03.4 - Applied Aerodynamics
MULTI-SCALE LINEAR FEATURE MODULATION DEEP NEURAL OPERATOR FOR RAPID PRESSURE COEFFICIENT PREDICTION
W.-F. Zhang¹, H. Zhao¹, S.-K. Wang¹; ¹Northwestern Polytechnical University, China
Strong nonlinear phenomena in transonic flow fields, such as shock wave-boundary layer interaction, pose a major challenge to aerodynamic shape optimization, introducing a fundamental trade-off between computational efficiency and prediction accuracy. While Computational Fluid Dynamics (CFD) methods provide high-fidelity results, they are computationally expensive, with a single simulation typically taking hours or days. In contrast, existing data-driven surrogate models, though fast, lack sufficient accuracy and sample efficiency, failing to capture the inherent multi-scale and high-frequency features of transonic flows. To address this challenge, this study proposes a novel Multi-Scale Feature-wise Linear Modulation Deep Neural Operator (MS-FiLM DON), aiming to establish a high-precision, end-to-end mapping from high-dimensional geometry parameters to wing surface pressure coefficient distribution. The core innovations lie in three synergistic designs. First, a multi-scale deep trunk network is constructed to conduct multi-level scale transformations and feature extraction on input spatial coordinates. It adopts parallel coarse-scale and fine-scale pathways to capture both large-scale macroscopic pressure gradients and subtle local flow features, enabling hierarchical and efficient representation of multi-scale physical phenomena in transonic flows. By providing multiple scale channels, the network autonomously selects and adapts to multi-frequency information, mitigating spectral bias. Second, an adaptive Feature-wise Linear Modulation layer is introduced. This mechanism parses global geometric deformation features embedded in input parameters from the branch network via a linear feature modulation layer, generating two sets of modulation parameters to adaptively linearly modulate the trunk network outputs. This allows the model to dynamically enhance sensitivity to steep gradient regions such as shock and separation zones based on specific wing geometries, achie
DOI: 10.71945/icas2026_0473