Conceptual Design of Green Propulsive Systems Using Reinforcement Learning

Abstract

Hybrid-electric powertrains offer a solution to significantly reduce aircraft emissions in flight. This study presents a method for automatically generating hybrid-electric architectures and optimizes their control parameters to maximize payload using reinforcement learning. Applied to an ATR 72-600 reference aircraft, with the Flightpath 2050 sustainability goals as constraints, the framework indentifies an optimal architecture: a gas turbine combusting conventional jet fuel and hydrogen powers the primary propulsive line, while fuel cells deliver the majority of the power to an auxiliary propulsive line. Compared with a conventional architecture, this design reduces CO2 and NOx emissions by up to 74% and 86%, respectively, incurring a payload mass penalty of only 24%.