Graph Representation Learning for Fast Detector Simulation

This project addresses the challenge of fast and accurate particle detector simulation in high-energy physics. Traditional Monte Carlo methods are highly accurate but computationally expensive, while existing fast simulation techniques often compromise on physical fidelity. The goal is to develop a model that achieves both high accuracy and significantly lower computational cost. Building on my GSoC 2025 work, I propose to extend a graph-based generative pipeline where particle jets are represented as graphs and processed using Graph Neural Networks (GNNs). The current architecture includes a Distance-based ChebNet encoder, a deterministic latent bottleneck, a physics-aware cluster decoder, and a conditional flow matching model with FiLM conditioning. A key innovation is that generation occurs entirely in a fixed 512-dimensional latent space, making computation independent of the number of particles and substantially more efficient than existing methods. For GSoC 2026, the focus is to refine this pipeline to publication quality. This involves improving key physical observables such as multiplicity, jet mass, and particle momentum using better loss functions and physics-based constraints. I will also integrate a more stable ChebNet and perform ablation studies to validate architectural choices. The project will include rigorous benchmarking against state-of-the-art models like EPiC-FM, EPiC-GAN, and PC-JeDi using metrics such as Wasserstein distance, FPND, and generation time. Deliverables include an optimized pipeline, detailed evaluation and ablation studies, an open-source implementation, and submission to venues like NeurIPS ML4PS.