Master Thesis Topics
Search for D** → D(*)γ
In semileptonic B meson decays (B→Xlν) some tensions exist between different measurements. Part of these decays involve so called D** mesons (excited c quark - light quark mesons with L=1). These mesons so far have only been observed to decay into D or D* meson plus up to 2 additional pions, and therefore are only considered via these decay modes in branching ratio estimations. The existence of the yet unobserved D**→D(*)γ decay would therefore be able to explain, at least in some extend, the observed tension. For this reason we would like to search for the D**→D(*)γ decay with the Belle II dataset. Two decay modes for this search are considered (each studied in a separate thesis):
1) The search for D1→Dγ in e+e-→cc events, where D1 is one of the D** mesons with a narrow width which makes it easy to identify. The observed rate for D1→Dγ will be compared to the rate of D1→D*Pi with D*→Dγ to factor out the normalization and some of the systematic uncertainties.
2) The other mode is the search for the decay B→D(*)γπ. The assumption here is that the B meson first decays into the D** via B→D**π, and the D** subsequently decays into D(*)γ. Since this method does not rely on the reconstruction of a specific D** meson (e.g. the narrow D1) it is more inclusive compared to the previously described method. This makes it also more difficult to distinguish individual D** mesons contributing. The existence of the B→D(*)γπ decay would be a strong hint for D**→D(*)γ decays, but the ability of this method to distinguish individual contributions would need to be studied.
For both modes the goal is to study on simulated data how sensitive a search with the current and future Belle II dataset is.
Gap mode studies
The semileptonic decay rate of B→Xclν (with Xc being a meson containing charm) can be estimated in two ways. One way, the so called inclusive method, does not explicitly reconstruct the Xc. The other way is to sum over all so far measured modes where the Xc has been explicitly reconstructed (e.g. B→Dlν; B→D*lν, ...). If all semileptonic decays of the B meson would be known both methods would yield the same result. But measurements result in a difference of around 14%, indicating that 14% of all semileptonic B meson decays are unknown. There are several candidates of so far unmeasured B→Xclν modes. For each of those modes a separate thesis topic is assigned. The goal of the thesis is to investigate the sensitivity of a search for one of these modes with the current and future Belle II dataset. The considered modes are:
1) B→D(*)ωlν with ω→π+π-π0 and hadronic tagging of the other B meson.
2) B→Λcplν with hadronic tagging of the other B meson.
3) B→D(*)η'lν with η'→π+π-η with hadronic tagging of the other B meson.
4) Inclusively reconstructed B→D(*)η'lν with η'→π+π-η.
The decay modes 1)-3) use hadronic tagging of the other B-meson in the BB-event. This allows to reconstruct the whole event except for the neutrino, which cannot be reconstructed by Belle II, by inserting its properties from the known kinematics of the rest of the event. This method provides a very clean sample, but is also hindered by a low selection efficiency. In contrast to that method 4) does not require the other B meson to be reconstructed and thus promises a much higher reconstruction efficiency, though accompanied with more background.
Search for Heavy Neutral Leptons in Displaced Semileptonic B-Meson Decays at Belle II
Heavy neutral leptons (HNLs) appear in many extensions of the Standard Model that explain the smallness of neutrino masses, and can also serve as a portal to a strongly coupled dark sector in which the sterile state is a composite excitation. In both cases, semileptonic B-meson decays B → D ℓ N offer a large kinematical window for producing GeV-scale sterile states, whose long lifetimes lead to displaced decay vertices with charged tracks inside the detector. This thesis develops a search for such displaced signatures at Belle II, using a prompt charged lepton to tag the B decay and requiring at least one displaced vertex formed by a dilepton pair. The clean e⁺e⁻ environment and precise vertexing of Belle II make this topology nearly background-free, while the multiplicity of displaced vertices provides a unique handle to discriminate between a weakly coupled HNL (a single vertex) and a strongly coupled composite sector producing dark jets (multiple vertices). The work comprises signal simulation and reconstruction, the design of the displaced-vertex selection, background studies, and an estimate of the sensitivity in the HNL mixing and compositeness parameter space. Interest in particle physics and curiosity in physics beyond the Standard Model is required, Python experience is very helpful.
