Dedicated AI/ML Accelerator Test Facility

15 Jul 2026, 08:30 → 17 Jul 2026, 18:00 America/Los_Angeles

48/1-112A/B/C/D - Redwood A/B/C/D (SLAC)

Auralee Edelen (SLAC), Brendan O'Shea (SLAC)

Registration is free. Hybrid will be available.

There is strong and growing interest across the accelerator and AI/ML community in an accelerator facility purpose-built and dedicated to AI/ML testing and operations, that could be called an AI Test Facility (AITF).

 

The impact of AI/ML on accelerator-based science has been well documented in numerous community reports in recent years. Yet a gap remains: no document has articulated what would be required to realize a fully AI/ML-integrated accelerator facility. What does the control system architecture look like? What computing infrastructure, both local and HPC, is needed? How should data acquisition, metadata tagging, and system configurability be designed? What types of access best serve the diverse communities that would use such a facility? Furthermore, accelerators are among the most complex engineered systems in science, making them an ideal environment for implementing AI/ML at scale and advanced algorithm testing.

 

We are also interested in examining the needs of fusion systems and any potential overlap with the needs of accelerators.

 

With such a broad range of topics to cover, the input needed to design a successful AITF extends to a wide variety of important stakeholders. A successful AITF will draw on the expertise of operations staff, accelerator physicists, data/computing and control systems infrastructure experts, facility users and the broader scientific community, academic researchers in AI/ML, robotics, controls, and related fields.

 

The time is right to develop a public-facing white paper that articulates the role, need, and requirements for a dedicated AI/ML test facility. The goal of this workshop is to produce a document that the broad community can endorse. A document that will serve as a reference and roadmap for the accelerator community at large.

Sessions:

Operation of scientific systems: Accelerators and Tokamaks

Operation of scientific systems: User Challenges and Needs

AI/ML Applications: AI/ML and Accelerators

Artificial Intelligence and Machine Learning: Complex Systems

Important dates 

April 14, 2026	Registration Opens
July 15/16/17, 2026	Workshop

Request for submission AI Workshop.pdf

Wednesday 15 July

Discussion: Welcome and Introduction

Day2_intro.pptx

Day 3 intro.pptx

Folder for Vision Statements

1 Intro

Day1_intro.pptx

Operation of scientific systems: Accelerators and Tokamaks

2 AI/ML Needs for a Commercial EUV Light Source

xLight Inc. is developing an accelerator-driven extreme ultraviolet (EUV) light source for the semiconductor manufacturing industry. This machine must operate autonomously, a goal we believe will be achieved through AI/ML techniques. In this talk, we offer an industry perspective on the proposal for a dedicated AI/ML accelerator test facility. We share some of the unique challenges in deploying a commercial EUV light source that AI/ML techniques may address, and suggest capabilities that would help such a facility address the needs of commercial partners. We end with a high-level picture of how AI/ML fits into the project of developing a commercial particle accelerator to revolutionize the chip manufacturing industry.

Speaker: Christopher Pierce (xLight Inc.)

Christopher Pierce SLAC Dedicated AIML Test Facility.pdf

3 Online Deployment and Operational Experience of ML-Based Beam Tuning at the ATLAS Heavy Ion Linac

Developing a multifaceted ML tool for day-to-day operation of a particle accelerator requires close collaboration with the machine operators to incorporate their experience and feedback. The tool must also ensure that all safety protocols and operational redundancies established by the operations team are fully integrated. In the initial part of this talk we briefly present the framework of an ML tool currently in use and its supporting control system architecture. Details on the underlying Bayesian Optimization will be highlighted for tuning various sections of the accelerator along with recent implementation of the capability to use prior experiment data for any beamline of the ATLAS accelerator significantly improving the tuning procedure. This has been made possible by recent availability of complete ATLAS database including information of beam current, element settings, operator notes associated with every experiment going back to 2005. Improvements made to the tool based on extensive accelerator operator feedback during its use over the past year will be discussed. Requested features by the operators such as the capability for sequential simultaneous tuning of multiple beamlines will be discussed. With the recent interest in increasing operational beam current at ATLAS, various pain points of the tool in this high-risk operating condition are presented. The overall evolution of the tool from its initial PRO version to the currently implemented LITE version for rapid use by operators will be discussed.

