Paper Spotlight

Point cloud-based diffusion models for the Electron-Ion Collider

This paper uses machine learning to advance full event generation at the EIC through a novel diffusion model that combines point cloud representation with transformer modules to generate all particle species while preserving physical conservation laws. The ML-based approach significantly outperforms previous generative models across all evaluation metrics and demonstrates potential for broader applications in collider phenomenology 

Design and simulation of a SiPM-on-tile ZDC for the future EIC, and its performance with graph neural networks

This NIM A article presents a novel Zero-Degree Calorimeter design and uses Graph Neural Networks to dramatically improve its energy/angle resolution. The GNN-based reconstruction meets or exceeds the stringent requirements from the EIC Yellow Report, showcasing how cutting-edge deep learning can optimize calorimeter performance even with complex detector geometries (an important 

Deep (er) reconstruction of imaging Cherenkov detectors with swin transformers and normalizing flow models

This work presents Deep(er)RICH, an advanced deep learning approach for fast and accurate particle identification in DIRC Cherenkov detectors.
It enhances the original DeepRICH framework using Swin Transformers and normalizing flows to handle complex hit patterns and simulations.
The method enables near real-time reconstruction and simulation, addressing key bottlenecks in PID for experiment

The optimal use of segmentation for sampling calorimeters

This work investigates the impact of calorimeter segmentation on energy reconstruction using deep neural networks that process calorimeter data as point clouds. Applied to a simulated forward calorimeter system for the ePIC detector at the future Electron-Ion Collider, the study shows that fine longitudinal segmentation is essential for achieving sub-10% energy resolution for isolated charged pions. These findings offer a key benchmark for

Towards a RAG-based Summarization Agent for the Electron-Ion Collider

This work introduces RAGS4EIC, a Retrieval Augmented Generation (RAG)-based AI system designed to help researchers at the Electron-Ion Collider efficiently navigate and summarize vast experimental documentation. By combining a vector database with a large language model, the system delivers concise, citation-enriched summaries in response to user queries. Built on LangChain and evaluated with dedicated RAG assessment tools, thi

Diffusion model approach to simulating electron-proton scattering events

In this paper, authors use machine learning to advance electron-proton scattering event simulation at the Electron-Ion Collider through a diffusion model with U-Net architecture, achieving high-quality generation of sparse collider events with good agreement on kinematic observables and momentum conservation. The ML-based approach outperforms traditional simulation methods by effectively handling the unique kinematics of ele

Design and simulation of a SiPM-on-tile ZDC for the future EIC, and its performance with graph neural networks

This NIM A (2024) article presents a novel Zero-Degree Calorimeter design and uses Graph Neural Networks to dramatically improve its energy/angle resolution. The GNN-based reconstruction meets or exceeds the stringent requirements from the EIC Yellow Report, showcasing how cutting-edge deep learning can optimize calorimeter performance even with complex detector geometries (an important 

ELUQuant: event-level uncertainty quantification in deep inelastic scattering 

This work introduces a physics-informed Bayesian neural network with flow-based posterior approximations to achieve detailed event-level uncertainty quantification in high energy physics. Applied to deep inelastic scattering events, the model accurately extracts key kinematic variables while capturing both aleatoric and epistemic uncertainties, enabling more informed decision-making in tasks like event filtering.  Its 

Machine learning-based jet and event classification at the Electron-Ion Collider with applications to hadron structure and spin physics

This paper uses machine learning to advance jet flavor classification at the Electron-Ion Collider (EIC), demonstrating that ML algorithms significantly outperform traditional observables like jet charge for identifying quark types. The study shows ML-based classification provides crucial insights for EIC detector design, particularly highlighting that particle i

AI-assisted optimization of the ECCE tracking system at the Electron Ion Collider 

The work demonstrates AI-driven multi-objective optimization to design the EIC’s tracking detector. The approach navigates a complex multidimensional parameter space (performance, cost, etc.) to maximize tracker efficiency within engineering constraints. This makes EIC one of the first large experiments to incorporate deep learning into detector design from the outset, improving the detector’s overall performance.