Conventional Fourier modal methods (FMMs) remain computationally intensive for simulating and inversely designing free-form meta-atoms. This work introduces a convolutional neural network (CNN)-based method for generating normal vector (NV) fields to improve FMMs’ convergence. Our proposed models predict NV fields from free-form shapes with high accuracy and reduced computational demand. Theoretical analysis and experimental validation using the rigorous diffraction interface theory (R-DIT) demonstrate equivalent accuracy with significant speed improvement, particularly up to 3x speedup on GPU platforms, utilizing only minimal convolutional layers. This integration of machine learning with traditional simulation techniques illustrates the potential for pioneering advanced tools in photonics de