Atmospheric Radiative Transfer Models (RTMs) are key physical tools used for simulating the propagation of electromagnetic radiation through the atmosphere. While accurate, they are computationally expensive, limiting their use in large-scale or real-time applications such as hyperspectral atmospheric correction. To address these limitations, Gaussian Process Regression (GPR) offers strong predictive performance, but it suffers from poor scalability due to its cubic computational complexity. In this study, we investigate neural network-based alternatives to GPRs, focusing on a conditional Variational Autoencoder (cVAE) architecture. Trained on a high-resolution MODTRAN6-simulated dataset of radiative transfer functions, the cVAE models each function separately in a PCA-reduced latent space enabling efficient and compact representations. Our experiments show that neural networks emulators achieve accuracy comparable to GPRs while being better suited for large datasets. These results highlight the potential of physics-informed neural network emulators as scalable, accurate, and uncertainty-aware surrogate models for RTMs, making them well suited in operational atmospheric correction workflows.