Session: 34-13 High-fidelity CFD for Compressors II

Submission Number: 178108

Aeroacoustic Optimization of a Fan Stage Using GPU-Accelerated Large-Eddy Simulations and Deep Learning 

Propulsion noise is one of the major noise sources of aircraft, and enormous efforts have been made to reduce the noise level to enhance the in-cabin environment, improve passenger experience, and reduce the negative impacts on the surrounding communities of airports. The optimization of turbomachinery aeroacoustics is particularly challenging due to the difficulties in simulating the complicated flow and acoustic field and the lack of effective optimization algorithms for aeroacoustics. The current design and optimization tools in industry mostly rely on Reynolds-averaged Navier-Stokes equations and thus have inherent difficulties in tackling engine noise problems. On the other hand, most flow solvers using high-fidelity numerical approaches, such as direct numerical simulation and large-eddy simulation, can hardly enter the industrial design process due to either prohibitively high computational cost or inability of handling complex engineering geometries. 

In the present work, an optimization framework for turbomachinery noise reduction based on wall-modeled large-eddy simulation (WMLES) and deep learning is developed. Leveraging the recent development in numerical algorithms and computer hardware, particularly the graphic processing unit (GPU), the flow solver Fidelity CharLES has evolved to the level that an efficient and affordable solution can be provided for turbomachinery simulations. It has been demonstrated that using the GPU-accelerated Voronoi-diagram based moving mesh solver, a full-wheel simulation of a transonic fan stage can be completed in hours, in contrast to weeks using conventional CPU solvers (Wang et al. J. Turbomachinery, 2025). At the same time, the rapid development in deep learning has rendered artificial neural network an efficient and powerful tool for numerical optimization. Utilizing these technical advancements, an affordable aeroacoustic optimization tool using high-fidelity simulation approach is developed to fulfill the industrial demand. 

The optimization study is performed on the NASA Fan Broadband Source Diagnostic Test (SDT) model. The model consists of a fan with 22 blades and a row of outlet guide vane (OGV) which has three configurations: baseline, low count, and low noise. The aerodynamic performance of these configurations was tested and the sound radiation was measured at NASA Glenn Research Center. Using the GPU-accelerated moving mesh solver, WMLES calculations were conducted for the three OGV configurations at the approach, cut-back and take-off conditions (Brès et al, AIAA-2023-4299, AIAA-2024-3162). The predicted pressure power level (PWL) agreed well with the experimental measurements. The study also assessed the solver performance and reported that for a mesh with 142 million control volumes, 10 full rotations, over which the acoustics data were collected, can be simulated in approximately 6 hours on a moderate number of GPUs. 

Based on these previous studies, the developed optimization method for aeroacoustics is applied to the low noise OGV at the approach condition. The geometry of OGV is perturbed from the original design by exploring the parametric space spanned by the blade stagger angle and sweep angle. A series of WMLES is performed and the acoustics at far field are computed using the permeable formulation of the Ffowcs-Williams-Hawkings equation. The process, starting from geometry perturbation, mesh generation, to flow simulation and acoustic post processing, is automated in the Fidelity CharLES tools. The acoustic spectra are used to train a neural network-based surrogate model. The optimization is converged to NASA’s optimal OGV design (low noise), and the identified optimum is within robust design space for both the fan stage efficiency​ and noise level. With modern GPUs, it only takes approximately 4000 GPU-hours to complete the entire optimization process, demonstrating the efficiency of the approach. The database is used for flow and acoustic analysis to help understand the noise reduction mechanism. The analysis and comparison will be presented in the full paper. 

Presenting Author: Christopher Ivey Cadence Design Systems, Inc.

Presenting Author Biography: Christopher Ivey is a Distinguished Engineer at Cadence Design Systems. He received his PhD in Mechanical Engineering from Stanford University. He specializes in numerical methods, mesh generation and GPU acceleration. He is key developer of Fidelity charLES and stitch.

Authors:

Kan Wang Cadence Design Systems, Inc.
Christopher Ivey Cadence Design Systems, Inc.
Liam Heidt Cadence Design Systems, Inc.
Guillaume Brès Cadence Design Systems, Inc.
Sanjeeb Bose Cadence Design Systems, Inc.