Speaker: Adwaith Ravichandran (Argonne National Laboratory)

SLAC_Workshop_Slides_7_14_2026.pptx

4 Outlook for AI/ML tools for SNS

This talk will highlight the operational challenges within the SNS complex. We will cover facility priorities, as illustrated by on-going efforts to implement ML solutions for beam control and facility health. We also describe capabilities that we expect to have benefits and transferability to SNS operations.

Speaker: Dr Sasha Zhukov (Oak Ridge National Laboratory)

ML_SNS.pptx

5 Practical Requirements for AI in Accelerator Operations

AI and machine learning tools have the potential to significantly improve operational efficiency, but there are important challenges standing in the way of their deployment and widespread adoption. Rather than focusing on new AI algorithms or applications, this discussion will examine what SLAC’s operators have learned through early experience with these technologies: where AI can provide meaningful value, where it falls short of operational needs, and what capabilities are required for operators to develop confidence in AI-assisted workflows. The goal is to offer an operations-centered perspective on how AI tools should be designed, integrated, and deployed to become practical, trusted components of future accelerator facilities.

Speaker: Benjamin Ripman (SLAC)

Practical_Requirements_for_AI_in_Accelerator_Operations.pdf

Practical_Requirements_for_AI_in_Accelerator_Operations.pptx

6 Genesis Activities and MOAT

Speaker: Jean-Luc Vay (Lawrence Berkeley National Laboratory)

FACET_AI_Vay.pptx

10:45 Coffee

Operation of scientific systems: Accelerators and Tokamaks

7 ILLUMINE, Agents, and the Path to Closed-Loop Control at LCLS

At LCLS, delivering a stable X-ray beam to the experiment still depends on skilled operator time across a long chain of accelerator and beamline systems; and with LCLS-II-HE, manual setup is becoming harder to sustain. At MFX we are building autoMFX around three pieces that are starting to work together: BES-funded ILLUMINE for autonomous photon-beam tuning; language-model assistants that let operators ask questions and issue goals in natural language while tracking how upstream changes affect what the detectors actually measure; and online analysis workflows so tuning is judged against diffraction and spectroscopy quality, not instrument controls readbacks alone.

We have been exercising this stack during real operator shifts at MFX—testing agent tools for device access, shift logging, and analysis setup alongside ILLUMINE tuning runs. With MFX instrument colleagues we are now aiming for a July 2026 demonstration in which agents help carry beam delivery, alignment, geometry, and online crystallography analysis, keeping expert operators in the loop but spending less time on repetitive steps.

What is still difficult on scarce user beamtime is closing the loop between the accelerator and the photon side: fusing source, beamline, and science-side signals, and trying control strategies in a digital twin before actuating hardware. An AI/ML accelerator test facility is where we would iterate on that bridge: controlled detune-and-recover tests, shared controls across the delivery chain, and staged rollout from shadow operation to bounded closed-loop control, before scaling to MFX and LCLS-II-HE.

Speaker: Frederic Poitevin (SLAC)

Discussion: Operations and Users

Day2_intro.pptx

Day 3 intro.pptx

Folder for Vision Statements

12:45 Lunch

Operation of scientific systems: User Challenges and Needs

8 AI for Neutron Sciences

We will discuss the data pipeline that supports user experiments at neutron scattering facilities, and how AI workflows can help accelerate science. AI can play a role in accelerating most areas of our data pipeline. While numerous efforts are ongoing to address individual needs, we will focus on the common infrastructure that an AI-enabled user facility requires and the current open questions to develop it.

