The ESO Science Archive offers millions of data files to its science community. It is, then, crucial to its success that it also provides advanced query capabilities to guide the researchers to identify which of those data are of interest for their specific use. The broad hypothesis behind the work presented here is that letting the abundant real astrophysical data speak for itself, with minimal supervision and no labels, can reveal interesting patterns and distil information in ways that may facilitate data discovery. The need for supervision to be minimal is a must, given the sheer size and diversity of the archive contents. To this end, we have applied Deep Learning techniques to large datasets in the ESO Science Archive, most notably the entire data history of the HARPS high-resolution spectrograph. In our first round of experiments, we have obtained promising results in two directions. On the one side, the networks have learned to infer physical parameters of the celestial sources, such as radial velocity and effective temperature, just by being exposed to a large number of stellar spectra, without being explicitly asked to do so. On the other side, networks with the same basic architecture have shown an inclination to separate stellar and telluric spectral components, again without being instructed to do so. The lessons learned from the exercise and prospects will be presented.