Speaker
Description
The Latin American Giant Observatory (LAGO) is a ground based observatory studying solar or high energy astrophysics transient events. LAGO takes advantage of its distributed network of Water Cherenkov Detectors (WCDs) in Latin America as a tool to measure the secondary particle flux reaching the ground. These secondary particles are produced during the interaction of the modulated cosmic rays flux with the atmosphere.
The LAGO WCDs are sensitive to secondary charged particles, high-energy photons through pair creation and Compton scattering and even neutrons thanks to the deuteration of protons in the water volume. The pulse shape generated by these particles depends on detector geometry, water purity, reflectivity and diffusivity of the inner coating. Due to the decentralized nature of LAGO, these properties are different for each node. Additionally, the pulse shape depends on the convolution between the response of the detector central photomultiplier (PMT) to individual photons and the time distribution of the Cherenkov photons reaching the PMT. Typically, a WCD gives pulses with a sharp rise time (~10 ns) and a longer decay time (of ~70 ns). The WCD data used in this work was acquired using the first version of the LAGO DAQ that digitizes pulses at a sampling rate of 40 MHz and 10 bits resolution on windows of 300 ns.
In this scenario, we applied unsupervised machine learning techniques to find patterns in the WCDs data and subsequently create groups, through clustering, that can be used to provide particle separation. We used LAGO data acquired with individual WCDs, showing that density-based clustering algorithms are suitable for automatic particle separation producing good candidate groups. Improved separation would help LAGO to reconstruct in situ the properties of the secondary cosmic rays flux. The obtained outcomes were validated with a dedicated Monte Carlo simulation that takes into account the effects expected during the regular operation.
These results open the possibility to deploy machine-learning-based models in our distributed detection network for onboard data analysis in a semi-operative manner, and allow the installation of detectors at very remote sites.
