Open-Source AI Framework Cuts Particle Beam Losses at GSI/FAIR from 45% to 12%

Particle accelerators are among the most complex machines ever built, and tuning them has long been a craft practiced by highly experienced operators who adjust magnets, timings and beam lines by hand. Now a team at the GSI Helmholtz Centre for Heavy Ion Research and the Facility for Antiproton and Ion Research (FAIR) in Darmstadt, Germany, reports that much of this craft can be automated. In an open-access study published on 4 August 2026 in EPJ Research Infrastructures, the researchers describe how the open-source Geoff framework, developed with CERN, delivered real-time optimization of beam parameters across a series of real machine experiments, cutting injection losses at the SIS18 synchrotron from 45 percent to as low as 12 percent and sharply accelerating the setup of the facility’s fragment separator.

Geoff is a modular, Python-based framework designed to bring machine learning and classical optimization techniques into accelerator control rooms on equal footing. Earlier attempts to apply machine learning to accelerator controls foundered, the authors explain, because the ecosystem at the time was anchored in Java, where most modern machine learning research happens in Python with tools such as TensorFlow and PyTorch. Individual scientists patched solutions together case by case, producing duplicated work and incompatible code. Geoff answers this by offering standardized interfaces, shared utilities and a reference graphical interface. Optimization problems are implemented as classes following the Gymnasium API, the same interface widely used in reinforcement learning research, while algorithms expose a single solve method that runs the entire optimization loop. Adapters connect the framework to established libraries such as SciPy, and the whole package is available as open-source software.

The framework’s flexibility is more than an engineering nicety. Because Geoff treats optimization problems as plugins that can be entire software packages, a single tuning task can span multiple control systems. At GSI, the five injection parameters of the SIS18 multi-turn injection are handled by the new GSI/FAIR control system, while steerer magnets in the preceding transfer line answer to the legacy control system. Geoff orchestrated both simultaneously. A graphical application, originally developed for the CERN accelerator complex, has been ported to the GSI control room and adapted to the FAIR control system based on CERN’s LSA framework, giving operators a live view of optimization runs with embedded monitoring plots.

The headline result concerns multi-turn injection into SIS18, the synchrotron that will act as a booster for FAIR’s flagship SIS100 ring. During multi-turn injection, beam pulses are accumulated over many turns, and losses at the injection septum or at the ring’s transverse acceptance can degrade the vacuum, a critical limit for intense intermediate-charge-state ion beams. Using the BOBYQA algorithm, a derivative-free method that builds a quadratic approximation of the objective within a trust region, the team randomized all starting parameters to avoid bias and used the median of three measurements per iteration to damp noise. In a November 2023 beam run, manual tuning by experienced operators left 45 percent of the beam lost; automatic optimization of all injection parameters reduced this to 15 percent. In May 2024, losses fell from 30 percent after manual tuning to 20 percent after partial optimization and 15 percent after full optimization, with the transfer-line steerers contributing a further 5 percent reduction. An optimization run takes roughly 15 to 20 minutes. Across the campaign, the team reports injection losses reduced from 45 percent to 12 percent.

The same machinery tamed slow extraction, the process by which synchrotrons deliver continuous beams to fixed-target experiments and hadron cancer therapy. Extraction relies on exciting a third-integer resonance with sextupole magnets, and uncontrolled losses at the electrostatic septum, which should stay below 5 to 10 percent, irradiate and damage accelerator components. Starting from identical initial machine states, Bayesian optimization, BOBYQA and Nelder-Mead all converged to low-loss settings within about 15 optimization steps, with good agreement between online results and a simulation study, though discrepancies remained because the simulation used an ideal magnet lattice and the measured beam parameters carried uncertainties.

The fragment separator FRS presented a different kind of bottleneck. Setting it up manually consumes up to one-third of the experimental time at GSI, and certain planned experiments at the Experimental Storage Ring are infeasible if optical mismatch between the separator and the ring drives the fragment beam intensity too low, a mismatch that cannot be fixed by hand because diagnostics are insufficient and too many parameters interact. In a November 2023 test, beam steering with four corrector magnets converged after 20 iterations. A harder task followed: automatic focusing in the dispersive focal plane, where a tantalum target split a uranium beam into three charge-state spots. A time projection chamber reconstructed individual particle trajectories, a classification algorithm from the Python tool scikit-learn grouped them into beam spots, and Twiss parameters derived for each spot were combined into a single score describing how upright the beam spot stood. The BOBYQA algorithm found the optimum after about 15 of 25 planned iterations. With the next-generation Super-FRS carrying four times as many magnets as the FRS, the authors argue automation will be essential for its commissioning.

