The Future of Aircraft Performance Engineering Includes Artificial Intelligence

As an aircraft performance engineer that started his career journey 40 years ago, I’ve seen a few changes. From the slide rule to punch cards to spreadsheets – technology keeps advancing and changing how aircraft performance data is developed, presented, and used. What hasn’t changed in response to new technology, however, are the rules used to certify the performance data and, consequently, the differences between certified aircraft performance and actual aircraft performance. Change is inevitable. Although I can't claim to be prescient or have insider knowledge of what regulators may be thinking, it is easy to imagine a scenario descriptive of where we may be heading.

'Artificial intelligence in the form of machine learning can be applied to the data that are available from the operation of every flight conducted.’

When a new airplane type is certified, performance data are generated using a series of flight tests designed to describe fundamental aerodynamic and operational information such as lift, drag, thrust, stall speeds, minimum unstick speeds, buffet margins, etc. These data are then used in basic aero and performance equations to develop the aircraft performance information necessary for operations. The data elements are collected during flight tests, then reduced or normalized for use in an equation. Once validated, the data are then expanded into a form required for certification. The regulatory agencies approve the data as part of the airplane type certificate.

There are rules to be followed during each step of the certification process. Many of the rules require the applicant (the manufacturer of the aircraft) to make a conservative assessment of the data during the normalization process. What does this mean? Effectively, this means that the aircraft is likely to be able to perform better in actual operation than the certified performance data would indicate. Additionally, there are several operational rules that effectively add more conservatism to the certified data when used that ensure the safety of operations to a very high level of confidence. In fact – it is extremely rare for inadequate aircraft performance to be cited as a contributing factor in an airplane accident or incident. But this conservatism was necessary since we didn’t know the variation between individual airplane builds (i.e., the fiftieth airplane may not fly exactly like the test airplane) and pilotage (i.e., average line pilots won’t fly like test pilots). Notice I said was and didn’t?

When an airplane is certified, it is essentially equipped with all the instruments, recording, and transmission devices you can think of to ensure all the necessary data elements are captured. In the past, the test airplane would record several parameters for certification. The ability to measure these parameters would be eliminated with a standard build for customer delivery. Not to worry, though – the airplane still delivered all the necessary information to the flight crew for safe operation and the performance information used was based on a conservative application of flight test data results. Today – modern aircraft, as delivered to the operators, have the ability to transmit much more data than what was once used in the certification. This begs the question – why do we continue to apply conservative factors and adjustments to performance data to cover unknown variations from flight test data when we actually know the variations? This is somewhat of a rhetorical question at this point since there is simply no acknowledged regulatory mechanism in place to allow operators to operate outside the boundaries created by the certified data. This is where I see the opportunity for advancement in the future of aircraft performance engineering.

Artificial intelligence in the form of machine learning can be applied to the data that are available from the operation of every flight conducted. These outcomes could be compared to predicted based on certified data and adjustments to the performance algorithms could be made such that (1) the variation between actual and certified performance is narrowed and (2) the level of confidence for meeting the required level of performance is a known entity. This would allow for a more meaningful operational application of aircraft performance. We just need a regulatory path (and a cadre of aircraft performance engineers proficient in data science) to make it happen.