Speaker: Mathieu Doucet (Oak Ridge National Laboratory)

9 Bayesian Exploration and Surrogate Emulation of Nonlinear Beam-Response Geometry in the LBNF Beamline

Next-generation long-baseline neutrino experiments aim to achieve multi-MW proton beam power while reducing accelerator-induced systematic uncertainties. At Fermilab, the LBNF beamline is designed for 1.2 MW operation with PIP-II and is upgradeable to 2.4 MW. DUNE will probe the three-flavor neutrino paradigm and search for CP violation, requiring precise neutrino-flux normalization and improved control of accelerator-related uncertainties.
Within the LBNF beamline, the System for On-Axis Neutrino Detection (SAND) will constrain flux uncertainties using precision near-detector measurements, while the Muon Monitor System (MuMS) will provide beamline diagnostics sensitive to the proton beam, target, and horn configuration. However, the pion phase space relevant for DUNE depends simultaneously on many correlated parameters, including beam centroid, beam width, horn current and alignment, target position, optics shifts, and radiation-induced changes. Consequently, MuMS observables exhibit nonlinear and coupled responses that are difficult to characterize using traditional one-parameter scans.
To address this challenge, we are developing a Bayesian Exploration framework coupled to physics-informed surrogate emulators trained on Geant4 beamline simulations. Gaussian-process emulators provide both fast predictions and uncertainty estimates, enabling adaptive selection of new simulation points in beam-parameter space. As an initial demonstration, we construct surrogate emulators for MuMS response observables using a verified simulation campaign spanning proton-beam steering conditions. The emulators reproduce the simulated dependence of MuMS centroid and gradient observables while providing predictive uncertainties, and serve as the foundation for future multidimensional exploration including beam width, horn current, and additional beamline parameters.
This work establishes a framework for uncertainty-aware beam monitoring, adaptive simulation campaigns, and rapid beam-response inference for future DUNE operations.

Speaker: Sudeshna Ganguly (Fermi National Accelerator Laboratory)

facet-ii-SG-talk.pdf

10 Bluesky and Photon Beamline Integration Challenges in AI/ML‑Driven Experimental Facilities

As AI/ML‑enabled workflows become increasingly central to accelerator and photon‑science experiments, the need for flexible, metadata‑rich, and automation‑friendly control systems has grown substantially. Bluesky provides a modern Python‑based framework for experiment orchestration, data acquisition, and real‑time feedback, making it a strong candidate for integration with next‑generation AI/ML accelerator test facilities. This contribution will outline Bluesky’s architecture and its role within photon‑science operations, including its capabilities for streaming data, coordinated device control, and enabling adaptive experimentation.

The talk will highlight practical integration challenges observed at photon beamlines: managing heterogeneous hardware interfaces, enabling continuous scanning, scaling data workflows for ML‑based analysis, and aligning metadata and data management with facility‑level requirements. We will discuss strategies that have proven effective in real deployments, as well as opportunities for deeper coupling between Bluesky and dedicated AI/ML accelerator infrastructures to support autonomous experimental pipelines.

Speaker: Max Rakitin (National Synchrotron Light Source II, Brookhaven National Laboratory)

FACET-II-AI-mrakitin-20260715.pdf

11 AI-Driven Advances Toward Intelligent Tokamak Operation

As fusion devices move toward reactor-scale conditions, robust, real-time plasma control is essential for sustaining high-performance operation and tight stability margins. This talk presents recent AI-driven control advances on the DIII-D and KSTAR tokamaks, highlighting successful demonstrations of RMP-induced ELM control [1], preemptive tearing mode suppression [2], and radiation-front-based detachment control [3,4]. To make these control architectures resilient for Future Pilot Plants (FPP), we introduce foundational AI-driven approaches that further strengthen system reliability against diagnostic degradation and actuator failures. Advanced plasma monitoring tools like Diag2Diag [5] and real-time kinetic equilibrium reconstruction (rt-CAKE) [6] provide fault-tolerant virtual diagnostics and continuous state evaluation. Alongside robust plasma monitoring, efficient actuator coordination is equally crucial. Complementing this need, neural network surrogates such as TORBEAM-NN [7] enable fast, scalable, real-time electron cyclotron heating (ECH) optimization, even during gyrotron loss. Together, integrating these high-fidelity diagnostics and adaptive actuator models establishes a crucial pathway toward intelligent, failure-tolerant tokamak operation.

