Deep learning-enhanced aerodynamics design of high-load compressor cascade at low Reynolds numbers

Abstract

This study addresses the challenges of designing high-efficiency compressors for long-endurance, high-altitude unmanned aerial vehicles (UAVs) under low Reynolds number (Re) and high load conditions, exploring the aerodynamic performance limits of compressor blades under extreme conditions. The research integrates advanced numerical simulations, experimental methods, and deep learning technologies to minimize profile losses and deviation angle on compressor blades. An orthogonal experimental design systematically explored the impact of key geometric factors on aerodynamic performance. A deep learning model incorporating neural networks with spatial attention mechanisms was developed to significantly enhance the accuracy of aerodynamic predictions. This model adeptly captured the complex nonlinear interactions between aerodynamic and geometric parameters. Sobol sensitivity analysis revealed that the dimensionless maximum thickness is the most critical factor, accounting for 44 % of the total variance in total pressure loss and deviation angle. The position of maximum thickness and the aspect ratio of the elliptical leading edge also significantly influenced performance. The optimized high-load compressor blade profile was validated through experimental data and detailed computational fluid dynamics analysis using large eddy simulation methods. This analysis revealed flow separation and reattachment mechanisms, shedding light on turbulence and vortex dynamics critical to performance. This research deepens the theoretical understanding of compressor cascade fluid dynamics and provides practical insights for designing more efficient compressors, especially for micro axial compressors in high-altitude UAVs.