Towards an automated data cleaning with deep learning in CRESST.

Area of Science:

• * Particle Physics
• * Astrophysics
• * Data Science

Background:

• * The CRESST (Cryogenic Rare Event Search and Superconducting Spectrometer) experiment uses cryogenic calorimeters to detect nuclear recoils from potential dark matter particles.
• * Accurate reconstruction of recoil energies is crucial, but susceptible to pile-up and detector artefacts.
• * Current data cleaning methods are manual and time-consuming, necessitating automated solutions.

Purpose of the Study:

• * To automate the data cleaning process for CRESST experiment's nuclear recoil signals using neural networks.
• * To evaluate the effectiveness of different neural network architectures for this time series classification task.
• * To assess the performance of the trained models in terms of accuracy, recall, and selectivity.

Main Methods:

• * A dataset of over one million labeled signal records from 68 CRESST detectors (2013-2019) was used.
• * Four common neural network architectures were trained and tested for time series classification.
• * Performance was evaluated using balanced accuracy on real data and recall/selectivity on simulated data.

Main Results:

• * The best performing neural network model achieved a balanced accuracy of 0.932.
• * Analysis revealed that approximately 50% of misclassified events were due to incorrect initial labeling.
• * Further investigation showed a significant portion of remaining misclassifications had context-dependent ground truths.

Conclusions:

• * Neural networks provide a highly effective automated solution for cleaning CRESST data, significantly improving signal processing.
• * The developed models demonstrate strong performance, confirming their suitability for real-time data cleaning in dark matter detection.
• * Findings highlight the potential for machine learning to refine the analysis of complex experimental data in particle physics.