Tricky Tracks with the CATFinder at Belle II

Precision tracking of charged particles is a key ingredient for the physics programme of the Belle~II experiment at the SuperKEKB.
In particular, decays in flight of charged kaons and pions, such as $K^{+} \to \mu^{+} \nu_\mu$ and $\pi^{+} \to \mu^{+} \nu_\mu$ and their charge conjugates, pose a challenge for standard track-finding algorithms.
These tracks, hereafter referred to as tricky tracks, exhibit a kink topology with a change of momentum and direction at a displaced decay vertex.
They are relevant for a range of flavour-physics and new-physics analyses, since a successful reconstruction leads to an improvement in the particle identification and $\text{d}E/\text{d}x$ calibration.
A successful reconstruction of tricky tracks would also make new processes measurable.

This thesis studies the reconstruction of tricky tracks using the CDC AI Track Finder, called CATFinder, a graph neural network-based track-finding algorithm for the central drift chamber, and a dedicated configuration for tricky tracks.
The focus is on kink topologies from $K^{+} \to \mu^{+} \nu_\mu$ and $\pi^{+} \to \mu^{+} \nu_\mu$ decays, where both the mother and daughter tracks are reconstructable in the central drift chamber.
To this end, simulated Monte Carlo samples are used that combine realistic physics events with specialised particle-gun configurations.

On the algorithmic side, the CATFinder is extended with two additional binary classifiers trained to identify mother and daughter particles in tricky tracks.
Three training strategies are explored that differ in the relative weight assigned to the new classifier terms in the loss function and in whether a pretraining phase is used without these terms.
The performance of the resulting models is evaluated on prompt pion samples and mixed $B$ events.
The evaluation considers general tracking metrics, kink-specific metrics, momentum resolution, hit-based metrics, and classifier-quality measures such as receiver operating characteristic curves.
Comparing the training strategies provides insight into the trade-offs between generic tracking performance and specialised sensitivity to tricky tracks.