In the realm of energy research, a team of scientists from various institutions, including the University of Coimbra, INFN-Laboratori Nazionali di Legnaro, and the University of Bologna, among others, have been exploring innovative machine learning techniques to enhance the capabilities of the CYGNO experiment. This experiment is designed to search for rare, low-energy interactions, which could potentially include dark matter particles.
The CYGNO experiment utilizes an optical-readout Time Projection Chamber (TPC) to capture finely resolved scintillation images. These images, while rich in topological information, are large and sparse, posing challenges for real-time data processing and background discrimination. To address these challenges, the researchers have developed two complementary machine-learning approaches.
The first approach is an unsupervised strategy for online data reduction. This method employs a convolutional autoencoder, a type of neural network, trained exclusively on pedestal images—frames acquired with the detector’s amplification disabled. The autoencoder learns the detector’s noise morphology and highlights particle-induced structures through localized reconstruction residuals. From these residuals, compact Regions of Interest (ROIs) are extracted. The researchers found that this configuration retains approximately 93% of the reconstructed signal intensity while discarding about 97.8% of the image area, with an inference time of around 25 milliseconds per frame on a consumer GPU.
The second approach involves a weakly supervised application of the Classification Without Labels (CWoLa) framework. Using data acquired with an Americium-Beryllium neutron source, the researchers trained a convolutional classifier to identify nuclear-recoil-like topologies. Notably, the classifier was trained using only mixed AmBe and standard datasets, without event-level labels. The performance achieved approaches the theoretical limit imposed by the mixture composition, isolating a high-score population with compact, approximately circular morphologies consistent with nuclear recoils.
These machine learning techniques could have significant implications for the energy sector, particularly in the field of nuclear and particle physics research. By improving the efficiency and accuracy of data processing, these methods could enhance the search for rare interactions and potentially lead to new discoveries in energy-related research. The research was published in the journal Physical Review D.
This article is based on research available at arXiv.