be scarce, noisy, or fragmented. The new Turing Institute project directly addresses this challenge by developing models purposefully designed to learn from imperfect datasets.

Crucially, the models will also be able to quantify uncertainty—allowing researchers to understand how confident an AI system is in its own predictions. This capability proves particularly important when AI informs scientific discovery, experimental design, or engineering decisions that may be costly or safety-critical.

Open-Source Toolkit in Development

Over the six-month project, researchers will create open-source techniques for multi-scale physical systems that support multiple data modalities whilst remaining computationally efficient. The toolkit will allow scientists to switch between AI model architectures, adapt to different data types, and apply the system across a wide range of physical problems.

“By focusing on how we learn from limited data and enabling AI models to declare the confidence in their predictions, we are hoping to create ready-to-use AI tools that will help scientists build models in minutes rather than months,” says Professor Jason McEwen, Mission Director for Fundamental Research in AI for Physical Systems at the Alan Turing Institute.

Accessibility by Design

The techniques are being designed with accessibility in mind, enabling scientists with limited AI expertise to use advanced machine learning models in their research. Real-world case studies will demonstrate the techniques and provide templates that can be adapted across other scientific domains.

Looking Forward

This research addresses a fundamental limitation in current AI approaches and could significantly accelerate scientific progress in data-poor priority areas. For UK research institutions and businesses working with limited datasets, the resulting open-source toolkit may provide a practical path to AI-enabled insights.

Source: The Alan Turing Institute