Rowan Kelleher

Bio

Hi! My name is Rowan Kelleher and I am currently a 4th year undergraduate at Duke University studying physics. I grew up in Portland, Oregon and have spent most of my time since then in Durham either in class or doing research. I am also planning to graduate with minor degrees in philosophy and computer science as I find both of those disciplines quite interesting. In my free time I try to play a lot spikeball, ping pong, and pickleball. I also enjoy cooking when I have time.

Research

I started my research journey the summer after my first year of undergraduate when I began working with Prof. Vossen through the HEP 101 program. My work has focused on CLAS12 experiments, first with a project investigating the applicability of TMD factorization in phase space, and later with a project trying to improve Lambda identification. Recently I have been working on implementing a design for a iron-scintillator sandwich calorimeter in DD4HEP as well as creating a parameterization for the detector simulation to reduce computation time.

---

Affinity

A theory group recently proposed a tool, referred to as affinity, that estimates how well a SIDIS factorization theorem applies to a particular phase space. My work focused on adapting this tool to study the applicability of TMD factorization for CLAS12 data at JLab. Because the affinity tool requires partonic momenta, I created a ROOT macro that read CLAS12 monte-carlo simulation files and produced ROOT files with the relevant kinematics. I then used these kinematics to calculate TMD affinity in different kinematic bins, producing projections that suggest how the factorization improves or declines with changes in kinematics

When comparing my results with the original theory projections, we found differences between the projections, indicating that the distributions used to model partonic kinematics in the original proposal could be updated to improve the model. We are actively working on implementing distributions that approximate those produced by the monte-carlo event generator which we believe will provide a useful improvement in the model.

Improving Lambda Signal Extraction

Lambda identification at CLAS12 is difficult due to a number of factors, one being a large proton-pion background. Matthew McEneaney, a graduate student in Vossen Group, has worked on improving the Lambda signal extraction by using different approaches like GNN classifier networks and DAGNNs. My project builds on Matthew's work by performing domain adaptation between the GNN feature extraction and classification steps. I chose a normalizing flow network to implement the domain adaptation, and produced a working network based on the RealNVP architecture.

In order to train a classifier to distinguish signal and background events, a labeled training dataset must be used. Because we do not have labeled experimental data, we must train the classifier on labeled simulation data produced by a monte-carlo event generator. While the classifier will learn how to classify the simulated events, the simulation differs from the experiment, meaning the classifier will not perform as well on the experimental data. To remedy this miss-match, I created a network that attempts to make the data "look like" the simulation data, allowing the classifier to perform better.