Expected Outcome:Project results are expected to contribute to the following expected outcomes.

Environment: the proposed solutions shall have a positive impact on the environment (i.e. in terms of emissions, noise and/or local air quality) and on the aviation environmental footprint e.g., AI will enable the optimisation of aircraft trajectories;Capacity: AI will play a fundamental role in aviation/ATM to address airspace capacity shortages, enabling dynamic configuration of the airspace and allowing dynamic spacing separation between aircraft;Operational efficiency: the proposed solutions are expected to improve the synchronisation and predictability of the ATM system;Cost efficiency: AI will enrich aviation datasets with new types of datasets unlocking air/ground AI-based applications, fostering data-sharing and building up an inclusive AI aviation/ATM partnership;Safety: the proposed solutions are expected to maintain at least the same level of safety as the current ATM system;Security: the proposed solutions are expected to maintain at least the same level of security as the current ATM system. Scope:Tomorrow’s aviation infrastructure will be more data-intensive an... ver más

Expected Outcome:Project results are expected to contribute to the following expected outcomes.

Environment: the proposed solutions shall have a positive impact on the environment (i.e. in terms of emissions, noise and/or local air quality) and on the aviation environmental footprint e.g., AI will enable the optimisation of aircraft trajectories;Capacity: AI will play a fundamental role in aviation/ATM to address airspace capacity shortages, enabling dynamic configuration of the airspace and allowing dynamic spacing separation between aircraft;Operational efficiency: the proposed solutions are expected to improve the synchronisation and predictability of the ATM system;Cost efficiency: AI will enrich aviation datasets with new types of datasets unlocking air/ground AI-based applications, fostering data-sharing and building up an inclusive AI aviation/ATM partnership;Safety: the proposed solutions are expected to maintain at least the same level of safety as the current ATM system;Security: the proposed solutions are expected to maintain at least the same level of security as the current ATM system. Scope:Tomorrow’s aviation infrastructure will be more data-intensive and thanks to the application of Machine Learning (ML), deep learning and big data analytics aviation practitioners will be able to design an ATM system that is smarter and safer, by constantly analysing and learning from the ATM ecosystem. Artificial intelligence (AI) is one of the main enablers to overcome the current limitations in the ATM system. AI is a breakthrough technology that could radically influence or transform the aviation/ATM industry value chain, potentially impacting all stakeholders, including original equipment manufacturers (OEMs) and their business models. The impact of transformative AI will be felt throughout the industry, and beyond. The challenge is to develop potential innovative and breakthrough AI solutions that will help addressing capacity issues in ATM by enabling better use of data, leading to more accurate predictions and more sophisticated tools, increased productivity and enhancing the use of airspace and airports. Considering the extent of these challenges, the proposals shall define and develop potential innovative AI-based solutions that may come up with innovative responses based on non-straightforward correlations of parameters, while improving the scalability, efficiency and resilience of the system.

The SESAR 3 JU has identified the following innovative research elements that could be used to meet the challenge described above and achieve the expected outcomes. The list is not intended to be prescriptive; proposals for work on areas other than those listed below are welcome, provided they include adequate background and justification to ensure clear traceability with the R&I needs set out in the SRIA for the AI for aviation flagship.