Speaker: SangKyeun Kim (Princeton Plasma Physics Laboratory)

15:45 Coffee

Discussion: Report - Operations and Users

Day2_intro.pptx

Day 3 intro.pptx

Folder for Vision Statements

Discussion: Report Writing

Day2_intro.pptx

Day 3 intro.pptx

Folder for Vision Statements

Thursday 16 July

Discussion: Report Discussion

Day2_intro.pptx

Day 3 intro.pptx

Folder for Vision Statements

AI/ML Applications: AI/ML and Accelerators

12 Experience with Digital Twin Deployment on the CLARA Facility

High-fidelity virtual replicas of accelerator facilities, otherwise known as digital twins, are becoming increasingly desirable tools that many labs are seeking to add to their arsenal. Numerous hurdles can prevent the successful integration of a digital twin into the day-to-day operation of the machine, including control system integration, model fidelity, and processing speed, not to mention the generalizability of any specific solution for a given accelerator. The CLARA facility at Daresbury Laboratory has recently deployed a prototype digital twin using the JANUS framework, which seeks to provide a general-purpose digital twin solution for accelerators. This contribution outlines the architecture of the framework, the steps taken to connect the tool with the real control system, and the work required to deliver a fully integrated, generic digital twin. A dedicated AI/ML test facility could be pivotal for assessing deployment techniques needed for wider implementation.

Speaker: Matthew King (ASTeC, Daresbury Laboratory, Science and Technology Facilities Council)

Deployment of Digital Twin at CLARA - FACET-II Workshop.pdf

Deployment of Digital Twin at CLARA - FACET-II Workshop.pptx

13 AI-Assisted Accelerator Operation at the Advanced Photon Source

Artificial intelligence and machine learning are becoming increasingly practical tools for accelerator operation and diagnostics. This talk will present an overview of ongoing AI/ML activities at the Advanced Photon Source. Topics include differentiable simulation methods for beam sigma matrix reconstruction, hysteresis-aware modeling and control, anomaly detection in power-supply and vibration systems, and retrieval-augmented tools for operational knowledge extraction. These activities demonstrate a broader direction toward integrating AI/ML with accelerator physics expertise to enable more robust diagnostics, improved operational awareness, and more automated workflows for accelerator operation.

Speaker: Chenran Xu (Argonne National Laboratory)

2026-07-16-AITestFacility-APS.pdf

14 ML Applications and Experiences at BNL hadron injector complex

The BNL Hadron Injector Complex serves as an active testbed for applying AI/ML techniques to accelerator operations, including beam tuning and optimization, predictive modeling, anomaly detection, virtual diagnostics, and operator decision support. Here we provide an overview of ongoing AI/ML efforts at BNL and the practical experience gained from integrating machine learning with accelerator control systems and operational workflows. We will discuss challenges and lessons learned in data access, model deployment, real-time control interfaces, and operational reliability, as well as our involvement in the ongoing NARAD project, which is developing community standards for accelerator data access and control system information models. These experiences help identify the infrastructure, software, and interoperability requirements needed to enable robust AI/ML deployment at future accelerator facilities and dedicated AI/ML test stands.

Speaker: Weijian Lin (Brookhaven National Laboratory)

WLin_FACET-II_talk.pdf

WLin_FACET-II_talk.pptx

15 First Experimental Demonstration of Machine Learning-Based Tuning on the PSI Injector 2 Cyclotron

We present the first live deployment of reinforcement learning for closed-loop control of an operational cyclotron, demonstrated at the PSI Injector-2 during a 12-day beam-development campaign. The agent controlled multiple magnetic and RF actuators using phase, loss, and current diagnostics to minimize phase deviations and radial losses directly on the machine.
For a fixed operating configuration, real-machine training converged within a few hours, and surrogate-based pretraining reduced online training time substantially. The campaign also showed that policies are highly configuration-specific: transfer between nearby turn numbers was limited, consistent with measured changes in machine sensitivity. Despite this, the agent restored final beam phases to within about ±1° of target, significantly reduced losses, and maintained safe operation under interlock-aware training. Overnight evaluation further demonstrated drift compensation and recovery from deliberate perturbations. The results provide a first validation of RL-assisted tuning for cyclotron operation and motivate future work on safer transfer and multi-configuration learning.

Speaker: Dr Malek Haj Tahar (Transmutex)

16 Toward Autonomous Beamline Tuning and Control at the Argonne Wakefield Accelerator

The Argonne Wakefield Accelerator (AWA) is a flexible user facility and testbed for accelerator physics, radiofrequency technology, and advanced accelerator concepts. Recent efforts include migrating controls to EPICS for greater standardization and portability across facilities, alongside the addition of a new low-level RF system. These hardware upgrades have enabled the deployment of Bayesian optimization for automatic tuning and generative phase-space reconstruction. Further expanding these capabilities to integrate HPC resources for online analysis, deploy online digital twins, and further extend autonomous control to complex tasks is our main goal. This contribution will report on current progress and near-term plans.

Speaker: Anthony Tran (Argonne National Lab)

2026_07_16_SLAC_Workshop.pptx.pdf

17 LUME Ecosystem Updates: Standardized Virtual Accelerators and Digital Twin Infrastructure

Operating scientific user facilities demands continuous optimization and rapid decision-making to maximize throughput and scientific output. To address this, we present major updates to the LUME ecosystem that standardize the implementation and deployment of virtual accelerators and digital twins across heterogeneous simulation backends and control system interfaces at SLAC National Accelerator Laboratory.

At the core of these updates is the expanded LUMEModel abstraction, which provides a fixed, simulator-agnostic API with a rich variable system encoding metadata such as units and data types. This design enables standardized interaction with physics-based simulators, surrogate models, and differentiable simulations, supporting both Python-native workflows and EPICS IOC-based operation. Facility- and simulator-specific details are encapsulated through extensible transformer layers, allowing consistent control-system semantics across diverse simulation engines.

To enable production deployment of these digital twins, including complex workflows that chain ML surrogates with physics simulations, we leverage a robust MLOps infrastructure built on Kubernetes. This framework provides rigorous model versioning and tracking, containerized deployment templates, and seamless transitions from model development to online serving with continuous inference capabilities. The system supports continual learning workflows and flexible configuration for rapid adaptation to evolving control scenarios.

We present the updated LUME package architecture, deployment infrastructure, and representative use cases including model interchangeability, staged and chained simulators combining surrogates with physics codes, and continuous integration testing. This integrated ecosystem accelerates the deployment of virtual accelerators while enhancing automation, reliability, and scientific output.

Speaker: Sara Miskovich (SLAC)

lume_2026_aiml_test_fac.pdf

11:00 Coffee

AI/ML Applications: AI/M and Accelerators

18 Machine Learning for Accelerator Online Tuning and Anomaly Detection

Machine learning (ML) is emerging as a powerful tool for improving accelerator operations by reducing tuning time, increasing machine reliability, and enabling more autonomous operation. At the National Synchrotron Light Source II (NSLS-II), we have applied ML to two important operational tasks: accelerator tuning and anomaly detection. In this talk, I will present our work on using ML for efficient online tuning of the injector and storage ring. I will also present a model-based anomaly detection method that monitors magnet performance and identifies overheating issues, providing early alerts so engineers can flush the abnormal magnets before problems become more serious. The presentation will also describe how these ML tools are integrated into the NSLS-II control system and share practical lessons from online deployment, including real-time data access, safe operation, and interaction with operators. These experiences help identify the infrastructure and software capabilities needed to support future AI/ML-enabled accelerator facilities.

