Improved machine learning models for local optics corrections in the LHC triplet magnets

The thesis aims to develop a beam optimization process in LHC, particularly the correction of the triplet magnet at each interaction point. This study will construct a machine learning algorithm to predict the quadrupole field errors around IPs, making the correction process more precise and efficient. The models include two types of features: phase advances and β-beating around IPs. The impact that each feature has on precise predictions will be explored. The thesis also explores models' robustness while incorporating the various hyperparameters, compares their global and local performances, and assesses their ability to make local corrections. Models are also trained to predict different error sources, specifically energy offsets for both beams of LHC. Studies have shown that the triplet field errors are predicted with the high precision, especially while including β-beating around IPs into the features. Prediction of energy offset showed that it can be predicted by the model very well, and moreover, can develop the field error precisions since these errors are presented in the real-time beam. Studies have shown that incorporating β-beating makes models very robust in the global case, but this is no longer the case locally. The beam's performance response to the different hyperparameters does not match in the global and local cases, which motivates further studies and suggests incorporating a local approach into the training process.