Integrating 'Trustworthy AI' Principles into Machine Learning for Aviation Weather Forecasting

Weather forecasting for aviation allows tens of thousands of flights to operate daily with prior warning of global hazards including in-flight icing, turbulence, convection, and fog. Machine learning (ML) methods have begun to be utilised across the aviation weather forecasting sector and can provide greater skill, lower false alarm rates, and cheaper running costs than conventional equivalent products. Often these products are built on existing numerical weather prediction techniques, but can also be standalone products that make predictions only based on observations. Aviation is a highly regulated and safety-critical industry, so weather forecasting products must meet stringent quality-control standards, and machine learning processes must be trusted by customers.

The Aviation Applications Team at the Met Office has developed a set of 'Trustworthy AI' principles that ML products must strive to adhere to. These principles have guided the recent development of a range of ML driven weather forecasting solutions for aviation including convective cloud detection at UK airfields, auto-TAF (Terminal Aerodrome Forecast) verification, and global convective forecasting capability. In each of these use cases the aviation sector end-users have been considered to ensure the products are trustworthy and explainable.

This study showcases a range of aviation weather forecasting case studies and how they have utilised trustworthy AI techniques including,  explainable AI (XAI), representative AI, and considered how existing 'research to operations' pipelines can be exploited to add trust to machine learning models. The research group has worked with the UK's aviation regulator, the Civil Aviation Authority, to consider what the industry requires to be able to use machine learning safely in UK aviation operations and what can be learned from the long-standing collaboration between the two organisations in developing trusted weather forecasting products.

Future challenges in operationalising ML driven weather forecasting products in the aviation sector include; sparsity of observations for some hazards, shifting baselines for long-term deployment of products, and regulatory hurdles for the approval of AI products.