numine

This repository contains algorithmic and ML code to select (the mine part) neutrino interactions (the nu part) in the ICARUS liquid argon time projection chamber (LAr-TPC) imaging detector at Fermilab, using neutrinos from the BNB and NuMI beams (the numi part).

The repository covers multiple selection channels (1eNp0π, 1μNp0π, NC, Michels) and provides two complementary approaches: a CAFAna-based cut selection framework (see cafana) and a machine learning pipeline based on XGBoost BDTs (see ml).

Repository structure

numine/
├── cafana/       # CAFAna-based selections and tree makers
    └── cc1e0pi/, cc1mu0pi/, ncpi0/, ...    # per-channel selections and tree makers
└── ml/           # ML-based selections (e.g., XGBoost BDTs)
    ├── src/      # shared Python modules (preprocessing, training, evaluation, plotting)
    ├── models/   # base model definitions, if needed
    └── cc1e0pi/, ...    # per-channel notebooks and evaluations

CAFAna setup

The CAFAna-based code requires a SL7 environment, available at FNAL or CNAF. Set up sbnana:

source /cvmfs/icarus.opensciencegrid.org/products/icarus/setup_icarus.sh
setup sbnana v10_01_00 -q e26:prof

The CAFAna modules provide tools to select neutrinos, plot distributions for relevant variables (organized by true interaction type or final-state topology), and produce selection efficiency and purity curves at different cut stages, accounting for pile-up. CAFAna tree makers are also provided to flatten the CAF structure into ROOT TTrees, correctly handling systematic uncertainties.

Run any CAFAna macro with:

cafe -bq <CAFAna_Macro.C>

making sure input files (defined via TargetFile) are accessible via path, wildcard, or sam.

ML setup

The ML pipeline is Python-based and self-contained. Set up the environment and install the src package:

pip install -e ml/

Per-channel BDT trainings and evaluations live in the channel subdirectories as Jupyter notebooks. Shared utilities (feature definitions, preprocessing, training, evaluation, plotting) are in ml/src/.