Bayesian optimization of Laser-Wakefield Accelerators

Irshad, Faran (2025): Bayesian optimization of Laser-Wakefield Accelerators. Dissertation, LMU München: Fakultät für Physik

Abstract

Laser Wakefield Accelerators (LWFAs) are compact, high-gradient particle accelerators capable of producing ultra-relativistic electron beams over centimeter-scale distances. By driving large amplitude plasma waves with intense laser pulses, LWFAs can achieve accelerating fields orders of magnitude greater than those in conventional radiofrequency accelerators. Despite their potential for applications in medicine, industry, and ultrafast science, LWFAs remain limited by challenges in reproducibility, stability, and control. Their performance is highly sensitive to small changes in laser and plasma conditions, and they exhibit complex, nonlinear behavior. Traditional manual tuning and heuristic optimization strategies are insufficient to reliably access the high-dimensional parameter space and optimize multiple competing beam metrics such as energy, charge, and bandwidth. This thesis addresses these challenges by developing and implementing machine learning-based optimization strategies, specifically Bayesian optimization (BO), for the systematic control and tuning of LWFAs. After reviewing the theoretical foundations of Bayesian optimization, a novel multi-objective, multi-fidelity optimization framework termed Trust-MOMF is introduced. This method enables sample-efficient exploration of complex, high-dimensional parameter spaces using data sources with varying accuracy. Unlike traditional approaches, this method supports continuous fidelities without requiring monotonic improvement across fidelity levels, making it especially suitable for systems where measurements are noisy, expensive, and inconsistent. The performance of Trust-MOMF is first demonstrated in numerical simulations of LWFAs using particle-in-cell codes. The method shows an order of magnitude improvement in convergence times, substantially reducing the computational cost compared to standard multi-objective optimization techniques. The surrogate model learned during the optimization process is then inverted to enable energy tuning of the accelerator and uncover trade-offs between competing objectives such as bandwidth and efficiency. In the second part of the thesis, this framework is applied to real-world LWFA experiments using ATLAS which is a petawatt-class laser system. The number of laser shots per configuration is interpreted as a fidelity variable, allowing the optimization to balance information gain with experimental cost. For the first time, multi-objective multi-fidelity Bayesian optimization is demonstrated in LWFA experiments, yielding high-quality electron beams with significantly fewer laser shots. Furthermore, the trained models are used to perform fine-tuning and a posteriori single-objective optimization, enabling reproducible energy tuning between 150 and 400 MeV by navigating an 8-dimensional parameter space. Together, these results establish a generalizable and efficient approach for controlling LWFAs, enabling automated, data-driven operation tailored to specific user requirements. This work represents a significant step toward deploying laser-plasma accelerators in practical settings, including user facilities that require flexible and rapid re-optimization of beam parameters. It will further enable real-time tuning strategies, automated control, and eventual deployment of LWFAs in scientific user facilities.

• Dokumententyp: Dissertationen (Dissertation, LMU München)
• Keywords: Bayesian Optimization, Laser-Wakefield Accelerators, Laser-Plasma
• Themengebiete: 500 Naturwissenschaften und Mathematik ; 500 Naturwissenschaften und Mathematik > 530 Physik
• Fakultäten: Fakultät für Physik
• Sprache der Hochschulschrift: Englisch
• Datum der mündlichen Prüfung: 11. November 2025
• 1. Bericht­erstatter:in: Karsch, Stefan