1 of 8

Accelerators: AI/ML solutions

Session 2 summary

FAIRS2026 · Accelerator group · Facilitators: Fumihiko Tamura, Takahiro Nishi

2 of 8

Background: why accelerator operation is hard

2

  • ~10⁶ control variables (PVs) per facility — magnets, RF, vacuum, …
  • Conditions drift over time (temperature, magnet hysteresis) → continuous retuning
  • Beam loss must be avoided: it damages hardware and stops the machine
  • Many beam properties cannot be measured without stopping the beam

Topic

In one sentence

Digital twin

A virtual copy of the machine, calibrated with real data, faster than physics-based simulation

Virtual diagnostics

Infer beam properties that are hard or destructive to measure, from existing monitors

Anomaly detection

Learn “normal” from thousands of signals and warn before a fault stops the beam

Agentic AI

An AI assistant that reads manuals and logbooks, writes code and uses control tools

Optimization

Tune many parameters automatically (BO, RL, adaptive control) under safety constraints

FAIRS2026 · Session 2 · Accelerators

3 of 8

Digital twin

3

A virtual copy of the accelerator, calibrated with machine data, fast enough to use in operation

AI/ML solutions

  • Digital twin for operation / anomaly detection
  • Surrogate models
  • Foundation model, knowledge base (RAG / operation manual), control codes
  • e.g., Osprey (join the collaboration)

Comparison to other methods

  • Faster, cheaper and more accurate than physics-based simulations
  • Obtains information that is missing in the real machine

Steps to build the solutions

  • Build a high-fidelity simulation model:
    • integrate the several simulation codes now in use
    • correct it with experimental data
  • Emulate the control system
  • Common formats are important (PALS, openPMD)
  • Some simulations (e.g., PIC) are too heavy → surrogate model (faster)

Why are they not built yet?

  • Human resources
  • Lack of data: unknown parameters; new diagnostics required
  • Key features of a twin still unclear (all-in simulation is too heavy)
  • Differentiable simulation codes (Cheetah, SciBmad) as the framework
  • Iwasaki-san's NN-based method:
    • a system identification study — one of the solutions

FAIRS2026 · Session 2 · Accelerators

4 of 8

Virtual diagnostics

4

Use ML to infer beam properties that are hard or destructive to measure, from the monitors we already have

AI/ML solutions

  • e.g., phase-space reconstruction
  • Rich information from non-destructive monitors
  • Higher-resolution diagnostics

Comparison to other methods

  • Compared with tomographic methods:
    • richer information, non-destructive
    • faster / real-time; a small data set is enough

Steps to build the solutions

  • Use destructive monitors as the ground truth for training
  • In operation, use only non-destructive monitors
  • GPSR (generative phase-space reconstruction; based on differentiable simulation)

Why are they not built yet?

  • Some facilities have already started to build them
  • Hard to share between facilities

FAIRS2026 · Session 2 · Accelerators

5 of 8

Anomaly detection

5

Learn what “normal” looks like across many signals and flag deviations before a fault stops the beam

AI/ML solutions

  • Bayesian regression, autoencoder, PCA, random forest, SVM, digital twin
  • Monitoring the whole accelerator system (NN / transformer, digital twin)

Comparison to other methods

  • Compared with rule-based methods:
    • earlier detection; correlations can be taken into account
    • scales to 1000s of diagnostics
    • better performance (much higher recall, higher precision)
  • Spare parts can be prepared in advance

Steps to build the solutions

  • Supervised learning: annotation of operation data

　(Unsupervised learning may be used if annotation is too difficult.)

  • Select a good model (depends on facility / component)
  • Methods for time-series data (LSTM, transformer, embedding, agentic AI)

Why are they not built yet?

  • Human resources
  • In many cases rule-based detection is enough — but higher availability now requires improvement
  • Annotation is not trivial for operation data
  • SACLA has already started ML-based anomaly detection of thyratrons.

FAIRS2026 · Session 2 · Accelerators

6 of 8

Agentic AI

6

An LLM-based assistant that plans and uses tools: searches manuals and logbooks, writes code, operates control software

AI/ML solutions

  • Knowledge base (RAG / operation manual)
  • Control codes
  • Use other codes / manage the control

Comparison to other methods

  • Pros (compared with operators):
    • keeps quality; consistent, scalable, lower cost
    • preserves expert know-how (e.g., capturing the PC screen)
  • Cons: risk of faults (beam loss) — not only with agentic AI

Steps to build the solutions

  • Models, harness, skills, knowledge base
  • Sub-agents / MCP for each component or purpose
  • Local model is required for control
  • Foundation model: frontier model → Genesis open model
  • The agent can create the code / MCP

Why are they not built yet?

  • Agentic AI is new; physics-familiar (multimodal) models needed
  • Security: control network is isolated → local LLM / gateway with authentication
  • Budget for a local LLM (can start from ~1M JPY)
  • OCR of handwritten logbooks for RAG is difficult

FAIRS2026 · Session 2 · Accelerators

7 of 8

Optimization for operation (Additional)

7

We did not have the time to discuss this topic, so facilitator added this page.

AI/ML solutions

  • Bayesian optimization
  • Reinforce learning
  • Adaptive control
  • Physics informed NN

Comparison to other methods

Pros.

  • keeps quality; consistent, scalable
  • reduced dependence on tuning experience.
  • correlations can be taken into account
  • faster

Cons. risk of faults (beam loss) — not only with ML tuning

Steps to build the solutions

  • Solve the problem of the risk of faults�　Safe constraints such as SafeLineBO�　Limit search range based on the prior knowledge
  • Prepare GUI for operators and get feedback
  • time-dependent modeling, adaptive control
  • Training by digital twin

​

​

​

Why are they not built yet?

  • Already introduced at SACLA and other facilities; development is progressing
  • Safety requirements are high → tests must be repeated to guarantee safety
  • RL needs a digital twin first

FAIRS2026 · Session 2 · Accelerators

8 of 8

Summary

8

Topic

Key message

Digital twin

Integrate the several simulation codes into a high-fidelity model corrected with experimental data; surrogate models and NN-method as solutions

Virtual diagnostics

Already being built at some facilities (e.g., GPSR); trained with destructive monitors, used non-destructively; hard to share between facilities

Anomaly detection

Earlier and more scalable than rule-based methods (SACLA has started); needs annotation or unsupervised learning, and a model per component

Agentic AI

New: knowledge base + control codes via sub-agents / MCP; needs a local LLM for the isolated control network, budget and digitized logbooks

Optimization

Added by facilitators: BO already used at SACLA and others; safety must be guaranteed by repeated tests; RL needs a digital twin first

Common to all topics: human resources and data are the main bottlenecks; safety must be guaranteed before use in operation; the digital twin is the basis for other topics (e.g., RL); sharing needs common formats (PALS, openPMD) and collaborations

FAIRS2026 · Session 2 · Accelerators