Speaker: Minghao Song (Brookhaven National Laboratory)

19 Osprey: A Portable Platform for Developing and Deploying Agentic AI Across Accelerator Facilities

Agentic AI has the potential to assist accelerator operations through natural-language interaction, automated analysis, and execution of complex workflows. Yet deploying such capabilities at operating user facilities remains challenging. Development opportunities on production machines are limited, operational risk is high, and software developed for one facility often requires substantial effort to adapt to another.

This work presents Osprey, an agentic AI platform designed to address these challenges through a portable architecture that separates facility-specific knowledge and interfaces from the underlying agent framework. Osprey provides semantic access to control-system information, governed interaction with operational tools, and facility-grounded memory, enabling the same agent capabilities to be deployed across multiple accelerator facilities with minimal modification.

We argue that portable software platforms and dedicated AI test facilities play complementary roles in accelerating adoption. Test facilities provide an environment where agent capabilities can be developed, validated, and challenged without impacting scientific operations, while portable platforms ensure that validated capabilities can be transferred efficiently to production facilities. Drawing on experience deploying Osprey across multiple DOE accelerator facilities, we discuss architectural requirements for portability, safety validation, and scalable adoption of agentic AI in accelerator operations.

Speaker: Thorsten Hellert (LBNL)

20 ORION: A Supervisory Architecture for AI-Assisted Autonomous Beamline Operation

Recent advances in artificial intelligence and machine learning have created new opportunities for autonomous operation of accelerator facilities. While significant effort has focused on developing optimization algorithms and AI-enabled control frameworks, comparatively less attention has been devoted to the supervisory software architecture required to coordinate autonomous multi-section optimization workflows. As the accelerator community considers the development of dedicated AI/ML test facilities, practical experience with deployable control architectures can help inform future facility design.

This contribution presents ORION (Orchestrated Real-Time Intelligent OptimizatioN), a modular framework for autonomous multi-stage beamline tuning developed for the nuCARIBU radioactive ion beam facility at Argonne National Laboratory. ORION combines Bayesian Optimization for section-level tuning with a supervisory orchestration layer powered by locally hosted large language models. The framework employs a hybrid decision strategy in which deterministic rule-based logic governs well-defined operating conditions, while language-model reasoning is invoked only for ambiguous supervisory decisions, such as adaptive section transitions and warm-start initialization. The architecture further incorporates an independent retrieval-augmented reference system for operator assistance and a web-based interface providing synchronized monitoring of optimization progress and hardware status.

Recent efforts such as Osprey [1] have demonstrated the value of production-ready AI frameworks for safety-critical accelerator control systems. ORION addresses a complementary problem by focusing on supervisory coordination of autonomous multi-section optimization workflows rather than general-purpose agent capabilities. The framework illustrates how deterministic optimization, AI-assisted reasoning, hardware-aware workflow coordination, and operator-centered interfaces can be integrated into a unified architecture for autonomous beamline tuning.

Rather than focusing solely on optimization performance, this presentation discusses the architectural principles that emerged during the development of ORION. Topics include the separation of optimization from supervisory reasoning, coordination of software decisions with beamline hardware operations, operator-centered interface design, and practical considerations for deploying locally hosted language models in accelerator control environments. These experiences provide insight into the software infrastructure and operational paradigms that may inform the design of future AI-native accelerator test facilities and other complex scientific systems.

