The project of a Parallel Library for Identification and Study of Astroparticles (pLISA) was restarted in the reporting period. Open-source technologies like TensorFlow and Keras have become established and suitable to be used as a foundation to build upon. As a consequence, even with limited human resources it was possible to produce relevant scientific results. The goal of pLISA is to set up a continuous training framework, to complement and possibly replace traditional algorithms to identify and study astroparticles in event-based detectors.
While human-made reconstruction methods are clear to understand and control, they suffer from slow evolution and are heavily dependent on single developers. The task of recognising, classifying and studying astroparticle events is at the foundation of multimessenger astrophysics and astronomy and could profit from continuous improvement (also including reanalysis of older datasets). Machine learning offers an opportunity to continuously improve the performances with limited human efforts, while providing immediately applicable results.
pLISA is a Python-based library that uses Convolutional Neural Networks (CNNs) to provide particle identification and energy/direction estimation. Such features and quantities are relevant in relating astroparticle events to astrophysical and cosmological sources. Particle detectors are represented as geometrical models, removing much of the technicalities that are needed in conventional algorithms. Events are represented as multidimensional images or movies (3D, 4D, 5D or higher). User-defined CNNs can then be trained to extract the relevant information without any a priori knowledge. pLISA has been tested on KM3NeT simulated events and even with limited development and training time, it has shown competitive with conventional reconstruction techniques. In particular, pLISA’s reconstruction have shown inherent log-linearity in energy reconstruction and stability in direction estimation, and able to provide meaningful output even in highly noisy events.
The evolution of pLISA is only at the beginning and a massive evolution is planned for the next future well beyond the end of ASTERICS. The first results were obtained by single GPU boards (NVidia GTX 1080Ti), but more powerful boards (NVidia RTX 2080Ti) and small farms are being set up, while the capabilities of the networks are being refined and methods are being developed to model also higher-order effects such as detector deformation, inefficiency and noise.
Source code: https://baltig.infn.it/bozza/plisa/
The library is actually a new development undertaken in the final reporting period. It is still evolving as models are added.
Machine Learning techniques can shift the research activity from software production to data analysis. GPU farms can be set up to continuously improve algorithms, removing the need to handle technicalities and helping researchers to focus on their primary interests.
pLISA is now being applied to KM3NeT simulated data to assess its strengths and weaknesses with respect to traditional algorithms. Development time has been short so far and there is room for optimisations. Nevertheless, pLISA truly shows the power of the Machine Learning approach, because networks with few hours’ training can be competitive with algorithms that humans have been polishing for many years. pLISA is developed with portability in mind, so it could easily support several geometries for event-based detectors, bringing shared expertise and know-how and a common toolbase to astroparticle physics experiments.
pLISA started as a container for generic Machine Learning algorithms with its own data representation. Later it was found that focusing more on modelling and leveraging open source libraries boosts productivity in this field.
Technological evolution in this field is fast and there is plenty of ready-made platforms. Finding and training Machine Learning experts takes time.