Antihydrogen, composing an antiproton and positron, is the only bound state of two antiparticles yet to be synthesised, making for an enticing system to study the purported symmetry of matter and antimatter. As antihydrogen does not occur naturally in the observable universe, any study of this atom requires it to be synthesised in a lab, which the ALPHA experiment is routinely able to do. However, the absolute numbers are small and efficient detection is crucial for the experiment.

To detect these atoms, ALPHA deploys two main annihilation detectors: a silicon vertex detector and a new time projection chamber installed in 2018. A key challenge for both detectors is distinguishing between antimatter annihilations and background events (e.g. cosmic radiation), a task for which machine learning is well suited. Presently, for the silicon vertex detector, this is done with high-level variables, while the time projection chamber has no way of filtering these events. In the present work, we have developed the first models capable of filtering events in the time projection chamber, which proved vital in the first measurement of the effect of gravity on the motion of antimatter. A first-of-its-kind deep learning model trained on low-level data from the silicon vertex detector has been developed, and it can successfully classify events to a high degree of accuracy. Further, the newest models trained for the silicon vertex detector are presented. The use of these models on real data is included, and all results generated by ALPHA from the 2022-2024 experimental runs will use the models described in this thesis.

Finally, the transverse beam profile in the accelerators throughout CERN (such as the one used to provide antiprotons to ALPHA) is an important metric for successful operation. The significant increase in beam intensities poses a challenge that make the currently deployed correcting magnetic fields undesirable. The possibility of using machine learning to reconstruct beam profiles in the Proton Synchrotron is presented, and a first attempt at applying these models to real data is included which, despite a troubled dataset, shows promising results.