Building an AI in ATM we can all trust

Artificial Intelligence (AI), a concept capable of eliciting both great interest, and distrust in equal measure. Hailed for the potential it holds to carry out complex calculations in record times, yet feared for the opacity of its inner workings, AI has many industry experts working to understand how it can best be utilised, and to what extent.

Air Traffic Management (ATM), is of course, no exception. AI could significantly support and enhance several critical decision making tasks, leveraging the large quantities of data and combinatorial complexity. We have already exploited AI in order to reduce delays, minimise taxi times and favour continuous descent operations, thereby reducing fuel consumption and CO₂ and Non-CO₂ (e.g. contrails) emissions. However, ongoing worries about AI’s explainability continue to restrain the extent to which it can be integrated into an industry whose primary concern is safety. We work with the European Union Aviation Safety Agency (EASA) to be compliant to the new norms¹ for Machine Learning in Aeronautics, and to the EU AI Act². 

Thales has capitalised on its extensive ATM experience to work towards an AI in ATM we can all trust. Combining decades of experience in the ATM industry and cortAIx, our trusted AI accelerator for research, sensors, and systems, our teams are ideally positioned to leverage the benefits in order to get the most out of AI while continuously working to tackle remaining challenges.

Understanding the potential of AI for ATM

AI has the potential to drastically change the way humans and machines collaborate. New forms of AI, that are currently being developed and implemented, including in the ATM industry, differ to traditional AI, which is a more automated process. This AI does not learn, it simply performs calculations very quickly within a very specific set of parameters, speeding-up decision-making and reducing operators’ cognitive load. We have been using this form of AI for a long time, in various processes such as traffic flow optimisation and sequencing, trajectory prediction, and Conflict Detection and Resolution (CDR).

New types of AI are instead starting to include Machine Learning (ML), Reinforcement Learning (RL) and Hybrid AI (mixing knowledge and data).

As its name indicates, ML allows AI-powered machines to learn from the data they are being fed or harvesting in order to self-programme and improve, whilst RL enables the exploration of highly complex combinatorial problems and propose actions.

To harness its benefits

In a sector like ATM, where an operator must be able to make decisions based on numerous sets of data – e.g., flight information, atmospheric conditions, environment, etc – AI’s ability to contain, recall and combine more data than a human, in a shorter amount of time is a significant asset.

AI algorithms are already in use to improve Air Traffic Control (ATC). These have been used to facilitate pattern recognition and trajectory predictions, and to enhance ATC optimisation. The latter requires that AI algorithms be fed vast quantities of data – such as flight plans, aircraft information, environmental conditions, infrastructure, Estimated Time of Arrival (ETA), etc – in order to train the algorithm to make quick and informed decisions.

AI presents a fantastic opportunity to improve detection, alerting and resolution of situations, potentially influencing a flight’s safety. AI training for flight optimisation is intrinsically linked to CDR. As such, our SkyLab France team and its customers have also been working on training its ATC solution to suggest relevant, efficient and safe solutions to potential conflicts, based on Reinforcement learning . Compared to simpler AI models which run through all potential solutions and select the best one, this new form of self-training AI functions more like a human: based on all the data available, it is capable of predicting which options are preferable in order to make suggestions to the Air Traffic Controller.

This approach has proven to be far more accurate. Thales is also exploring a Hybrid Approach, which combines both classical and RL algorithms. Furthermore, to answer the EASA questions about RL algorithm qualification, we are collaborating with TU Delft to develop methodologies and tools for qualifying AI based on Reinforcement Learning in Air Traffic Management.

And continue tackling remaining challenges.

While there is no doubt that AI can significantly contribute to the ATM industry, one important challenge relating to AI based on ML subsists: the opacity of the decision-making process.

Complex calculations carried out by a simple AI algorithm or a human resulting in an error can be relatively easy to check and consequently, explain their performances. The same cannot be done for an AI-enabled machine because it provides a solution from multiple matrices, all computed and evaluated continuously. Imagine trying to identify a mistake in the midst of thousands of lines of calculations across hundreds of chalkboards… impossible!

Yet in an industry where safety is the primary objective, every single system contributing to that safety needs to be certified. But if it is humanly impossible to identify a potential mistake, then how can an AI-enabled machine be certified?

 “The only alternative is trust.”

I am convinced that working with several industry partners is key when looking to build trust in how AI is used. This entails understanding mathematical properties of AI algorithms to ensure they are operated under controlled conditions. Thus ensuring the quality and representativeness of training data, allowing us to observe the behaviour of AI algorithms in operations, and consequently detect and fix deviations. We will valorise AI Engineering methodologies and tools developed in the CONFIANCE.AI program and assets from cortAIx Labs. 

Using this approach, we will collectively develop a process to quantify – and therefore certify – the level of trust one can have in an AI-enabled machine. As the systems evolve, they will then be able to build on these developments to achieve progressively more complex types of AI.

In other words, we are working to not only bring efficiency to the ATM world, but also to build a safer AI-enabled eco-system, together.

¹ ED-324 / ARP6983: AI/ML standard in aviation

² https://www.europarl.europa.eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligence