Unsupervised particle tracking with neuromorphic computing

The core of this thesis is the development of a Spiking Neural Network (SNN) model for particle reconstruction in high-energy physics and for the Compact Muon Solenoid (CMS) experiment specifically, as detailed in Chapter 4. The main distinctive features of our SNN model are its adaptive learning capabilities via spike-timing-dependent plasticity (STDP) and the use of a genetic algorithm for hyperparameter tuning. Unlike traditional deep learning systems, the SNN model replicates the biological mechanics of actual neurons, allowing it to process ionization hits as temporal spike patterns. This event-driven paradigm enables low-power, high-speed monitoring and real-time trigger applications for future colliders, where managing massive data rates is a significant issue. The first chapters provide the historical and technical background necessary to contextualize the entity of this contribution. For instance, Chapter 1 focuses on traditional track reconstruction methods, highlighting the strengths and weaknesses of each approach to help readers understand how the SNN might improve this task. Chapter 2 shifts the discussion to machine learning methods, presenting a historical overview of AI approaches that led to neural networks, as well as an outline of their conceptual basis, strengths, drawbacks, and major advances in pattern recognition, as well as introducing frameworks and ideas that will be pivotal in our model. Chapter 3 describes Masquelier's work on spike-timing-dependent plasticity (STDP) in neural networks, which will serve as a powerful inspiration for our own study. Finally, Chapter 4 illustrates in detail our innovative network proposal, designed for performing unsupervised track learning in high luminosity environments. The positive results we obtained shed light on the potential applications of neuromorphic approaches in high-energy physics and call for future investigations and refinements.