Direct searches for sub-GeV dark matter are limited by the intrinsic quantum properties of the target material. In this proof-of-concept study, we argue that this problem is particularly well suited for machine learning. We demonstrate that a simple neural architecture consisting of a variational autoencoder, and a multi-layer perceptron can efficiently generate unique molecules with desired properties. In specific, the energy threshold and signal (quantum) efficiency determine the minimum mass and cross section to which a detector can be sensitive. Organic molecules present a particularly interesting class of materials with intrinsically anisotropic electronic responses and O(few) eV excitation energies. However, the space of possible organic compounds is intractably large, which makes traditional database screening challenging. We adopt excitation energies and proxy transition matrix elements as target properties learned by our network. Our model is able to generate molecules that are not in even the most expansive quantum chemistry databases and predict their relevant properties for high-throughput and efficient screening. Following a massive generation of novel molecules, we use clustering analysis to identify some of the most promising
molecular structures that optimise the desired molecular properties for dark matter detection.