INVERSE AERODYNAMIC DESIGN OF ROTORCRAFT AIRFRAMES VIA CFD-TRAINED MACHINE LEARNING

Abstract

This thesis introduces a workflow that enables airframe aerodynamic design to be integrated early into the rotorcraft design process by leveraging mid-fidelity Computational Fluid Dynamics (CFD) and Machine Learning (ML). Traditional rotorcraft conceptual design often relies on isolated airframe analyses that ignore important aerodynamic interactions between the main rotor (MR), propeller (PROP), and airframe. It is not until later in the design process that these nuanced interactions are studied more closely, often resulting in aircraft aerodynamic performance trending that can deviate from initial conceptual design expectations and/or missed opportunities to capture incremental, yet significant, system-level efficiencies. To address this, a robust parametrized geometry generation tool based on ROBIN elliptical expressions was developed. This tool was then exercised using an 180-case Design of Experiments (DOE). The geometry outputs were then integrated into STAR-CCM+ steady state Reynolds Averaged Navier Stokes (RANS) simulations. Ambient conditions were flight at 180kts Sea Level Standard, 0° pitch att, and 22k lbf gross weight. Virtual disks were used for the MR and PROP. The resulting CFD results were then used to train various ML regression models, out of which a Quadratic Support Vector Machine (SVM) regression model was selected for its accuracy as dictated by the Root Mean Square Error (RMSE) results. This trained model enabled rapid evaluation of approximately 906 million airframe geometry perturbations. The optimization process successfully reduced the required shaft horsepower (SHP) from a ROBIN baseline of 2,282 SHP to 2,113 SHP – a 7.4% improvement. Analysis of the optimized and baseline geometries shows that efficiency is maximized through a bluff cockpit design that ensures the boundary layer (BL) has high streamwise (inertial x-direction) momentum when it reaches the tailcone. This flow conditioning allows the boundary layer to overcome the adverse pressure gradient encountered at the aft fuselage, without triggering flow separation and associated pressure drag. The tailcone design that works synergistically with this bluff cockpit has a low boattail angle, minimizing the magnitude of adverse pressure gradients at the aft fuselage. From an interactional aerodynamics perspective, avoidance of pressure drag for this pusher propeller aircraft configuration is key. Since pressure drag is proportional to dynamic pressure, any separation yields especially large drag impacts as the propeller induction effect speeds up the flow in that region. Conversely, while aft flow separation is highly penalized, some shedding of low dynamic pressure structures is favorable as velocity deficits in the propeller plane translate to higher propeller efficiency (Eta). Main rotor and propeller sizing results highlight the importance of selecting radius, solidity, and rotational velocity, and loading such that the disks operate at the condition for best system efficiency, which doesn’t necessarily translate to best main rotor and/or propeller efficiency. Ultimately, this study shows that DOE facilitates large initial performance gains, and coupling it with machine learning allows designs to approach the aerodynamic optimum within the modeled physics– the so called “converged design space”.