Advances and challenges in aerodynamic optimisation design of fans using machine learning

As essential fluid machinery, fans play a critical role in industrial energy conservation. Traditional design methods, which rely on low-dimensional theories and empirical adjustments, face significant challenges in handling complex three-dimensional flows and high-dimensional parameter spaces, often requiring computationally expensive numerical simulations or experiments. In recent years, machine learning techniques have introduced transformative approaches to aerodynamic optimisation. From the artificial intelligence perspective, methods such as deep neural networks and reinforcement learning have enabled the construction of efficient surrogate models, automated feature extraction from flow fields, and adaptive optimisation strategies. From the engineering application perspective, these machine learning methods have been successfully employed in key areas including airfoil shape optimisation, three-dimensional blade stacking, and flow pattern control. Quantitatively, surrogate models can accelerate flow field predictions by up to three orders of magnitude compared with conventional CFD, and deep reinforcement learning has achieved over 10% improvement in aerodynamic performance while reducing computational cost by approximately 85% relative to traditional genetic algorithms. This review summarises current research progress, highlights synergies between data-driven and physical model-based approaches, and discusses future directions such as multidisciplinary design optimisation and physics-informed machine learning. These advancements provide valuable insights toward intelligent and efficient fan design.