Multi-facility virtual diagnostic for longitudinal phase space predictions.

Area of Science:

• Accelerator Physics
• Machine Learning Applications
• Beam Diagnostics

Background:

• Understanding electron beam longitudinal phase space (LPS) is crucial for linear accelerators (linacs) and free electron lasers (FELs).
• Traditional transverse deflecting structure (TDS) measurements for LPS characterization are often destructive and complex.

Purpose of the Study:

• To develop and validate a machine learning-based virtual diagnostic (VD) for online LPS prediction.
• To demonstrate the generalizability of the VD framework across different accelerator facilities.

Main Methods:

• Training a machine learning model on destructive TDS measurements.
• Applying the trained model to non-destructive measurements for LPS prediction.
• Developing simplified architectures for predicting key beam parameters like bunch length and slice energy chirp.

Main Results:

• Achieved high prediction accuracy (scores >= 90%) for LPS across MAX IV, FERMI, and SwissFEL.
• Demonstrated a single network architecture and training procedure effective for diverse facilities.
• Successfully predicted key beam parameters using simplified VD architectures.

Conclusions:

• A generalizable virtual diagnostic framework enables rapid deployment for online LPS monitoring in accelerators.
• Machine learning offers a powerful, non-destructive alternative to traditional beam diagnostics.
• The VD approach can be further tailored for specific operational needs at different facilities.