We propose the development of a deep clustering framework capable of blindly sifting through continuous seismic array data and identifying families of different behaviors. Given the diversity of seismic data recorded at volcanoes, we choose Erebus volcano as a first study area and exploit its multi-decadal data sets and pre-existing event catalogs that include icequakes, Strombolian eruptions, active source explosions, ash vent eruptions, and other exotic transient behaviors. We propose a multi-modal variational autoencoder as a base framework, which features the ability to uniquely determine representative basis vectors of high dimensional spectrogram data while also permitting the inclusion of array phase information, classically lost by the input convolutional layer processes. Initial results show a promising ability to robustly distinguish between icequakes and eruptions at single stations, a task typically outside the ability of a human operator, and we propose several ways to augment the framework with array data. Finally, we propose running a 3 day workshop for early career scientists and graduate students to intuitively introduce them to a variety of deep learning architectures, and present coding tutorials of base concepts all the way to high level results across fields.