The SuperCDMS dark matter detection experiment uses a multi-channel, multi-template (N×M) fitting algorithm to reconstruct events in the detector. This algorithm fits each signal from the 12 detector channels with a weighted sum of linearly independent templates and optimizes a χ² function for a set of N×M amplitudes for each respective template and channel. The standard energy reconstruction method uses a weighted sum of the N×M amplitudes along with a trained Boosted Decision Tree (BDT) to map each event to its energy and a correction factor respectively.
The proposed improvement to this method consist of an improved template generation technique along with an optimized energy reconstruction algorithm. N×M assumes linear scaling of the signals with the deposited energy inside the detector. At higher energies, however, the channels experience some saturation due to the temperature-resistance curve of the Transition Edge Sensors (TES). The revised template generation method fits for five templates. The first three templates are generated using events at lower energies where saturation is negligible, as in the original implementation of N×M. The final two templates, are derived from the residual of a saturated signal and the appropriate linear combination of the first three non-saturated templates. This creates a set of five templates, three non-saturated and two saturated, that match the resolution of the 5 standard templates at lower energies but have a 2-5 eV improvement at energies above 10 keV.
For energy reconstruction, events are separated into center and edge events using a decision tree. For energy reconstruction of center events using N×M amplitudes, we propose the use of a quadratic energy estimator that includes linear and quadratic functions of the N×M amplitude, and appropriate weights trained on a set of simulated event. The range of energies are split into discrete bins and a distinct set of quadratic weights is calculated for each bin. A general linear estimator trained on the 0.5-12 keV range of events is used to place each event in the appropriate bin. A MultiLayer Perceptron (MLP) neural network is used for the energy reconstruction of the edge events. This method provides a resolution of 0.7 - 2.3% within our energy spectrum.
Add to Calendar
2026-08-12T13:00:002026-08-12T14:00:00Optimizing the Energy Resolution of the SuperCDMS Multi-channel,Multi-template Event Reconstruction AlgorithmEvent Information:
Abstract:
The SuperCDMS dark matter detection experiment uses a multi-channel, multi-template (N×M) fitting algorithm to reconstruct events in the detector. This algorithm fits each signal from the 12 detector channels with a weighted sum of linearly independent templates and optimizes a χ² function for a set of N×M amplitudes for each respective template and channel. The standard energy reconstruction method uses a weighted sum of the N×M amplitudes along with a trained Boosted Decision Tree (BDT) to map each event to its energy and a correction factor respectively.
The proposed improvement to this method consist of an improved template generation technique along with an optimized energy reconstruction algorithm. N×M assumes linear scaling of the signals with the deposited energy inside the detector. At higher energies, however, the channels experience some saturation due to the temperature-resistance curve of the Transition Edge Sensors (TES). The revised template generation method fits for five templates. The first three templates are generated using events at lower energies where saturation is negligible, as in the original implementation of N×M. The final two templates, are derived from the residual of a saturated signal and the appropriate linear combination of the first three non-saturated templates. This creates a set of five templates, three non-saturated and two saturated, that match the resolution of the 5 standard templates at lower energies but have a 2-5 eV improvement at energies above 10 keV.
For energy reconstruction, events are separated into center and edge events using a decision tree. For energy reconstruction of center events using N×M amplitudes, we propose the use of a quadratic energy estimator that includes linear and quadratic functions of the N×M amplitude, and appropriate weights trained on a set of simulated event. The range of energies are split into discrete bins and a distinct set of quadratic weights is calculated for each bin. A general linear estimator trained on the 0.5-12 keV range of events is used to place each event in the appropriate bin. A MultiLayer Perceptron (MLP) neural network is used for the energy reconstruction of the edge events. This method provides a resolution of 0.7 - 2.3% within our energy spectrum. Event Location:
Hennings Room 318