In this thesis a Neural Net (NN) code is first presented from scratch and
applied to the Kaon-Pion matching in the rare Kaon decay
(K+ → π+νν¯ ) analysis of NA62 at CERN. The NN code showed increased efficiency in Kaon decay identification with respect to the standard algorithm based on statistical analysis. It is designed and trained on K+ → π+π+π− decay channel to optimize the statistical significance of K+ - π+ matching by
amplifying the association between parent Kaons and downstream Pions
over accidental beam particles (“Pileup”) and final state Pions. Essential
enhancement and evaluation processes using state-of-the-art techniques
of XAI (eXplainable Artificial Intelligence) are presented in the context
of choosing the optimal NN-discriminant that fits in the framework of
πνν analysis in NA62 based on necessary physics-related metrics. Another
XAI application of an innovative Calorimetric “Virtual Bubble Chamber”
technique, called NNODA (Neural Net Object Detection Approach), for
NA62’s LKr (Liquid Krypton Calorimeter) is constructed to analyze
images of clusters using DL (Deep Learning) Computer Vision (CV)
techniques. The idea is to use color tags on the cluster timing to veto
random activities and unwanted decay products (mainly π0 background)
allowing an unusual and flexible event selection time window of ±10 ns
around the arrival time of the charged single particle in the final state.
NNODA efficiently increased signal acceptance by controlling random cuts.
Additionally, practical data science skills in Robotics are presented, by
training algorithms that would help a drone to identify and locate endeffectors in unusual environments. Then, An AI-based vision system is proposed for an embedded device and presented in its full facets, and specifically uses DL CV in image classification and object detection. These XAI tools and others have been successfully transferred to NA62’s most precise measurement of
Br (K+ → π+νν¯ ) in a cross-disciplinary fashion.