Reconstruction of neutrons and search for B → K n nbar
Neutrons are rarely produced in particle decays and are often not considered in the detector design and reconstuction software of high energy physics experiments. Nevertheless they can play an important role. For example the B → K n nbar decay is a background in the measurement of the B → K nu nubar decay, but has not been experimentally measured so far. The neutron reconstruction at Belle II and the sensitivity to the B → K nu nubar decay should be studied.
Search for B decays with baryon number violation by two units
A precondition for the generation of a baryon asymmetry in our universe is baryon number violation. While the violation of baryon number by one unit (∆B = 1) involving quarks of the first or second generation is severely constrained by the non-observation of proton decay, limits on baryon number violating processes involving third generation quarks are considerably weaker. Processes violating baryon number by two units are even less studied. With Belle II data we can search for ∆B = 2 processes, such as B meson decays to a deuteron and light mesons or leptons.
Generic new physics search with machine learning
New physics can be established by measuring deviations from standard model predictions. Usually this approach is implemented for selected observables. New physics effects may be overlooked if they show up in observables that were not considered or only in correlations of observables. A detailed and more generic comparison of measurements and theoretical predictions may be achieved by machine learning algorithms. Besides the design of the machine learning model and challenge will be to decide if observed deviations should be attributed to new physics or imperfections in theory calculations or detector simulation. This topic is explorative and requires the combination of knowledge from multiple areas.
Agentic analysis to workflow translation @ Belle II
Current data analysis pipelines are built in individual, often unstructured code repositories with the goal of a single publication at the end. The knowledge transfer to future generations of physicists exists mostly through the documented text (paper and analysis reports). The goal of this project is to take existing analyses and build executable and structured workflows, which would enable us to build on existing work much more efficiently.
Agentic tools built on top of modern LLMs enable us to refactor such large code bases effectively. The outcome of the project would ideally be a team of agents which can reliably translate analysis code to reproducible workflows, including various subagents which take care of coordinating, building, validating, ensuring code quality etc. Proficiency in Python and version control (Git) are required, experience in particle physics data analysis and machine learning are desirable.
Reconstructing B Meson Decay Trees with Attention Networks at Belle II
Many of Belle II's most interesting searches, for example B→τν and B→Kνν̄, rely on reconstructing the second B meson in each event from its decay products. The current algorithm for the exclusive reconstruction of the second B meson, the Full Event Interpretation, works through about 100 hand-coded decay channels and reconstructs only a small fraction of events. In this thesis you will train an attention-based neural network that learns the decay tree structure directly from the measured final-state particles, with no hard-coded channels. The work runs on full Belle II simulation: you build the training pipeline, compare your model against the existing FEI and graFEI algorithms on a common benchmark, and run controlled experiments on open design questions, such as how the network should handle particles the detector never saw. Along the way you pick up modern machine learning practice (transformers, large-scale training, ablation studies), and the Belle II analysis software. Interest in particle physics and machine learning is required, Python and ML experience is very helpful.
Do neural networks learn the physics? Latent geometry and attention in particle decay models
Neural networks reconstruct particle decays well, but what they learn stays hidden, and in physics that is a real problem: a result only counts once it can be validated. This thesis looks inside attention-based networks trained on particle decay data. Decay trees are hierarchies, and hyperbolic geometry (the Poincaré ball) fits hierarchies unusually well, with depth in the tree mapping to radius and similar decays ending up near each other. You will train attention-based models on simulated decays and then probe them. Does the tree structure show up in the latent space at all? Attention maps seem to show which particles belong together, but do they survive faithfulness tests, or are they just pretty pictures? And does a geometry-aware prediction head beat a standard one in practice? One possible question defines the study: whether decays that differ only by an invisible neutrino land close together in the learned representation. You will learn representation learning, interpretability methods, and how to validate them to physics standards. Interest in machine learning and geometry as well as curiosity about what neural networks actually learn are required, Python and ML experience is very helpful.