AI for higher automation. This element covers the development of an AI-powered infrastructure and services (supporting higher levels of automation). In addition, the aim is to develop automation of ATM processes in which analysis and prediction are particularly likely to benefit from AI, and to develop AI-powered ATM environment requirements, infrastructure, and common regulation and certification guidelines. This may include the research on multi-agent deep reinforcement learning (RL) that has a great potential to enable a highly automated ATM, where functions, roles and tasks are allocated to human and artificial intelligence-based agents at both ground and airborne side based on the strengths and weaknesses of each type of agent. Research shall take into account the impact on the role of the human, responsibility and liability aspects, etc. (R&I need: human–AI collaboration: digital assistants).Exploring underuse AI paradigm in ATM. AI Paradigms (X-axis) are the approaches used by AI researchers to solve specific AI-related problems. Without trying to be exhaustive, a broad classification accounts for: logic-based tools, knowledge-based tools, probabilistic methods, machine learning, embodied intelligence, search and optimization. Latest projects applications have concentrated most of the research efforts on application of ML in ATM, in detriment of exploring the possibilities of what the other paradigms could do for ATM. Research aims at investigating these alternative possibilities (R&I need: human–AI collaboration: digital assistants).Transfer-learning and few-shot learning methodologies in ML ad XAI. Research focuses on transfer-learning and few-shot learning methodologies. In ATM domain, the transfer-learning methodology could be another essential research and development direction for utilizing machine learning and XAI. The lifelong machine can incorporate transfer learning for parameterizing to learn domain-invariant features (e.g., how existing AI models can be used for solving different tasks that share common features or attributes). Transfer-learning can also be used where there are some relations between ATM tasks, such as balancing arrival and departure capacity and take-off delay prediction. Few-shot learning (FSL) is a machine learning framework that enables a pre-trained model to generalize over new categories of data (that the pre-trained model has not seen during training) using only a few labelled samples per class e.g. models for the detection of objects in an image, etc. Research on this element shall consider the output of project ARTIMATION (R&I need: human–AI collaboration: digital assistants).Innovative methodologies for ATM safety, security and resilience. Research aims at developing methodologies (or evolution of existing ones) for safety, security and resilience that will contribute to ensure that ATM is robust against ever-evolving risks, threats and disruptive events in the physical and cyber worlds in an environment with automation levels 4/5. New and disruptive technologies, operations and business models to ensure ATM is resilient against internal and external threats, including health, natural disasters, terrorism and criminal activity. Research shall ensure coordination with EASA (R&I need: trustworthy AI-powered ATM environment).Ensuring the integrity of non-ATM data for AI/ML applications in ATM. For artificial intelligence and machine learning applications in aviation the integrity and quality of input data is critical. The benefits of AI in ATM can only be leveraged if the models are fed with great quantities of good quality data. While existing ATM data present certain homogeneity and is, by design, oriented to ATM uses and analysis, other data sources also needed for the development of AI models in ATM are heterogeneous and not adapted to ATM granularities. One example is meteorological information, which is presented in a variety of sources and formats that are not always of direct use in ATM solutions. There is a need to develop potential solutions to identify erroneous data injected from non-ATM sources that could introduce a safety risk in ATM and how to mitigate it. The research shall address these non-ATM data availability and format, proposing a framework for data curation, sharing and feeding oriented to ATM use cases, as well as developing new indicators at least for data quality and integrity (R&I need: Trustworthy AI-powered ATM environment).Enhancing robustness and reliability of machine learning (ML) applications. Research aims at enhancing machine learning (ML) applications to ensure they are technically robust, accurate and reproducible, and able to deal with and inform about possible failures inaccuracies and errors. Research aims at developing potential solutions to address this challenge, which shall include/refer to the EASA methodologies for certification of AI in aviation. The scope may address: Verification methods of robustness for machine learning (ML) applications. Due to the statistical nature of machine learning applications, they are subject to variability on their output for small variations on their input (that may even be imperceptible by a human). Research aims at proposing new methods to verify the robustness of machine learning applications, as well as to evaluate the completeness of the verification;Standardised methods for evaluation of the operational performance of the machine learning (ML). Research addresses the definition of reference methods and metrics to assess the accuracy or error rate of ML applications;Application of transfer learning and data augmentation techniques for the development of the proposed applications, thus guaranteeing their robustness. In addition, these systems would be continuously validated using ML Ops methodology and explainability techniques, to ensure system performance and detect as early as possible if concept drift is occurring;Identification, detection and mitigation means of bias in ML applications. Machine learning applications are subject to bias, which can compromise the integrity of their outputs. One of the most challenging aspects when collecting, preparing or using data, is the capability to identify, detect and finally mitigate adequately any bias that could have been introduced at any time during the data management and/or of the training processes. Research aims at developing potential solutions to address this challenge (R&I need: trustworthy AI-powered ATM environment). Accelerating AI implementation for ATC automation. AI implementation pace in ATC is far slow compared to other industries. Safety is the principal barrier in the ATC context. Research aims at developing concrete applications that can support the acceleration of AI implementation in Europe. The research seeks for environments where full (or close to full) ATC automation may become a reality in the short term without human supervision. Those scenarios could be very low complex situations like night shifts, where few flights need ATC service are the most suitable, but the research should explore the suitability of more complex scenarios. Research also addresses exploratory activities on solutions non-dependant of human supervision to take back control to solve contingency is necessary. Research may propose ML-based potential solutions to address specific operational use cases, relying on explainability techniques to validate the robustness and performance of the system in all types of situations(R&I need: Trustworthy AI-powered ATM environment).Just culture and AI. Before the introduction of AI/ML into the ATM system, it was difficult but possible to draw the red line between “gross negligence”, “wilful violations” and “destructive acts” on the one side and “honest mistakes” on the other side. State of the art algorithms for AI/ML systems such as neural networks are essentially “black boxes” in terms of explainability. Arguably, the best-known disadvantage of neural networks is their “black box” nature. Simply put, it is unknown how or why the neural network came up with a certain output given a certain input. In other words, they are tremendously successful in providing accurate predictions based on historical data, but no one can understand why. The introduction of AI/ML in essence clouds the drawing of a red line between “gross negligence”, “wilful violations” and “destructive acts” on the one side and “honest mistakes” on the other side. Research aims at redefining just culture and rewrite its procedures in the era of digitalization (R&I need: Trustworthy AI-powered ATM environment).Development of ATM specific ontologies. This research element focuses on special-purpose representation systems (e.g., semantic networks and description logics) that can be devised to help organizing a hierarchy of ATM related categories. There are many variants of semantic networks, but all are capable of representing individual objects, categories of objects, and relations among objects. Knowledge representation through a semantic network will enable ATM-related knowledge to be expressed not only in natural language, but also in a format that can be read and used by software agents; hence, permitting them to find, share and integrate information more easily (R&I need: AI Improved datasets for better airborne operations). ver menos