Physics-Aware Enhanced Implicit Neural Representation with Multimodal Feature Fusion for Zero-Shot Super-Resolution of Complex Flow Fields

Abstract

This study addresses the insufficient representation of key regions, such as shock waves and boundary layers, in classical implicit neural representations (INRs) for super-resolution reconstruction of complex aerodynamic flow fields. A physics-aware enhanced implicit neural representation (PA-INR) framework with multimodal feature fusion is proposed. In this study, multimodal features refer to data-driven deep flow features and physics-derived features, including pressure gradients, velocity divergence, and vorticity. The method improves the basic INR in two aspects: local interpolation and feature representation. Specifically, a shock-aware weight correction module is designed to generate corrected weights by combining local features with geometric information, thereby reducing the erroneous fusion of physical states across shock waves. In addition, a gradient-information feature expansion module fuses the data-driven and physics-derived features to enhance the model representation of high-gradient regions. Experimental results show that, on the dataset and evaluation metrics used in this study, PA-INR outperforms the comparison methods in terms of overall reconstruction error and shock-position error. These results indicate that the proposed multimodal feature fusion strategy can improve reconstruction accuracy in complex-flow super-resolution tasks and enhance the recovery of details in shock and high-gradient regions.