Advancing probabilistic machine learning to deliver safer, more efficient, and predictable air traffic control
Project Bluebird is a partnership between NATS and The Alan Turing Institute, supported through an investment from EPSRC. The research vision is to deliver the world’s first artificial intelligence (AI) system to control a section of airspace in live trials, working with air traffic controllers to help manage the complexities of their role. This system will use digital twinning and machine learning technologies, and will include tools and methods that promote safe and trustworthy use of AI.
Air traffic control (ATC) is a remarkably complex task. In the UK alone, air traffic controllers handle as many as 8,000 planes per day, issuing instructions to keep aircraft safely separated. Although the aviation industry has been hit by the pandemic, European air traffic is forecast to return to pre-pandemic levels within five years. In the long term, rising passenger numbers and the proliferation of uncrewed aircraft will mean that UK airspace is busier than ever, so next-generation ATC systems are needed to choreograph plane movements as efficiently as possible, keeping our skies safe while reducing fuel burn.
The project has three main research themes:
Develop a probabilistic digital twin of UK airspace. This real-time, physics-based computer model will predict future flight trajectories and their likelihoods – essential information for decision-making. It will be trained on a NATS dataset of at least 10 million flight records, and will take into account the many uncertainties in ATC, such as weather, or aircraft performance.Build a machine learning system that collaborates with humans to control UK airspace. Unlike current human-centric approaches, this system will simultaneously focus on both the immediate, high-risk detection of potential aircraft conflicts, and the lower risk strategic planning of the entire airspace, thus increasing the efficiency of ATC decision-making. To achieve this, researchers will develop algorithms that use the latest machine learning techniques, such as reinforcement learning, to optimise aircraft paths.Design methods and tools that promote safe, explainable and trustworthy use of AI in air traffic control systems. This will involve experiments with controllers to understand how they make decisions, so that these behaviours can be taught to AI systems. The project will also explore ethical questions such as where the responsibility lies if a human-AI system makes a mistake, how to build a system that is trusted by humans, and how to balance the need for both safety and efficiency.
Underpinning all of these themes is a drive to create a vibrant AI research community in this domain, enabled by knowledge sharing, a framework of peer review, and two-way communication with ATC researchers.
The project’s overarching objective is to deliver the first AI system to work with air traffic controllers and control a section of airspace in live trials, which will put the UK at the forefront of technical advances in this sector. More broadly, research of AI technologies in ATC will be a catalyst to scientific discovery providing ATC an insight into new and innovative ways to modernise UK airspace, how to increase its efficiency, and help the UK aviation industry achieve net zero carbon emissions by 2050.
Applications
The methods and theory within the project have direct applications to other areas of science, including:
Digital twins: these are transforming many aspects of engineering, enabling simulation and data-led decision-making in advanced engineering systems.Computational statistics and uncertainty quantification: new methods in advanced computational statistics for calibration and assimilation of high-dimensional uncertain systems will provide direct application across many scientific areas that use uncertainty quantification.Machine learning control and optimisation: theoretical work will explore issues around robustness in reinforcement learning and evolutionary optimisation problems, alongside new results in hierarchical multi-agent systems. These outputs will provide further results in nonlinear function approximation (e.g. deep neural networks) and optimisation methods in stochastic environments.Scientific and high-performance computing: the real-time nature of ATC requires extremely fast simulations and sampling techniques of a large complex system. This will necessitate algorithmic approaches for modern high-performance computing hardware.Computational ethics, society & AI: the project will provide new results on how social science studies can be technically embedded within new AI frameworks, and how new approaches can be developed to improve human-AI collaboration through explainable AI methods.Legalisation of AI control: the project will provide important data on the distribution of responsibility within a human-AI system.
Opportunities to join the team
We will be recruiting for additional team members as the project progresses, and will link to the opportunities below as they arise. These two roles are the first to be recruited and will work closely with the PI and partners in order to deliver this exciting cross disciplinary research partnership:
Project leads
Turing / university partner leads
Professor Tim Dodwell (PI, Turing/Exeter)Professor Mark Girolami (Co-I, Turing/Cambridge)Professor Richard Everson (Co-I, Turing/Exeter)Dr Adrian Weller (Co-I, Turing/Cambridge)Dr Edmond Awad (Co-I, Exeter)Dr Evelina Gabasova (Turing, Co-I)
NATS leads: