Bayesian Optimization of a Free-Electron Laser.

Area of Science:

• Accelerator physics
• Machine learning
• X-ray science

Background:

• The Linac Coherent Light Source (LCLS) requires frequent configuration changes, impacting experimental efficiency.
• Fast tuning strategies are crucial for minimizing setup time between experiments at the LCLS.

Purpose of the Study:

• To develop and implement a sample-efficient optimization routine for tuning the LCLS x-ray free-electron laser.
• To maximize x-ray laser pulse energy by controlling quadrupole magnets.

Main Methods:

• Utilized a Bayesian approach with Gaussian processes for probabilistic modeling of machine response.
• Incorporated historical data and accelerator physics knowledge for model parameter learning.
• Balanced exploration and exploitation in optimizing control parameters.

Main Results:

• Successfully learned model parameters from archived experimental data.
• Extracted correlations between control devices from beam transport data.
• Demonstrated a sample-efficient optimization routine outperforming existing methods.

Conclusions:

• The Bayesian optimization strategy significantly enhances the tuning efficiency of the LCLS.
• This approach offers a powerful tool for optimizing complex scientific apparatus in real-time.
• Combining machine learning with domain knowledge accelerates experimental throughput.