[1] T. Hellert, J. Montenegro, and A. Sulc, Osprey: Production-ready agentic ai for safety-critical control systems (2025), arXiv:2508.15066 [cs.MA

Speaker: Sergio Lopez-Caceres (Argonne National Laboratory)

lopez-caceres_ORION_aiml_test_fac.pdf

21 ML+LLRF: From Demonstrated Laser Control to SRF Resonance Control

Through years of HEP core investment and the AI-Hardware project, we have established a machine-learning + low-level RF (ML+LLRF) foundation for real-time control of lasers and accelerator subsystems. The foundation is built on the open, reconfigurable LBNL Marble platform and a set of general-purpose DSP repositories designed to be portable across plant classes, rather than tied to a single application. This substrate already underpins two demonstrated capabilities: ML-based laser combining and laser-pointing stabilization at BELLA. We are now applying the same approach to PIP-II SRF resonance control—a large collaboration with FNAL, SLAC, ANL, MSU, JLab and others—which shows strong promise in simulation and is slated for hardware demonstration, with early hardware tests at the Vertical Test Stand (VTS) showing promising results. Because the platform and DSP libraries are broadly reusable, they open a wide range of additional applications and test opportunities across facilities. This talk provides a status update on our demonstrated results, work in progress, and near-term plans, and outlines where an open, ML-ready control substrate can serve the broader community.

Speaker: Dan Wang (LBNL)

FACET-ML_LLRF-July2026.pptx (1).pdf

13:00 Lunch

AI/ML Applications

22 Autonomous Tuning at FACET-II

FACET-II requires operators to conduct multiple tasks for both injector startup and optimal performance at the experimental area. Injector startup tasks consume hours of beamtime, a significant cost for a multi-purpose facility. Moreover, beam tuning for plasma wakefield acceleration at the experimental area is a challenging and time-consuming task due to the longitudinal-transverse coupling at the W-chicane compressor. In this work, we present recent developments in software tools to automate injector startup tasks and sextupole tuning at the W-chicane. First, we demonstrate automated tasks for laser alignment, Schottky scan phasing, emittance optimization, beam steering, and energy spread minimization at the injector using Xopt routines. Additionally, we show an autonomous injector startup that integrates these tasks using an AI agent. Finally, we present sextupole tuning in the W-chicane using Xopt. These demonstrations can save a significant amount of beamtime spent on tuning and have been successful in producing desired outcomes in beam quality.

Speakers: Juan Pablo Gonzalez-Aguilera (SLAC National Accelerator Laboratory), Yiheng Ye (SLAC)

AI Test Facility Workshop - Autonomous Tuning at FACET II.pdf

Discussion: Operations + Users + AI/ML and Accelerators

Day2_intro.pptx

Day 3 intro.pptx

Folder for Vision Statements

15:30 Coffee

Discussion: Paragraph on what AI/ML Test Facility would look like, what you would it do for you?

Day2_intro.pptx

Day 3 intro.pptx

Folder for Vision Statements

Discussion: Report Writing

Day2_intro.pptx

Day 3 intro.pptx

Folder for Vision Statements

18:00 Dinner (no-host) - Dutch Goose

Dutch Goose
3567 Alameda de las Pulgas, Menlo Park, CA 94025

Friday 17 July

Discussion: Report Discussion

Day2_intro.pptx

Day 3 intro.pptx

Folder for Vision Statements

Artificial Intelligence and Machine Learning

23 RL for particle accelerators

Reinforcement learning (RL) is a unique learning
paradigm inspired by the behaviour of animals and humans to learn to solve tasks autonomously. Learning occurs
through interactions with an environment, exploring, and
evaluating strategies under various conditions. RL excels in
complex environments, can handle delayed consequences,
and is able to learn solely from experience without access
to an explicit model of the system. This makes RL particularly promising for particle accelerators, where the dynamic
conditions of particle beams and accelerator systems require
continuous adaptation, and modelling is challenging. Although RL applications are emerging in accelerator physics
and showing promising results, their widespread introduction faces critical challenges. Among the main obstacles
are the effective formulation of control problems, training,
and the deployment of solutions in real systems. In this talk I will present the promise and open challenges of RL in particle accelerators.

