S​urrogate Models at Airbus: From Development to Certification

Stephane Grihon (Airbus)

Abstract:

Surrogate Models is not a recent area and can be considered as grounded in applied mathematics and interpolation/approximation theory. Nevertheless it benefited from the progress of deep learning and is now a subfield of Artificial Intelligence. Many applications are running at Airbus especially in Airbus Engineering with various purposes: build models from empirical data, replace heavy simulations by a quicker model, export Airbus simulations to external users to provide services. These extensive applications required to formalize and standardize the process for sake of reuse and proper validation and verification. The topic of the V&V is particularly sensitive because a data-driven model is difficult to validate in a full deterministic way. In line with the growth of international standards (EUROCAE-SAE ARP 6893 anticipated by EASA Concept Paper), specific methods and processes have been introduced at Airbus to handle surrogate model validation.