Abstract: To achieve flow-field-based prediction of airfoil aerodynamic noise, this study investigates a straight NACA0018 wing equipped with artificial flexible serrated coverts, a bio-inspired trailing-edge treatment intended to attenuate aerodynamic noise. The wake velocity field near the trailing edge and the far-field noise were measured in wind-tunnel experiments using hot-wire anemometry, particle image velocimetry (PIV), and far-field microphones. Because trailing-edge noise is governed by both the near-wake turbulence structure and the global flow condition, features from complementary measurement techniques were combined to provide a more complete physical description of the noise-generation mechanism. Using case features (freestream velocity, covert thickness, and mounting position) together with characteristic wake velocity-field features as inputs, two deep neural network architectures, an early-fusion neural network (EFNN) and a late-fusion neural network (LFNN), were developed and compared in terms of their far-field noise prediction accuracy. The prediction target was the one-third-octave-band sound pressure level. Without a physical constraint, both architectures reproduced the overall trend of the noise spectra well, with EFNN achieving slightly higher accuracy than LFNN. After introducing an overall sound pressure level (OASPL) physical-constraint loss term, LFNN improved the OASPL prediction accuracy while retaining higher one-third-octave-band sound pressure level accuracy than EFNN. Furthermore, a SHAP-based feature-attribution analysis showed that the flow-field-to-noise mappings learned by the two architectures were largely consistent, with the Reynolds number identified as the dominant predictive feature, confirming the physical plausibility of the model predictions and pointing to a data-driven route for rapid aerodynamic noise estimation.