Enhanced Performance of AutomaticTuning in Isotope Separation Online Systems through Bayesian Optimization: Improved constrained-aware and physics-informed Bayesian optimization for application in ISOL@MYRRHA for fast and safe automatic tuning of the ion source and beam transport.

The Multi-purpose hYbrid Research Reactor for High-tech Applications (MYRRHA) is a subcritical nuclear reactor driven by a linear proton accelerator, currently under development at the Belgian Nuclear Research Centre (SCK CEN). Part of the beam from this accelerator will be directed to an Isotope Separator OnLine (ISOL) facility (ISOL@MYRRHA) for the production of radioisotopes in the form of radioactive ion beams (RIBs). These isotopes play a crucial role in several fields of science, such as nuclear and atomic physics, fundamental interactions, condensed matter, and biology, as well as applications in nuclear medicine.
This thesis focuses on tuning two key subsystems within the ISOL system: the ion source and the transport beamline. Tuning involves adjusting multiple parameters associated with components across both subsystems, all of which collectively influence the yield and purity of the radioactive ion beam (RIB). Accordingly, the tuning process belongs to the operational phase of the ISOL system, whereas its design phase lies beyond the scope of this thesis. The ion source is responsible for ionizing the radioactive isotopes, while the transport beamline purifies and delivers the RIB to the end user. Beyond successful delivery, strict requirements concerning beam quality and purity must be satisfied, which vary depending on user needs and applications. Traditionally, the tuning of an ISOL system is performed manually by experienced operators. However, this approach is time-consuming and may lead to suboptimal configurations due to the high dimensionality of the operational parameter space. As a result, optimization algorithms are gaining increasing attention for automating the tuning process, offering gains in both efficiency and performance.
While optimization algorithms have been extensively applied in accelerator tuning, their use in ISOL systems is still an emerging area of research. Although the tuning processes in accelerators and ISOL systems share some similarities, ISOL systems present unique challenges. For instance, optimal tuning of the ion source remains largely unexplored, and the application of optimization techniques to this subsystem has been limited.
This doctoral research focuses on the development of optimization algorithms for the automatic tuning of ISOL@MYRRHA, with a particular emphasis on advancing the state-of-the-art through Bayesian optimization (BO). The thesis is structured around four primary objectives. The first objective is to accelerate the convergence of BO by exploiting available archive data or simulation results to capture correlations among operational parameters. This data-driven strategy provides an alternative to conventional Hessian-based methods, improving convergence speed by effectively incorporating prior knowledge into the optimization process. The second objective addresses safety in transport beamline tuning. A safety-aware BO algorithm is proposed that leverages simulation-based information to enforce safety when constraints are non-quantifiable, i.e., expressed only as binary outcomes (safe/unsafe).
The third objective addresses the performance requirements of both the ion source and the transport beamline. Since ionization imparts an electrical charge to isotopes within the RIB, their yield is directly reflected in the RIB current. For the transport beamline, an optimization methodology is developed to ensure the required beam purity, expressed in terms of achieving specific alignment and beam size. Among these conditions, the beam size is reformulated as a constraint, thereby reducing the problem to a single-objective constrained optimization. For the ion source, a nested optimization approach is introduced that optimizes the ion source parameters to maximize the RIB current.
The fourth objective is the development of a practical framework that integrates the proposed algorithms into a tool accessible to ISOL operators for automatic ISOL system tuning. Although designed for ISOL@MYRRHA, the framework and algorithms are generalizable to other ISOL systems, as demonstrated through experimental validation on the ISOLDE Offline 2 mass separator system at CERN and SCK CEN’s ISOFF system.
The application of the developed methods demonstrates significant improvements in tuning efficiency. The implementation of automated algorithms reduces the transport beamline tuning time from some hours of manual adjustment to approximately 30 minutes, while decreasing the likelihood of convergence to local optima. Automated algorithm substantially accelerates system setup and improves overall operational turnaround, enabling users to access the beam sooner, facilitating more timely execution of their experiments. For ion source tuning, automatic tuning achieved improvements in RIB current of up to 87 % compared to the standard operational parameter settings. This improvement represents a considerable gain in isotope yield, particularly when optimal configurations are applied from the start of an experimental campaign.
In conclusion, this work proposes an efficient Bayesian optimization-based methodology for tuning the ion source and transport beamline of ISOL systems, with a focus on key performance requirements such as ion-beam shape, alignment, and current. Future work may extend the methodology to include additional objectives, such as beam emittance and beam losses, within the automatic tuning routine. Furthermore, the integration of adaptive Bayesian optimization offers a promising direction for compensating time-dependent variations during ISOL system operation.