Skip to content

Latest commit

 

History

History
91 lines (48 loc) · 3.95 KB

CNN1D.md

File metadata and controls

91 lines (48 loc) · 3.95 KB
Raw

CNN for Jet Constituents

This project implements a Convolutional Neural Network (CNN) designed to analyze jet constituents in high-energy physics datasets. The goal is to perform classification using jet particle features and explore how well CNNs can distinguish between different jet structures by leveraging constituent-level information. The project is implemented using TensorFlow and utilizes particle feature datasets stored in HDF5 format.

Table of Contents

  1. Part 1: Dataset Preparation and Exploration
  2. Part 2: Model Architecture
  3. Part 3: Training the Model
  4. Part 4: Evaluation and Analysis
  5. Part 5: Conclusion

Part 1: Dataset Preparation and Exploration

  • Dataset:

    The jet constituent data is stored in .h5 files within the JetDataset directory. Each dataset contains information for 50,000 jets, with up to 100 particles per jet. For each particle, 16 features are provided, including momentum components (px, py, pz), energy, and relative angular information.

  • Exploration:

    We explored basic dataset properties, including particle multiplicity per jet, momentum distribution, and energy patterns. This helped us understand the overall dataset characteristics before feeding it into the CNN.

Part 2: Model Architecture

The CNN architecture used for jet constituent classification includes:

  • Input Layer:

    The input shape is (100, 16), corresponding to 100 particles and their 16 features.

  • Convolutional Layers:

    Several 1D convolutional layers are applied to the constituent-level feature matrix, capturing patterns across the particle feature space.

  • Pooling Layers:

    Max pooling is used to down-sample the data and reduce computational complexity while retaining the most important information.

  • Dense Layers:

    After flattening, the convolutional layers are followed by fully connected (dense) layers, leading to a softmax output for classification.

  • Activation Functions:

    ReLU activations are used in convolutional and dense layers. The final layer uses a softmax function for multi-class classification.

Part 3: Training the Model

  • Data Augmentation:

    To enhance model generalization, various data augmentation techniques were applied, including random rotations and translations of jet constituents.

  • Loss Function:

    The categorical cross-entropy loss function was employed to measure the model's classification performance.

  • Optimizer:

    Adam optimizer was used with a learning rate of 0.001 for faster convergence.

  • Batch Size and Epochs:

    The model was trained using a batch size of 64 for 50 epochs, with early stopping to prevent overfitting.

Part 4: Evaluation and Analysis

  • Validation:

    The model's performance was evaluated using a validation set. Metrics such as accuracy, precision, recall, and F1-score were computed.

  • Confusion Matrix:

    A confusion matrix was generated to visualize the model's classification performance across different jet classes.

  • Visualization:

    Plots of training vs. validation loss and accuracy were generated to ensure that the model is not overfitting and that training proceeds smoothly.

Part 5: Conclusion

The CNN-based approach for jet constituent classification shows promising results, achieving a validation accuracy of over 85%. Future improvements may include experimenting with deeper network architectures and applying advanced techniques like attention mechanisms to better capture the hierarchical structure of jets.

References

  • Jet Dataset:

    The dataset used for training is available in the public HDF5 format.

  • TensorFlow Documentation:

    Refer to the official TensorFlow documentation for more details on the framework and tools used.