Dark Matter comprises 85% of the Universe, yet there is no experimental knowledge of its properties. Direct detection experiments generate Petabytes of extremely noisy spatiotemporal data, in which scientists search for the signatures of exceptionally rare particles at the noise level of the sensors. Currently used techniques do not readily allow for domain knowledge to be incorporated into the models that are being built and learned for supporting this search. As a way of embedding domain knowledge into the solution, we propose using the framework of probabilistic graphical models, which can readily incorporate physical constraints and prior knowledge. The learned Probabilistic Graphical Model can be used to infer interaction positions of experimentally measured events, and thus uncover the meaningful signals in noisy detector data.