Research

Parnassus: Fast Detector Simulation and Reconstruction

Synthetic datasets are essential for data analysis in all areas of particle physics. However, the detector simulation and reconstruction are significantly computationally expensive. Our group developed a novel deep learning tool — Particle-flow Neural Assisted Simulations (Parnassus) — to address this challenge.

Parnassus is a generative model, trained to perform fast simulation of reconstructed particles, using a set of stable truth particles as a condition. Based on the recent advancements of flow matching and transformer models, it accurately mimics the CMS particle flow algorithm, surpassing the existing Delphes framework.

Heavy Flavour Jet Tagging in ATLAS.

The Higgs Boson was directly observed, only in its decay modes to the Gauge Bosons (photons, W, and Z) and 3rd generation fermions. No clear, direct signs for Higgs decaying to the second generation quarks have been seen yet. Our group is developing advanced deep learning methods to improve the sensitivity to such processes.

CaloGraph - Fast Simulation for Detector Calorimeters

Simulating energy deposition in calorimeters is an essential part of event generation in high-energy physics, but traditional tools are computationally intensive. To address this, we develop fast, machine learning–based alternatives.

CaloGraph is a graph-based generative model for simulating calorimeter showers. It uses a diffusion model to generate realistic showers in a geometry-aware way, enabling efficient simulation in detectors with irregular geometries (e.g. ATLAS) and supporting large-scale data generation.

Track reconstruction in noisy environment

Accelerating Tracking with Symbolic Regression

Tracking charged particles through a detector is a key part of understanding what happened in a particle collision. Graph neural networks are a natural way to model this task because they can capture complex relationships between hits in the detector. However, these models are often too slow to use in real-time applications, such as trigger systems, where decisions are made within microseconds. In this work, we use symbolic regression to find simple mathematical expressions that approximate the behaviour of the graph network. This allows us to keep the advantages of graph-based approaches while making the model much faster and easier to run on hardware.

COCOA

An open calorimeter simulation for AI

Three different jet species projected onto 2-dimensional representation of learned contrastive space

Representation learning for detector data

Using self-supervised learning for rich, generic representations of jets

Graph representation of calorimeter and track data

Particle reconstruction

From detector readout to particles with deep learning

Super-resolution for calorimeters

Going beyond granularity using conditional generation

VHcc Analysis: Measuring the Higgs to Charm Coupling in ATLAS

Until we clearly observe the Higgs Boson decay to second generation particles, one would not get a clear evidence for the role of the BEH mechanism in giving mass to matter particles. We are joining effort with TAU in developing an analysis to measure the Higgs coupling to Charm Quarks via the associate production of the Higgs with a Vector Boson (W,Z).

Particle Flow Algorithms

We use machine learning to seperate overlapping energy deposits in the calorimeter into different particles.