Optimal Operation of Cryogenic Calorimeters Through Deep Reinforcement Learning.

Area of Science:

• Physics
• Astrophysics
• Machine Learning

Background:

• Cryogenic phonon detectors with transition-edge sensors offer high sensitivity for dark matter detection.
• Optimizing detector performance requires careful tuning of thermometer temperature and bias current, a complex manual task.

Purpose of the Study:

• To automate the optimization of cryogenic phonon detectors using reinforcement learning.
• To compare the performance of reinforcement learning optimization against human experts.

Main Methods:

• Developed a reinforcement learning framework to optimize detector parameters.
• Trained the model on simulated data from Cryogenic Rare Event Search with Superconducting Thermometers (CRESST) detectors.
• Validated the trained model by optimizing detectors in a live, underground CRESST setup.

Main Results:

• Reinforcement learning achieved comparable detector optimization results to human experts.
• The automated method significantly reduced the time required for detector tuning.
• Successful optimization was demonstrated in both simulated and real-world detector environments.

Conclusions:

• Reinforcement learning provides an efficient and effective method for optimizing cryogenic phonon detectors.
• This automation reduces manual intervention, enabling faster deployment and tuning of large-scale detector arrays.
• The developed technique advances the field of direct dark matter detection experiments.