topsbi

This repository uses the simulation based inference techniques outlined by J. Brehmer et al by training a DNN binary classifier between two points in Wilson coefficient space under the Standard Model effective field theory hypothesis. This classifier can be converted to a likelihood ratio between these two points which can then be used for a binned or unbinned physics analysis.

Installation

This code uses PyToch to for neural network training. As outlined in the PyTorch instalation instructions, Conda is no longer supported. Using the current environment.yml to install via Conda or Mamba will install the last Conda-supported PyTorch version (2.5.1). To install current versions of PyTorch and Cuda, use the corresponding pip command for your GPU. These instructions will cover the system architecture of camlnd.crc.nd.edu.

Start by cloning the repository.

git clone https://github.com/emcgrady/topsbi.git
cd topsbi

Then instal and activate the repository Conda/Mamba environment.

mamba env create -f environment.yml
mamba activate topsbi

Instal Conda for GPU use (Optional) Use the instructions found in the PyTorch documentation for your OS and GPUs. For camlnd.crc.nd.edu, use the following command.

pip3 install torch torchvision torchaudio

The validation framework uses topcoffea as a baseline to compare to when reweighting. In any directory, topcoffea can be installed by the following commands.

git clone https://github.com/TopEFT/topcoffea.git
cd topcoffea
pip install .

Network Training

All of the training is done through train.py which takes a single argument, a path to a configuration yaml. Examples of these yaml files can be found in the examples directory. As the fileds in these configuration files will be updated regulary as various features are added, a serpate README can be found in this directory.

There are two cases for netwrok training, quadratic and other, which are determined by the configuration fields signalTrainingTerm and signalTrainingPoint respectively. To train a discriminator for a quadratic field, put null in the signalTrainingPoint field and put the name of the quadratic term (i.e. sm, ctq8, cdq1, etc.) in the signalTrainingTerm field. Conversely, to train any other term, put null in the signalTrainingTerm field and a point in WC space (i.e. ctq8=2.5) in the signalTrainingPoint field. Note that for learning any cross-terms, the values for both BSM fields must be given.

Future Development Plans

• Add quantitative measure to LR ROC recreation
• Network combination