From neutrons to dark matter: Directional recoil detection and utilization of deep learning for gaseous time projection chambers

dc.contributor.advisor Vahsen, Sven E. Schueler, Jeffrey Thomas
dc.contributor.department Physics 2022-10-19T22:36:19Z 2022-10-19T22:36:19Z 2022 Ph.D.
dc.subject Particle physics
dc.subject Belle II
dc.subject Convolutional neural network
dc.subject Directional recoil detection
dc.subject Fast neutron backgrounds
dc.subject Head/tail effect
dc.subject SuperKEKB
dc.title From neutrons to dark matter: Directional recoil detection and utilization of deep learning for gaseous time projection chambers
dc.type Thesis
dcterms.abstract Modern gaseous time projection chambers (TPCs) with high readout segmentation are capable of reconstructing detailed 3D ionization distributions with voxel sizes of order (\SI{100}{\um})$^3$. This enables measurements of the 3D momentum vectors of short, mm-scale nuclear recoils, which is of interest for neutron measurements, as well as searches for dark matter, where directionality opens the possibility of identifying the galactic origin of weakly interacting massive particles (WIMPs), even below the so-called neutrino floor. We perform a variety of experiments and simulations with eight miniature TPCs filled with a 70:30 mixture of He:CO$_2$ gas at \SI{1}{atm} pressure. Each so-called BEAST TPC is of identical design and contains two gas electron multiplier (GEM) amplification devices and a $(2.00\times 1.68)~\text{cm}^2$ pixel-ASIC readout. We first detail the measurement of neutron backgrounds at the SuperKEKB $e^+e^-$ collider in Tsukuba Japan. We focus on measurements surrounding SuperKEKB's final focusing magnets (recorded in 2018) and in the accelerator tunnel surrounding the Belle II detector (recorded in 2020-2021). In our analyses we reject large X-ray backgrounds from the accelerator, resulting in $>$99$\%$ pure samples of nuclear recoils down to recoil energies as low as \SI{8.0}{keV_{ee}}. We find excellent agreement between measured and simulated nuclear recoil energy spectra indicating that our simulations model neutron production well. We additionally introduce a correction for charge integration bias in observed recoil tracks with high axial inclination. This correction leads to correct vector directional ``head-tail" (sign of 3D vector) assignment for $91\%$ of simulated He recoils ranging from $\SI{40}{keV_{ee}}$ to about $\SI{1}{MeV_{ee}}$, with a mean angular resolution of 8$^\circ$; a significant improvement over the $72\%$ head-tail efficiency achieved without these corrections. Applying this technique to measurement leads to an agreement between measured and simulated angular distributions that allows us to conclude the existence of a neutron production hotspot in the accelerator tunnel. While the BEAST TPCs are highly sensitive to ionization, and can detect even single electrons, extending directionality to the keV-scale, as is desirable for dark matter searches, requires operating the detectors with lower-density gases, at higher gains, and developing improved analysis techniques. We here focus on the two latter aspects. We improve on existing head-tail classification methods through the introduction of deep-learning computer-vision algorithms called 3D convolutional neural networks (3DCNNs). We first perform a simulation benchmark study where we train a 3DCNN to assign directional head-tail to simulated neutron recoils with energies up to \SI{515}{keV_r} and compare these results to three existing methods of head-tail assignment. We find a head-tail efficiency of $99.9\%$ on this sample using the 3DCNN, compared to $97.8\%$, $93.7\%$, and $79.0\%$ for existing methods. Next, we measure neutrons from a $^{252}$Cf source incident on separate sides of a TPC. We operate both at low gain and high gain. At low gain, the simulation-trained 3DCNN reliably identifies whether the observed recoil points toward or away from the $^{252}$Cf source. On a small sample of identified He recoils between \SI{39}{keV_{ee}} and \SI{49}{keV_{ee}}, before correcting for residual background such as back-scattered events, we observe a head-tail efficiency of $(62.1\pm 11.4)\%$. Using simulation, we show that the true head-tail efficiency after correcting for residual backgrounds should be greater than this, marking the first statistically significant observation of event-level head-tail sensitivity below \SI{50}{keV_{ee}}. At high gain, we attempt to improve our head-tail sensitivity to sub-10-$\rm keV_r$ recoils, and also introduce a 3DCNN for event identification. In simulation, we reject all X-ray backgrounds down to \SI{5}{keV_{ee}} at $50\%$ nuclear recoil selection efficiency and demonstrate head-tail efficiencies above $50\%$ for He recoils down to \SI{3}{keV_r}. These results do not yet generalize to measurement, which is currently being investigated. If the 3DCNN robustness can be improved, this would be the first demonstration of directional recoil detection at energies relevant for the directional detection of $\mathcal{O}(\text{GeV})$ dark matter particles. Finally, we perform a study comparing the keV-scale electron background rejection performance of a 3DCNN to the traditional discriminant of track length, as well as discriminants obtained from state-of-the-art shallow learning methods in a simulated detector with an 80:10:10 mixture of He:CF$_4$:CHF$_3$ at \SI{60}{torr}. We train the 3DCNN classifier using recoil charge distributions with ionization energies ranging from 0.5-\SI{10.5}{keV_{ee}} after \SI{25}{cm} of drift. The charges are initially segmented into (\SI{100}{\um})$^3$ bins when determining track length and the shallow learning discriminants, but are rebinned with a reduced segmentation of (\SI{850}{\um})$^3$ for the 3DCNN. Despite the courser binning, compared to using track length, we find that classifying events with the 3DCNN reduces electron backgrounds by an additional 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 \SI{2}{keV_{ee}} reduction in the energy threshold. Collectively, the results in this thesis highlight the unique measurements enabled by high-resolution ionization imaging, and how 3D convolutional neural networks appear ideally suited to maximally utilize the rich 3D data from detectors with this capability.
dcterms.extent 216 pages
dcterms.language en
dcterms.publisher University of Hawai'i at Manoa
dcterms.rights All UHM dissertations and theses are protected by copyright. They may be viewed from this source for any purpose, but reproduction or distribution in any format is prohibited without written permission from the copyright owner.
dcterms.type Text
Original bundle
Now showing 1 - 1 of 1
No Thumbnail Available
57.23 MB
Adobe Portable Document Format