Schematic overview of the autoencoder architecture. Observed spectra x_n are given as input and passed through the encoder into a lower dimensional latent space, which is subsequently decoded into the reconstruction . After training by minimizing the reconstruction error through a gradient descent algorithm, the endmember matrix M is extracted as the weights of the decoder and the abundance vector ω_n is extracted as the latent representation h_n. P is the number of pixels for each spectral order in the observed spectrum. The network is illustrated for R = 3 endmembers representing the solar (orange, top), H₂O (blue, middle) and O₂ (green, bottom) endmembers.