Modeling the LHCb tracking with Deep Neural Networks

This package defines the training procedure for the neural networks contributing to the parametrization of the LHCb tracking performance as a pipeline of notebooks.

The pipeline is defined as a Direct Asynchronous Graph (DAG) using Snakemake, according to the rules presented in the Snakefile.

The logical flow of the tracking parametrization is as follows:

• Evaluate whether a particle is in geometrical acceptance
• Evaluate whether a particle is reconstructed as a track, and of which track-type (long, upstream or downstream)
• Smear the Monte Carlo parameters into track parameters according to a resolution function that may depend on the kinematics of the particle
• Assess the reconstruction uncertainty on the track, encoded into the covariance matrix of the Closest-To-Beam track state, as a function of both kinematic features and measures of the track reconstruction quality (e.g. the track χ²).

The notebooks are divided in two main categories:

• Filter functions, defining the probability of a particle to survive some fixed selection, such as being in geometrical acceptance or pass the track quality requirements
• Feature functions, defining new features with arbitary distribution and therefore involving the usage of generative models (and in particular GANs).

Filter functions (acceptance and tracking efficiency)

• Data preprocessing: Preprocessing.ipynb
• Training of the acceptance model: Acceptance.ipynb
• Validation of the acceptance model: Acceptance-validation.ipynb
• Training of the efficiency model: Efficiency.ipynb
• Validation of the efficiency model: Efficiency-validation.ipynb

Feature functions

• Data preprocessing: Preprocessing-GANs.ipynb
• Training of the resolution model: Resolution.ipynb
• Validation of the resolution model Resolution-validation.ipynb
• Training of the covariance model: Covariance.ipynb
• Validation of the covariance model: Covariance-validation.ipynb

Deploying the trained models

The deployment of the trained model relies on scikinC as described in the Deploy.ipynb notebook.

Running the code

To run the code you will need to set up the environment. It is recommended to use the Docker image landerlini/lhcbaf:v0p8 defining two conda environment:

• LHCb Analysis Facility, to be used for preprocessing and validation notebooks
• TensorFlow on GPU, to be used for training

Using the docker image, the whole pipeline runs issueing

bash snakemake --resources gpu=1

from the folder containing this README.md file.