I'm Matthew McEneaney

This project arose out of necessity for an ML project I was working on. I needed a fast python library for reading and writing data with the CLAS12 binary file format HIPO, and the existing library was outdated and clunky. I used pybind11 to expose C++ classes and methods in python, and then created the interface on top of that. The interface allows for fast batching and globbing large sets of files and has helped myself and many others to quickly implement ML projects. It also boasts a speed boost by a factor 22.8 compared to the old python libraries.

I found myself constantly rewriting the same analysis code and so one day decided to expand it into a generalizable framework. This framework implements solutions in C++ ROOT for complex multi-dimensional binning, maximum likelihood fits for distinguishing signal and background mass spectra and extracting asymmetries, and much more. It also includes python libraries for configuring jobs to run with SLURM, applying bin migration corrections, and plotting binned results in a systematic way. All told, this has saved me over half the time I would have spent writing new analysis code and configuring job submissions. Installation is made easy via a container image.

This project expands on the domain adversarial GNN approach for event classification that I developed and published. It implements machine learning pipelines for domain-adversarial, contrastive adaptation, and Transfer of Imbalance-Tolerant Knowledge (TIToK) domain adaption approaches in training via PyTorch and standard python libraries such as numpy and matplotlib. The GNN event classifier approach is able to improve signal yield by a factor of 1.95 in simulation. Hyperparameter optimization studies may be run via Optuna and the results are stored in an SQL database. The project also allows one to deploy trained GNN event classifier via Flask and is available as a container image.

This tool, similar to SAGA, arose out of the realization that I was always rewriting the same event selection code for physics analyses. This tool is written in groovy and uses a JNI bridge developed by a colleague to write data to C++ ROOT. It can select any final state topology desired from physics events and compute and store particle kinematics, momenta, and detector level information in ROOT Trees. Installation is made easy through distribution as a container image. This tool has saved me countless hours writing event selection code for new analyses.

This project is a framework of java and groovy libraries run with bash scripts for a data quality assurance database developed by a postdoc in our group, and to which I have contributed several key features. Recently, I implemented a CSV reader for taking in data from an outside source in the case it had been corrupted. I also tracked down and fixed a bug in timeline creation when computing interquartile ranges that resulted in a factor of 100 times faster speed.