Slides for AI4EIC - ePIC AI/ML Overview

ePIC AI/ML Overview October 28, 2025 Dmitry Kalinkin on behalf of the ePIC Collaboration AI4EIC Workshop, October 27-29, 2025 Electron-Ion Collider 2 Diverse energies •ep: 28 – 141 GeV •eA: 28 – 63 GeV At high luminosity •1033 – 1034 cm-2s-1 •10 – 100 fb-1 / year Various ion species from p to U Polarized e, p and light ions AI4EIC Workshop, October 27-29, 2025 ePIC Experiment 3 The detector for the EIC project. It is to deliver on EIC science goals: •Origin of Nucleon Mass & Spin •Confinement •Nucleon/Nuclear Femtography •Dense Gluon States •Beyond Standard Model AI4EIC Workshop, October 27-29, 2025 Project detector timeline Collaboration working hard towards pre-TDR (for CD-2, ~60% design readiness) ePIC Streaming Computing Model publication in preparation 4 AI4EIC Workshop, October 27-29, 2025 ePIC Software Software and Computing organized around principles of openness and collaboration with other communities in NP, HEP and CS 5 ePIC Software stack built with community components AI4EIC Workshop, October 27-29, 2025 Streaming DAQ and Computing 6 AI workflows are to play key role in accelerating science •Rapid turnaround (~2 weeks) goal for integrated Compute-Detector system with AI control •Workflows for autonomous alignment, calibration, and validation AI4EIC Workshop, October 27-29, 2025 Agentic workflows Undoubtedly, AI will be a major part of the control loop for the collider and the experiment. We need to prepare today. ePIC collaboration conducts testbeds for Streaming Orchestration that can grow to include challenges to implement AI control over the data processing. Existing tools such as PanDA WMS have initial implementations for Model Context Protocol (MCP) that can serve as basis for such experimentation. 7 AI4EIC Workshop, October 27-29, 2025 AI for simulation ePIC relies on Geant4 for simulation through DD4hep interface (TGeo-based), we are evaluating running simulations on GPUs for optical and EM physics as well as surrogate modeling Geant4 built-in parameterized model interface is ML-ready (e.g. “Par04” example for inference with CaloDit) 8 Simulation time budget at ePIC central detector Compute consumption for our monthly simulation campaigns AI4EIC Workshop, October 27-29, 2025 AI for reconstruction (SRO) Streaming readout assumes data recorded as continuous stream of time-frames Ongoing work on implementing frame → physics event building using JANA2 is to complete in 2025 We plan to hold an AI/ML challenge for developing algorithms for physics event discrimination from backgrounds. 9 AI4EIC Workshop, October 27-29, 2025 Deploying ML at ePIC Our containerized environments include Torch, Tensorflow, ONNX. ePIC Data Model using PODIO provides standardized data structures for applications in simulation, reconstruction and analysis. In ePIC, we prioritize the integration of AI/ML approaches and methods into our production workflows. This includes not only those with strong benchmark performance, but also those that scale well and are sufficiently generic for our scientific use cases – initially in simulation campaigns and later in data processing. Our reconstruction framework uses for model exchange. 16 AI4EIC Workshop, October 27-29, 2025 MLOps at ePIC ePIC will deploy more and more AI models, that will all need to be kept up to date with the latest simulations and calibrations. 17 Attribution CC BY-SA 4.0 CI pipelines Image artifacts delivered to image_browser Standartized ONNX factory for EICrecon with model training and validation on CI AI4EIC Workshop, October 27-29, 2025 Multi-Objective Optimization Bulk of work on “macroscopic” detector optimizations have been performed before ePIC. There is still a room to contribute to critical designs at a various scales. 18 dRICH and B0 optimization for ePIC by AID2E M. Diefenthaler et al 2024 JINST 19 C07001 Sci-Glass calorimeter optimization for Detector II J. Crafts et al 2024 JINST 19 C05049 Modelling of EIC Streaming Computing infrastructure Kuan-Chieh Hsu, EIC Echelon 0-1 Workshop AI4EIC Workshop, October 27-29, 2025 AI/ML activities at ePIC community ePIC does not have a dedicated AI/ML WG, unlike EICUG. Instead, WG apply AI/ML methods where appropriate. AI Town Halls support community building. 19 Two problems given: Low-Q2 tagger calibration and PID in DIRC ePIC AI/ML Hackathon @ Frascati 1st , 2nd , 3rd ePIC AI Town Hall meetings AI4EIC Workshop, October 27-29, 2025 Conclusion •ePIC will be one of the first complex detectors constructed for the AI era, we an AI strategy in place, even at this early stage •Only few highlights of the ePIC-related works can fit into one talk •Our software design and implementation support collaboration with data science, as well as AI integration within the software stack •AI/ML solutions are key to exploring its full potential towards upcoming CD-2 and TDR milestones •There are a lot of opportunities to get involved and contribute! 20