Update TrkPID to use track-based observables

.gitignore

@@ -6,6 +6,17 @@ __pycache__/
 # C extensions
 *.so
 
+# Temporary files
+*.log
+*.tmp
+*~
+
+# Image files
+*.pdf
+*.png
+*.jpg
+*.gif
+
 # Distribution / packaging
 .Python
 build/

TrkOnlyPID/.gitignore

@@ -0,0 +1,6 @@
+*.keras
+*.csv
+*.root
+*.onnx
+*.json
+figures

TrkOnlyPID/README.md

@@ -0,0 +1,35 @@
+# TrkOnlyPID
+
+## Introduction
+
+TrkOnlyPID is a machine learning algorithm is trained to differentiate between electrons and muons using information only from the tracker.
+
+## Workflow
+
+### Dataset creation
+EventNtuple data files are used as input to the training.
+The ntuples are skimmed using [make_inputs.py](make_inputs.py) which drops unnessary data as well as adds branches if any are needed.
+
+### Model training
+The python code provided in the file [TrkOnlyPIDTrain.py](TrkOnlyPIDTrain.py) is used to:
+* skim track information from the input ntuples and store the data in local csv files
+* define the neural network architecture
+* train the algorithm and save the model weights into an ONNX file named "TrkOnlyPID.onnx"
+* test the algorithm, providing performance metrics
+* generate plots to provide more information on the dataset, how the training went, and how the model perform
+
+Once the model is trained and the weights are saved in an ONNX file, this file can be used by TMVA:SOFIE to generate the inference code
+that can be used in Offline (for details about this process, check [this documentation](https://github.com/Mu2e/MLTrain/blob/main/TrkQual/README.md#converting-a-model-for-use-in-offline)).
+
+## Version history
+
+### v0 (current version)
+This version has been trained using MDC2025 datasets:
+* signal: nts.mu2e.FlateMinusOnSpill-reco-ntuple.MDC2025-004.root
+* background: nts.mu2e.FlatMuMinusOnSpill-reco-ntuple.MDC2025-004.root
+It was trained using electrons and muons generated with a flat momentum spectrum from muon stop vertices in the target.
+The input features are:
+* p(chi^2) of the track
+* tracker hit dt/dz slope divided by the expected slope from the track fit assuming an electron mass
+* N(active hits) / N(hits)
+* N(null ambiguity hits) / N(active hits)