Deep learning for improved keV-scale recoil identification in high resolution gas time projection chambers
Recoil-imaging gaseous time projection chambers (TPCs) with directional sensitivity are attractive for dark matter (DM) searches. Detectors capable of reconstructing 3D nuclear recoil directions would be uniquely sensitive to the predicted dipole angular distribution of DM recoils that would unambiguously establish the galactic origin of a claimed DM signal and provide powerful discrimination against background recoils from solar neutrinos. These advantages can only be exploited however, if electron recoil backgrounds from gamma rays can be sufficiently suppressed. We introduce a deep learning-based recoil event classifier that uses a 3D convolutional neural network (3DCNN) to identify event species based on their recoil images. We compare electron background rejection performance of the 3DCNN both to the traditional discriminant of track length, as well as discriminants obtained from state-of-the-art shallow learning methods. We train the 3DCNN classifier using recoil charge distributions with ionization energies ranging from 0.5-10.5 $\rm keV_{ee}$, for 25 cm of drift in an 80:10:10 mixture of $\rm He$:$\rm CF_4$:$\rm CHF_3$. The charges are initially segmented into $(100\times 100\times 100)$ $\rmμm^3$ bins when determining track length and the shallow learning discriminants, but are rebinned with a reduced segmentation of about $(850\times 850\times 850)$ $\rmμm^3$ for the 3DCNN. Despite the courser binning, compared to using track length, we find that classifying events with the 3DCNN reduces electron backgrounds by a factor of up to 1,000 and effectively reduces the energy threshold of our simulated TPC by $30\%$ for fluorine recoils and $50\%$ for helium recoils. We also find that the 3DCNN reduces electron backgrounds by up to a factor of 20 compared to the shallow machine learning approaches, corresponding to a 2 $\rm keV_{ee}$ reduction in the energy threshold.