DeeLeMa (Deep Learning for Mass estimation)

Overview

$\textsf{DeeLeMa}$ is a deep learning network designed to analyze energies and momenta in particle collisions at high-energy colliders. Built with a foundation on symmetric event topology, $\textsf{DeeLeMa}$'s generated mass distributions demonstrate robust peaks at the physical masses, even after accounting for combinatoric uncertainties and detector smearing effects. With its adaptability to different event topologies, $\textsf{DeeLeMa}$'s effectiveness shines when corresponding kinematic symmetries are adopted.

The current version of $\textsf{DeeLeMa}$ (v1.0.0) is constructed on the $t\bar{t}$-like antler event topology which is shown below figure.

$t\bar{t}$-like antler event topology

Requirements

Using Pip

pip3 install -r requirements.txt

Using UV (Recommended)

If you haven't installed uv yet:

pip3 install uv

With UV installed:

# Create virtual environment (venv)
uv venv

# Sync dependencies
uv sync requirements.txt

# Activate virtual environment (venv)
source .venv/bin/activate

Getting Started

1. Clone the Repository

   git clone https://github.com/Yonsei-HEP-COSMO/DeeLeMa.git

2. Install Dependencies:

   Follow the Requirements section for instructions.

3. Training:

   ⚠️ Caution

   Before training, ensure you modify the data path in train.py to point to the location of your data. For more details, refer to train.py.

   To train the model, execute the following command:

   python train.py

4. Monitoring:

   To monitor the training process, run tensorboard:

   tensorboard --logdir=logs/

   ⚠️ Caution

   If you use huak then should run tensorboard in activated virtual environment.

5. Testing:

   • Load the saved checkpoint using the load_from_checkpoint() method:

     checkpoint_path = "DeeLeMa_Toy.ckpt"
     model = DeeLeMa.load_from_checkpoint(checkpoint_path)

   • Set the model to evaluation mode:

     model.eval()

   • Use the loaded model for inference or further analysis:

     from deelema.utils import decode_missing_momentum
     
     output = decode_missing_momentum(model, dl_test, m_C) # m_C is the pre-determined mass

Citation

If $\textsf{DeeLeMa}$ benefits your research, please acknowledge our efforts by citing the following paper:

@article{Ban:2023mjy,
    author = "Ban, Kayoung and Kang, Dong Woo and Kim, Tae-Geun and Park, Seong Chan and Park, Yeji",
    title = "{Missing information search with deep learning for mass estimation}",
    doi = "10.1103/PhysRevResearch.5.043186",
    journal = "Phys. Rev. Res.",
    volume = "5",
    number = "4",
    pages = "043186",
    year = "2023"
}

Reference

• K. Ban, D. W. Kang, T.-G. Kim, S. C. Park, and Y. Park, Missing Information Search with Deep Learning for Mass Estimation, PhysRevResearch.5.043186

License

$\textsf{DeeLeMa}$ is released under the MIT License. For more details, see the LICENSE file in the repository.

5. Loading a Trained Model:

   • Load the saved checkpoint using the load_from_checkpoint() method:

   checkpoint_path = "DeeLeMa_Toy.ckpt"
   model = DeeLeMa.load_from_checkpoint(checkpoint_path)

   • Set the model to evaluation mode:

   model.eval()

   • Use the loaded model for inference or further analysis:

   predictions = []
   with torch.no_grad():
       for batch in dl_test:
           outputs = model(batch)
           predictions.append(outputs)

Citation

If $\textsf{DeeLeMa}$ benefits your research, please acknowledge our efforts by citing the following paper:

@article{Ban:2023mjy,
  author = "Ban, Kayoung and Kang, Dong Woo and Kim, Tae-Geun and Park, Seong Chan and Park, Yeji",
  title = "{Missing information search with deep learning for mass estimation}",
  doi = "10.1103/PhysRevResearch.5.043186",
  journal = "Phys. Rev. Res.",
  volume = "5",
  number = "4",
  pages = "043186",
  year = "2023"
}

Reference

K. Ban, D. W. Kang, T.-G. Kim, S. C. Park, and Y. Park, Missing Information Search with Deep Learning for Mass Estimation*, PhysRevResearch.5.043186

License

$\textsf{DeeLeMa}$ is released under the MIT License. For more details, see the LICENSE file in the repository.