A Data-Driven Machine Learning Framework for Aerodynamic Analysis of Next-Generation Aircraft Propulsion Systems

Alternative propulsive systems impose new constraints on aircraft configuration, particularly due to the geometric and structural requirements of hydrogen tanks and battery systems. These constraints can drive changes to wing and fuselage shapes, making rapid aerodynamic characterization essential within the aircraft design loop. This paper presents a CFD-to-surrogate workflow for generating aerodynamic coefficient predictions suitable for integration with Argonne’s Aeronomie aircraft energy modeling framework. A targeted CFD campaign was conducted in the open-source software OpenFOAM to generate a database of lift, drag, and pitching-moment coefficients for a set of NACA 4- and 5-series airfoils at representative climb and cruise Reynolds numbers. Two data-driven surrogate approaches were evaluated. A recurrent neural network (RNN) formulation demonstrated strong lift prediction capabilities, but showed reduced robustness for drag and moment, particularly when multiple outputs were learned simultaneously. A geometry-driven multilayer perceptron (MLP) approach, trained separately for each coefficient using NACA-derived parametric descriptors and the angle of attack, provided more consistent predictions across airfoils and operating conditions. The trained MLP models were coupled to a post-processing workflow that generates aerodynamic initialization files for Aeronomie, enabling rapid evaluation of new airfoil profiles in mission-level energy consumption studies. Ongoing work focuses on porting the surrogate models to C# for full Aeronomie integration and expanding the training dataset and operating envelope.