SIMBAD references

We describe an optimal signal extraction process for imaging X-ray polarimetry using an ensemble of deep neural networks. The initial photoelectron angle, used to recover the polarization, has errors following a von Mises distribution. This is complicated by events converting outside of the fiducial gas volume, whose tracks have little polarization sensitivity. We train a deep ensemble of convolutional neural networks to select against these events and to measure event angles and errors for the desired gas-conversion tracks. We show how the expected modulation amplitude from each event gives an optimal weighting to maximize signal-to-noise ratio of the recovered polarization. Applying this weighted maximum likelihood event analysis yields sensitivity (MDP₉₉) improvements of ∼10% over earlier heuristic weighting schemes and mitigates the need to adjust said weighting for the source spectrum. We apply our new technique to a selection of astrophysical spectra, including complex extreme examples, and compare the polarization recovery to the current state of the art.