Scholarlink

<p>We investigate the performance of various deep learning architectures for classifying Beyond the Standard Model (BSM) events against Standard Model (SM) backgrounds using simulated high-energy physics data. Four models were evaluated: a Deep Neural Network (DNN), a Convolutional Neural Network (CNN) on 2D image-like data constructed from tabular data, TabNet, and a hybrid Autoencoder + TabNet pipeline. The data were pre-processed through dimensionality reduction and optimized via Optuna for architecture-specific hyperparameters. All four models achieved high classification performance, with AUC values exceeding 0.998. The CNN and hybrid models performed best in both standard metrics and in signal efficiency vs. background rejection analysis. In particular, the Autoencoder + TabNet model achieved the highest accuracy and F1 score, highlighting the effectiveness of combining unsupervised feature extraction with attention-based decision-making. These results demonstrate the applicability of deep learning, especially hybrid and interpretable architectures, for enhancing event classification in collider-based BSM searches.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Beyond the background: deep learning models for BSM signal discovery

  • 期刊论文

  • 发表日期：

    2025年9月26日

查看全文
  • Ali Celik,
  • Aslıhan Aybar

摘要

We investigate the performance of various deep learning architectures for classifying Beyond the Standard Model (BSM) events against Standard Model (SM) backgrounds using simulated high-energy physics data. Four models were evaluated: a Deep Neural Network (DNN), a Convolutional Neural Network (CNN) on 2D image-like data constructed from tabular data, TabNet, and a hybrid Autoencoder + TabNet pipeline. The data were pre-processed through dimensionality reduction and optimized via Optuna for architecture-specific hyperparameters. All four models achieved high classification performance, with AUC values exceeding 0.998. The CNN and hybrid models performed best in both standard metrics and in signal efficiency vs. background rejection analysis. In particular, the Autoencoder + TabNet model achieved the highest accuracy and F1 score, highlighting the effectiveness of combining unsupervised feature extraction with attention-based decision-making. These results demonstrate the applicability of deep learning, especially hybrid and interpretable architectures, for enhancing event classification in collider-based BSM searches.