Beyond classical optimizers, the study reports the first application of multi-objective and multi-fidelity Bayesian optimization to SIS18 injection tuning. Multi-objective optimization treats the multiplication factor of the injected current and the injection loss as two conflicting objectives, mapping out a Pareto front of equally valid trade-offs rather than a single best point. In an online run in June 2024, this approach identified more solution sets and a more favorable trade-off between loss and multiplication factor than BOBYQA. Multi-fidelity Bayesian optimization goes further by combining cheap low-fidelity simulations with expensive real machine measurements. In a proof-of-concept machine experiment constrained to 25 iterations, 14 simulation evaluations were combined with just 11 real machine measurements, reducing beam losses from about 45 percent of operator-defined settings to about 35 percent, a result the authors present as evidence that simulation can meaningfully guide tuning while sparing precious beam time on a multi-user machine.

The team also demonstrated data-driven model predictive control, which marries model-based reinforcement learning with control theory. A Gaussian process learns a probabilistic model of the ring’s response, capturing uncertainty, and the controller plans over a finite horizon, selecting actions that minimize a beam-loss cost while respecting hard limits such as septum current and aperture.

The framework’s reach extends beyond GSI through a collaboration on the PUMA experiment, which aims to study the neutron-to-proton composition at the surface of atomic nuclei by mixing antiprotons with radioactive ions in a transportable cryogenic Penning trap. Because PUMA runs on the EPICS control system used at more than 50 scientific facilities worldwide, it served as a testbed for porting Geoff beyond CERN- and FAIR-specific infrastructure. On a dedicated test setup at TU Darmstadt, the team automated a hot-cathode electron source with nine independently controllable electrode potentials, maximizing the beam current on a downstream Faraday cup. Bayesian optimization consistently outperformed Nelder-Mead, BOBYQA and even random search, which itself proved competitive, hinting at a landscape riddled with local optima. The infrastructure was then transferred to the PUMA offline ion source and its multi-reflection time-of-flight mass spectrometer. Optimizing five mirror electrode potentials, the ion storage energy and the extraction timing with BOBYQA narrowed the signal width by 10 percent, from 600 to 540 nanoseconds, while boosting detected ion intensity by 27 percent. Starting instead from random parameters, Bayesian optimization converged in two of ten runs, and subsequent BOBYQA refinement improved signal widths by 10 and 13 percent while raising intensities by 62 and 85 percent, reaching parameter settings far from the standard operating point that manual tuning would be unlikely to find.

The authors are careful to frame these as demonstrations within defined scopes: the model predictive control results are simulation studies, the multi-fidelity machine experiment ran under a tight iteration budget, and one hypothesis about logarithmic objective transformation could not be tested because the experimental campaign could not be repeated. Even so, the pattern across injection, extraction, fragment separation and ion-source tuning is consistent. By letting operators begin with simple, interpretable optimizers and graduate to Bayesian and reinforcement-learning methods without rewriting the problem definition, Geoff bridges traditional accelerator operations and modern data-driven optimization. Next stops include integration into GANIL under the iRIS project and the TwinRISE initiative, which will develop a digital twin of the Super-FRS as a pilot case, pushing automated tuning from a control-room convenience toward a core operating principle of future accelerator facilities.

Subject of Research: Automated tuning of particle accelerators using the open-source Geoff optimization framework at GSI and FAIR

Article Title: Automating Accelerator Tuning at GSI/FAIR

Article References: Appel, S., Alsmeier, H., Bajzek, M., Boine-Frankenheim, O., Dingeldein, L., Findeisen, R., Halilovic, B., Hirlaender, S., Hirt, S., Kallendorf, D., Kazantseva, E., Kazinova, O., Kraeft, M., Lenz, E., Madysa, P., Obertelli, A., Pfefferkorn, M., Pietri, S., Reinwald, C., … Wienholtz, F. (2026). Automating Accelerator Tuning at GSI/FAIR. EPJ Research Infrastructures, 10(1), Article 23. https://doi.org/10.1007/s41781-026-00175-6

Image Credits: AI Generated

DOI: 10.1007/s41781-026-00175-6

Keywords: particle accelerators, GSI, FAIR, machine learning, Bayesian optimization, reinforcement learning, beam tuning, SIS18 synchrotron, fragment separator, PUMA experiment, control systems, optimization