Generating high-fidelity hyperspectral data is an essential component of observation missions but traditionally involves expensive numerical simulations. This challenge can be addressed by constructing a statistical model approximating the distribution of simulated data. This process is referred to as emulation. This paper introduces a novel hyperspectral emulation approach pairing a Variational Autoencoder and a Neural Interpolator for fast hyperspectral image (HSI) emulation. The VAE is trained to reproduce data, and an interpolator is then used to link biophysical parameters to the latent space learned by the VAE. This method enables the sampling of HSI from biophysical parameters efficiently and accurately. This approach is evaluated on Sentinel-3 Ocean Colour Images and shows significant improvements in accuracy, efficiency, and scene-level information synthesis compared to existing methods.