The polarization of massive vector bosons is highly sensitive to the electroweak symmetry breaking and the Standard Model gauge symmetry structure. The comparison of polarization measurements with theoretical predictions provides a valuable insight into the Standard Model as well as Beyond the Standard Model physics.

Due to their short lifetimes, massive vector bosons cannot be observed directly but appear as intermediate particles. Their polarization is obtained indirectly from measurable, unpolarized observables using theoretically constructed polarization templates.

In this thesis, a machine learning-based approach for event-by-event vector boson polarization tagging is studied. Polarized samples of vector boson production processes are generated using the Monte Carlo event generator Sherpa where polarization fractions, defined as the ratio of polarized to unpolarized cross sections, are determined for each event. Neural networks are trained on the simulated samples to predict the event-wise polarization fractions of intermediate massive vector bosons based on experimentally accessible final-state kinematics. With this approach, the NN output can be directly compared to the true polarization distribution and the NN performance can be evaluated quantitively.

The performance of two neural network architectures, a multilayer perceptron and a more advanced transformer model, is systematically compared. The study is conducted for Z+jets and W+jet production processes, as well as for the electroweak production of a 𝑊 + 𝑍 boson pair in association with two jets at fixed leading order and matched to the parton shower. Additionally, the effect of higher-order corrections in the strong coupling constant and the influence of different phase spaces are examined, with Z+jets serving as the test bed.