Introduction

This repository contains the code necessary for constructing a muon neutrino selection at the ICARUS experiment using the outputs of a machine learning reconstruction (see lartpc_mlreco3d). The selection is implemented using the CAFAna framework (see sbnana), which takes its name from the Common Analysis Files (CAFs) that contain the input analysis-level informatino necessary for developing a selection.

CAFs

CAF Format

The CAF format employs a hierarchical organization of C++ objects stored in ROOT files (traditionally using .caf.root as an extension). At a basic level, the hierarchy follows something similar to the below:

• StandardRecord - The top-level object representing a single "spill" or event.
  • SRSlice (slc) - A single reco interaction or "slice" as defined by Pandora (vector).
  • SRInteractionDLP (dlp) - The reco interactions as defined by the ML reconstruction (vector).
    • SRParticleDLP (particles) - The reco particles comprising the interaction (vector).
  • SRInteractionTruthDLP (dlp_true) - The true interactions as defined by the ML reconstruction (vector).
    • SRParticleTruthDLP (particles) - The true particles comprising the interaction (vector).

The CAF format contains many nested layers, and so is often considered to be a bit unwieldy or slow to use directly - a cost associated with navigating the complex structure. For this reason, a "flattened" version of the files is often used instead (traditionally using .flat.root as an extension) which broadcasts all branches to match the deepest level. This greatly simplifies the navigation and results in a significant speed up for any framework using them as input (e.g. CAFAna). The flattening is performed by an executable that ships with sbnana called flatten_caf.

Generating CAFs

The analysis-level output of the machine learning reconstruction is stored in the HDF5 format. The advantage of the HDF5 format is its portability and the "self-describing" nature of the dataset format. The disadvantage is that it requires a little bit of work to get the analysis outputs into a CAF file. This functionality is implemented in sbn_ml_cafmaker and won't be described in great detail here. At a basic level, it reads the HDF5 input and organizes the truth and reco information in the new branches dlp_true and dlp within the StandardRecord. The resulting CAFs have been verified to work with CAFAna directly (albeit with some reduced functionality) and are able to be flattened using the flatten_caf executable.