DIV-1 · Divanshu Sharma

Work / Fig. 07 · Scientific ML · ML4SCI / CERN CMS · 2024

Deep Learning for CMS Detector Physics

80% AUC — electron-vs-photon classification, custom ResNet-15.

Fig. 07

Deep Learning for CMS Detector Physics

Scientific ML · ML4SCI / CERN CMS · 2024

Problem

Event reconstruction at the CMS detector needs classifiers that separate electrons from photons and quark from gluon jets on raw detector images — where physics-blind architectures plateau.

Approach

Developed a custom ResNet-15 CNN on raw CMS detector data; implemented a graph convolutional network for particle momentum regression that matched a heavier GAT baseline's ROC curve; applied targeted VGG-style CNNs to jet classification.

Results

  • 80% AUC

    electron-vs-photon classification, custom ResNet-15

  • −30% training time

    GCN matching the GAT baseline for momentum regression

  • +7% jet classification

    quark/gluon separation via targeted VGG-style CNNs