Speaker: Andrea Santamaria Garcia (University of Liverpool and Cockcroft Institute)

2026-06-17-RL-accelerators.pdf

24 Industry Needs for ML in Safety Critical Systems

Safety-critical systems — particle accelerators, reactors, enrichment cascades — occupy a category apart. Failure is not a degraded user experience but a physical hazard, and the properties that define these systems (hard real-time constraints, formally verified interlocks, regulatory traceability, and a low tolerance for opacity) are precisely the properties that make standard machine-learning deployment untenable. This talk draws on our experience deploying and pitching AI/ML control and optimization at the Paul Scherrer Institute and at nuclear fuel-cycle facilities to argue that the automation of safety-critical systems is a distinct engineering discipline, not an application domain.
We present an architecture built on separation of concerns for safety: a deterministic safety layer retains exclusive enforcement authority; the model operates at a deliberately throttled cadence, well below the machine's reaction envelope; every proposed action is simulated before it is surfaced; and no action reaches the plant without an explicit green light from the machine-safety layer. We describe the methodology behind this — the requirements for high-fidelity historical data, a validated simulation environment, and deep operator expertise — and the confidentiality protections required to work inside sensitive facilities.
Central to the approach is a progression continuum for the human-in-the-loop boundary: automation earns trust incrementally, moving from advisory to supervised to autonomous only as evidence accumulates. We close on operator engagement as the decisive factor — the humans who run these machines are not obstacles to automation but its most rigorous validators.

Speaker: Edouard Servan-Schreiber (Transmutex)

20260716-Transmutex-SLAC-ML-Deck.pdf

25 IGNITE - Tokamak World Model for Full-Discharge Simulation

IGNITE is a multi-modal world model for full discharge simulation, trained on a large-scale dataset of DIII-D experimental data. This world model has been developed as part of the DOE Fusion Science & Technology Roadmap [1]. Inspired by the success of scientific world models for weather and climate forecasting [2,3,4], IGNITE is an end-to-end AI model that ingests heterogeneous diagnostic signals, videos, and actuators and autoregressively forecasts the plasma state of a DIII-D experiment, relying on a pre-defined actuator trajectory and on an initially measured plasma state. The model was trained on a large dataset of DIII-D discharges, covering more than ten years of experimentation (DIII-D shot numbers 160000 - present). For dataset collection and model training, the Fusion AI Toolkit and Hub (FAITH) [5,6] was used, and the source-code is publicly available in the git repository [5].
IGNITE can deal with signals at different time scales, from slowly evolving signals such as electron temperature measured by Thomson Scattering to high resolution signals such as interferometer and electron cyclotron emission radiometer. In addition, IGNITE can deal with camera data like tangential TV. The autoregressive predictions are window based, meaning that IGNITE processes windows of 50ms of data. Different sampling frequencies are taken into account by encoding each modality independently. Figure 1 shows an example of autoregressively predicting different modalities with IGNITE.
The authors gratefully acknowledge the collaboration of the DIII-D Team supported by the U.S. Department of Energy, Office of Science, Office of Fusion Energy Sciences, using the National Fusion Facility, a DOE Office of Science user facility, under Award No. DE-FC02-04ER54698 and DE-SC0024527. Additional support was provided by the Princeton Laboratory for Artificial Intelligence under Award No. 2025-97. In addition, this work was supported by the US Department of Energy Contract DE-AC02-09CH11466.
This research used resources of the Oak Ridge Leadership Computing Facility at the Oak Ridge National Laboratory, which is supported by the Advanced Scientific Computing Research programs in the Office of Science of the U.S. Department of Energy under Contract No. DE-AC05-00OR22725.

Speaker: Peter Steiner (Princeton University)

26 Continual learning for particle accelerators

Speaker: Kishan Rajput

27 Co-design for Accelerators

Speaker: David Bizzozero

11:15 Coffee

Discussion: Report Discussion

Day2_intro.pptx

Day 3 intro.pptx

Folder for Vision Statements

13:00 Lunch

Discussion: Report Discussion and writing

Day2_intro.pptx

Day 3 intro.pptx

Folder for Vision Statements

15:30 Coffee

Discussion: Report Closeout

Day2_intro.pptx

Day 3 intro.pptx

Folder for Vision Statements