Autonomous operation of the DIAG0 diagnostic line for 6D phase-space monitoring at LCLS-II

This paper presents the first fully autonomous 6-dimensional beam-tomography system deployed on the LCLS-II DIAG0 line, which utilizes machine learning for adaptive control and generative analysis for rapid reconstruction to enable real-time, high-fidelity monitoring of the photoinjector's phase-space distribution.

The Gist

Imagine a massive, high-speed train station where the "trains" are actually beams of electrons, and the "destination" is a machine that creates incredibly bright X-rays for scientific experiments. This is the LCLS-II at SLAC National Accelerator Laboratory.

For this system to work perfectly, the electron beams need to be shaped and focused with extreme precision. If the beam is even slightly "out of shape" or drifting off course, the X-rays won't be as bright, or the machine might get damaged.

Here is the problem: The beam changes constantly. It's like trying to photograph a speeding car that is also changing its shape and color every second. To fix it, you need to know exactly what the car looks like right now.

The Old Way: The Manual Mechanic

Previously, checking the shape of this electron beam was a slow, manual process.

1. The Setup: A team of human experts would have to manually tweak dozens of magnets and knobs to steer the beam into a special "diagnostic tunnel" (called DIAG0).
2. The Scan: They would take measurements, then tweak the magnets again, take more measurements, and repeat this for hours.
3. The Analysis: After gathering all this data, they would spend many more hours on a computer trying to reconstruct a 3D (actually 6D!) picture of what the beam looked like.

By the time they figured out what was wrong, the beam had already changed again. It was like trying to fix a car by looking at a photo taken three days ago.

The New Way: The Self-Driving Car

This paper describes a breakthrough: The first fully autonomous system that can check, fix, and map the electron beam all by itself, in real-time.

Think of it as upgrading from a manual mechanic to a self-driving car with a super-intelligent navigator.

1. The "Self-Driving" Pilot (Bayesian Optimization)

Instead of humans turning knobs, the system uses a smart AI algorithm called Bayesian Optimization.

• The Analogy: Imagine you are trying to find the perfect spot to park a car in a dark garage. You don't want to crash into the walls (which would damage the machine).
• How it works: The AI "feels" its way around. It tries a position, checks if it's safe, and learns from that attempt. It quickly figures out the best path to center the beam without crashing. It does this for steering, focusing, and timing, all while obeying strict safety rules (like "don't let the beam hit the walls").
• The Result: The system can re-configure the entire diagnostic tunnel in minutes, adapting instantly if the incoming beam drifts.

2. The "Super-Vision" Camera (Generative Phase Space Reconstruction)

Once the beam is in position, the system takes pictures. But these aren't normal photos; they are complex "shadows" of the beam from different angles.

• The Analogy: Imagine trying to figure out the shape of a complex sculpture inside a box by only looking at its shadow on the wall.
• How it works: The system uses a powerful AI called GPSR (Generative Phase Space Reconstruction). It's like a 3D printer that works backward. It takes the 2D shadow data and "prints" a full 6-dimensional model of the beam in its mind.
• The Speed: In the past, this took hours. Now, thanks to super-fast computers (GPUs), it does it in 5 to 10 minutes.

Why This Matters

The paper demonstrates that this system can run autonomously for hours while the main machine is running experiments for real users.

• Continuous Monitoring: It's like having a doctor constantly monitoring a patient's heartbeat, rather than just checking them once a year.
• Instant Fixes: If the beam starts to drift, the system notices immediately, re-optimizes the magnets, and gets back on track without human help.
• Deep Insight: It doesn't just tell you the beam is "bad"; it shows you exactly how the shape is changing, revealing hidden details that humans would miss.

The Big Picture

This is a major step toward fully autonomous accelerators. In the future, these machines might not need human operators to tune them at all. They will be able to sense their own health, diagnose problems, and fix themselves, ensuring that the world's most powerful X-ray machines run smoothly, safely, and efficiently 24/7.

In short: They taught the machine to drive itself, check its own mirrors, and fix its own engine while the passengers (the scientists) are busy doing their work.

Technical Summary

1. Problem Statement

Characterizing the full 6-dimensional (6D) phase-space distribution of electron beams is critical for optimizing the performance of the Linac Coherent Light Source II (LCLS-II) free-electron laser (FEL). However, continuous, real-time monitoring of this distribution faces two primary challenges:

• Operational Dependency: Traditional tomographic measurements require dedicated personnel to manually tune beamline parameters (steering, focusing, cavity phases) to maintain beam transport and measurement quality. This is unsustainable during long user runs where machine drifts or reconfigurations occur.
• Computational Latency: Conventional 6D phase-space reconstruction methods (e.g., SART algorithms) are computationally expensive, often requiring hours of processing time and thousands of measurements. This latency makes them unsuitable for online monitoring or feedback loops, as they cannot track the temporal evolution of the beam distribution during operation.

2. Methodology

The authors developed a fully autonomous framework combining Bayesian Optimization (BO) for control and Generative Phase Space Reconstruction (GPSR) for analysis.

A. Autonomous Control Framework (DIAG0 Beamline)

The system autonomously configures the DIAG0 parasitic beamline using the Xopt Python package. It employs BO algorithms to solve four sequential optimization tasks without human intervention:

1. Beam Steering: Adjusts 14 corrector magnets to center the beam on Beam Position Monitors (BPMs). The system models each BPM signal independently using Gaussian Processes (GPs) to reduce dimensionality and efficiently identify irrelevant magnets.
2. Beam Focusing: Uses Constrained Log Expected Improvement (logEI) to tune quadrupole magnets (QDG007–QDG011) to focus the beam onto diagnostic screens (OTRDG02 and OTRDG04), maximizing resolution.
3. TCAV Phasing: Optimizes the phase of the Transverse Deflecting Cavity (TCAV) to ensure the beam experiences the zero-crossing of the transverse kick. A periodic kernel is used in the GP model to accelerate convergence to the sinusoidal optimum.
4. Tomographic Scans: Uses a Constrained Upper Confidence Bound (UCB) acquisition function to autonomously select quadrupole scan ranges. This balances exploration (finding the minimum beam size) and exploitation while adhering to constraints (maximum beam size, signal-to-noise ratio) to prevent beam loss.

Robustness: The system utilizes the tenacity package for error handling, allowing it to retry failed measurements or restart workflows from initial settings if beam interruptions occur, ensuring long-term operation.

B. Generative Phase Space Reconstruction (GPSR)

Once tomographic data is collected, it is streamed to the SLAC Shared Scientific Data Facility (S3DF) for reconstruction:

• Model: A neural network (4 layers, 100 neurons) parameterizes the 6D beam distribution, transforming random noise into macro-particle coordinates.
• Differentiable Simulation: The Cheetah simulation package is used to track these particles through the DIAG0 beamline. The simulation is differentiable, allowing the system to solve the inverse problem: finding the input distribution that best matches the experimental images.
• Data Flow: Raw images are pre-processed (background subtraction, filtering, cropping) and fed into the GPSR pipeline. Reconstructions are performed on NVIDIA A100 GPUs.

3. Key Contributions

• First Fully Autonomous 6D Monitoring: This is the first demonstration of a fully autonomous system operating a parasitic diagnostic beamline (DIAG0) at LCLS-II to continuously monitor 6D phase space.
• Adaptive Control: The system dynamically re-optimizes scan ranges and beamline parameters in response to incoming beam drifts, eliminating the need for manual re-tuning.
• High-Cadence Reconstruction: By combining GPU-accelerated GPSR with autonomous data collection, the system achieves a reconstruction cadence of one 6D distribution every 5 to 10 minutes.
• Uncertainty Quantification: The offline analysis employs ensemble methods (running 10 independent reconstructions per dataset) to estimate confidence intervals and distinguish between algorithmic uncertainty and physical beam fluctuations.

4. Results

The system was tested during a 9-hour user shift on November 25, 2025, delivering 70 pC beams at 35 kHz.

• Operational Performance: The system successfully completed 21 full 6D phase-space measurements over the shift. In the longest continuous run, it performed 9 measurements in 52 minutes (avg. 5.75 min/measurement).
• Reconstruction Quality:
  • Profile Agreement: Reconstructed beam profiles at diagnostic screens (OTRDG02, OTRDG04) showed strong qualitative agreement with experimental measurements, including complex features like beam halos and non-linear correlations.
  • Temporal Evolution: The system tracked the evolution of scalar RMS quantities (emittance, bunch length, energy spread) over time. It detected drifts in vertical emittance and changes in the $y-\delta$ (vertical-energy) correlation structure that were invisible to standard RMS monitoring.
  • Confidence Metrics: The ensemble approach revealed high confidence in the beam core but lower confidence in the beam halo, correlating with the signal-to-noise ratio of the experimental images.
• Comparison to Conventional Methods: The paper demonstrates that GPSR captures fine-scale structures and non-Gaussian correlations that traditional Multivariate Normal (MVN) fitting misses. MVN fits yielded significantly different emittance values (e.g., 0.5 mm·mrad vs. 2.8 mm·mrad for horizontal emittance), highlighting the necessity of generative models for accurate inference.

5. Significance

• Real-Time Diagnostics: This work bridges the gap between slow, offline tomography and the need for real-time beam characterization. It enables operators to observe how accelerator component drifts affect the beam structure during user operations.
• Path to Autonomous Accelerators: The framework demonstrates that complex beamlines can be operated autonomously using AI-driven control and analysis, a critical step toward "self-driving" accelerators for current and next-generation facilities.
• Foundation for Virtual Diagnostics: The systematic data collection enabled by this autonomous system provides the high-fidelity dataset required to train future "virtual diagnostics" (machine learning models that predict beam properties from machine settings without direct measurement).
• Scalability: The modular nature of the BO control and GPSR analysis suggests this approach can be adapted to other diagnostic lines and accelerator facilities globally.