As part of the CMS Anomaly Search Effort (CASE), my current research involves developing advanced unsupervised machine learning architectures aimed at significantly enhancing anomaly detection capabilities. Given that signatures of new physics could manifest anywhere within the multidimensional collider data phase space, our goal is to instead equip our ML models with a deep understanding of what the Standard Model (SM) looks like. By training such an AI model to accurately recognize and characterize SM phenomena, we effectively prepare them to flag any deviations, which could be potential indicators of new physics, without relying on predefined Beyond Standard Model (BSM) Physics models. This inherently model-agnostic approach makes unsupervised anomaly detection particularly powerful, enabling simultaneous searches for rare or unexpected signals across multiple final states, kinematic regimes and event topologies within LHC collision data.
Either 1's or 0's, or maybe both? In collaboration with Imperial College and Durham University, I worked on the development of novel encodings for particle physics data - usually jets produced at the LHC, onto a quantum circuit. Quantum computers differ fundamentally from classical computers through their use of quantum bits, or qubits, instead of classical bits. While classical bits store information as binary states (0 or 1), qubits exist as a superposition of two states, which is to say, both 0's and 1's. Once this data has been encoded, one can perform, in a sense, machine learning on a system of qubits using quantum mechanical operations such as superposition, entanglement and parameterized (trainable) rotations. I also work on the development of novel quantum algorithms that can be used for classification and anomaly detection.
Pileup at the Large Hadron Collider (LHC) refers to the phenomenon where multiple proton-proton collisions occur simultaneously within a single detector readout, complicating the reconstruction and analysis of collision events. These overlapping interactions create additional noise and background signals, posing significant challenges for accurately identifying particles and measuring their properties. In the past, I worked on mitigating the effects of pileup using Graph Attention (GAT) based neural networks. I also supervised a bachelor thesis at KIT to explore the possibility of distilling knowledge from a sophisticated GAT model into a simpler deep neural network (DNN), enabling comparable performance with significantly reduced computational complexity.