Troels Petersen
Associate Professor of Physics
Head of section for Particle & Nuclear Physics
Niels Bohr Institute, University of Copenhagen
Introduction
My research:
ATLAS physics: H to Zgamma search
ATLAS calibration (b-) jet energy estimates
IceCube reconstruction (GraphNeT project)
Various smaller projects (medical, ice cores)
Supervising:
Johann Ioannou-Nikolaides
My expertise is:
ATLAS: Electroweak- and Higgs physics, Electron and jet energy estimation with ML
IceCube: ML-based neutrino classification and reconstruction (energy & pointing)
ML: Training algorithms with input from real data, Graph Neural Networks, ML bottlenecks
A problem I’m grappling with:
How to get optimal ML performance in real data. Dealing with (highly) variable input data size, and anomaly detection in such data.
I’ve got my eyes on:
Foundation models for science data, adversarial training, and KAN-networks.
I want to know more about:
New transformer architectures, if LHC data can be used as the basis for foundation models outside LHC physics.
Projects
General searches with (GNN) AutoEncoders
Develop general Graph Neural Network (GNN) reconstruction algorithm(s) for all event types in experiment; based on this GNN architecture, use Variable AutoEncoder (VAE) to encode the dataset in a latent space (LS); apply the VAE to simulated data, also including theoretically motivated but unobserved physics signatures.
Transfer learning for jet energy scales
Include (in-situ) data information into the calibration of jet energy scales (JES) at an early ML-based stage, and improve the jet energy resolution due to considering the correlations among the variables relevant for the JES calibration and to integrating the in-situ constraints at the step of the training of the ML algorithm.
Explainable AI for Online and Transferable Learning
Develop XAI techniques for online and transfer learning applied to experimental data from physics and signal and image understanding; handle efficiently real-time massive sensor data in online and transfer learning; develop XAI techniques for high robustness and accuracy in multi-sensor environments.
