Particle identification in the GlueX detector with machine learning: software and data
Description
In this project we test the effectiveness of machine learning techniques to identify particles in the GlueX detector at Jefferson Laboratory. We trained multi-lyared percpetron (MLP) and boosted decision tree (BDT) models on smeared monte carlo data and compared these models to standard particle identification (PID) techniques.
Inside of the `mlp_pid` folder the MLP models of the charged and neutral particles are trained and saved using Tensorflow. The classification of particles is then carried out in the `mlp_pid.py` script where arrays of the outputs of this anlysis are saved to the `paper_plots` folder. All MLP results can be recreated in the `paper_plots.ipynb` notebook.
The BDT models are trained using XGBoost. The classification of particles is then carried out in the `bdt_pid` folder and outputs from this anlysis are saved to the `bdt_plots` folder. All BDT results can be recreated in the `bdt_plots.ipynb` notebook.
We use SHAP (SHapley Additive exPlanations) values to calcualte the feature importance of each label used in the MLP and BDT models. These plots are saved inside of the `bdt_plots` folder.
As for manual PID, all analysis is carried out in the `manual_pid.py` script in the `manual_pid` folder. This includes both dE/dx - p and timing cuts, both of which are optimized in this notebook with the optimal dE/dx - p cuts being saved to the `paper_plots` folder along with the manual PID results. Similarly, all manual PID plots can be recreated in the `paper_plots.ipynb` notebook.
The `data_processed.zip` file contains the data files that were used in the analysis while the `data_raw.zip` file contains the raw outputs from the MC simulation.
This code accompanies the paper "Particle identification in the GlueX detector with machine learning" (arXiv:2505.14706).
Files
pid_mlp.zip
Additional details
Identifiers
- arXiv
- arXiv:2505.14706
Related works
- Is new version of
- Software: https://github.com/erichabjan/pid_mlp/tree/v0.1.0 (URL)
- Is version of
- Software: https://github.com/erichabjan/pid_mlp (URL)
Software
- Repository URL
- https://github.com/erichabjan/pid_mlp
- Programming language
- Python