ML4CPD | Felix Wagner

Machine learning for cryogenic particle detectors (ML4CPD) is a project funded by the industrial PhD program of the Austrian research promotion agency for the years 2020-2023. In the course of this project we are building machine learning methods to automate data analysis and control of cryogenic calorimeters that are used by the CRESST and COSINUS experiments for the direct detection of dark matter. Below I highlight some pieces of the scientific output of our project.

Cait: analysis toolkit for cryogenic particle detectors

Cryogenic solid state detectors are widely used in dark matter and neutrino experiments, and require a sensible raw data analysis. For this purpose, we present Cait, an open source Python package with all essential methods for the analysis of detector modules fully integrable with the Python ecosystem for scientific computing and machine learning. It comes with methods for triggering of events from continuously sampled streams, identification of particle recoils and artifacts in a low signal-to-noise ratio environment, the reconstruction of deposited energies, and the simulation of a variety of typical event types. Furthermore, by connecting Cait with existing machine learning frameworks we introduce novel methods for better automation in data cleaning and background rejection.

Cait was published in the CSBS journal (Comput Softw Big Sci 6, 19) and is online on PyPI, Zenodo, ReadtheDocs and GitHub.

Towards and automated data cleaning with deep learning in CRESST

Our preprint about automated data cleaning is on online (arXiv:2211.00564) and was accepted for publication in the EPJ+ journal.

The signals recorded with cryogenic calorimeters need to undergo a careful cleaning process to avoid wrongly reconstructed recoil energies caused by pile-up and read-out artefacts. We frame this process as a time series classification task and propose to automate it with neural networks. With a data set of over one million labeled records from 68 detectors, recorded between 2013 and 2019 by CRESST, we test the capability of four commonly used neural network architectures to learn the data cleaning task. Our best-performing model achieves a balanced accuracy of 0.932 on our test set. We show on an exemplary detector that about half of the wrongly predicted events are in fact wrongly labeled events, and a large share of the remaining ones have a context-dependent ground truth. We furthermore evaluate the recall and selectivity of our classifiers with simulated data. The results confirm that the trained classifiers are well-suited for the data-cleaning task.

Different from standard approaches, we do not rely on prior knowledge of individual detectors, such as their characteristic pulse shape, or manual interventions, such as finding individual cut values. Our results can have the following impacts: 

For large-scale detector setups the fine-tuning of dedicated cuts for each detector can produce non-negligible delays in the analysis or even become infeasible. The application of our models instead can produce equivalent cuts instantly.
Recorded data can be monitored in real-time. This can uncover unwanted shifts in the measurement setup, e.g. an increased rate of artefacts, immediately and enable fast interventions. Furthermore, features in the event distribution, e.g. a peak in the PH spectrum, can be identified as particle-like or artefact-like, without the need for designing cuts first.

Optimal operation of cryogenic calorimeters through deep reinforcement learning

Our method for automated and optimal detector operation was presented at ACAT22 and will be published in 2023.

Cryogenic phonon detectors are used by direct detection dark matter experiments to achieve sensitivity to light dark matter particle interactions. Such detectors consist of a target crystal equipped with a superconducting thermometer. The temperature of the thermometer and the bias current in its readout circuit need careful optimization to achieve optimal sensitivity of the detector. This task is not trivial and has to be done manually by an expert. In our work, we created a simulation of the detector response as an OpenAI Gym reinforcement learning environment. In the simulation, we test the capability of a soft actor critic agent to perform the task. We accomplish the optimization of a standard detector in the equivalent of 30 minutes of real measurement time, which is faster than most human experts. Our method can improve the scalability of multi-detector setups.