Geoff: The Generic Optimization Framework & Frontend for Particle Accelerator Controls Journal: SoftwareX

P. Madysa, S. Appel, V. Kain, M. Schenk · 2025

Read Geoff: The Generic Optimization Framework & Frontend for Particle Accelerator Controls Journal: SoftwareX by P. Madysa, S. Appel, V. Kain, M. Schenk (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-03796.txt.

Introduction

Penelope Madysa and co-authors introduce in 2025 Geoff, an open source software framework that standardizes machine tuning of particle accelerators. Labs worldwide are testing learning algorithms to squeeze more beam time and better performance from their machines, but each group builds its own harness. Comparing or swapping methods then means weeks of friction.

Geoff, a set of Python packages born at CERN, gives optimization problems a shared interface, stocks helper tools to speed development, and ships a reference graphical app that ties the parts together. Maintained with a partner institute under a European labs project, it is already in use at its home labs. The paper lays out the design and features for control room teams weighing adoption. One common workbench, the authors argue, lets the field spend its effort on algorithms rather than plumbing.

Catalog record

CREATORP. Madysa, S. Appel, V. Kain, M. Schenk
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE41 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-03796

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Geoff: The Generic Optimization Framework & Frontend for Particle Accelerator Controls Journal: SoftwareX P. Madysa, S. Appel, V. Kain, M. Schenk Abstract Geoff is a collection of Python packages that form a framework for automation of particle accelerator controls. With particle accelerator laboratories around the world researching ML techniques to improve accelerator performance and uptime, a multitude of approaches and algorithms have emerged. The purpose of Geoff is to harmonize these approaches and to minimize friction when comparing or migrating between them. It provides standardized interfaces for optimization problems, utility functions to speed up development, and a reference GUI application that ties everything together. Geoff is an open-source library developed at CERN and maintained and updated in collaboration between CERN and GSI as part of the EURO-LABS project. This paper gives an overview over Geoff’s design, features, and current usage. Metadata C1 Current code version 0.17.11 C2 Permanent link to code/repository used for this code version https://github.com/geoff-project/coi C3 Permanent link to Reproducible Capsule none C4 Legal Code License GPL-3.0-or-later OR EUPL-1.2+ C5 Code versioning system used Git C6 Software code languages, tools, and services used Gitlab, Python, PyTorch, Ruff C7 Compilation requirements, operating environments & dependencies Python 3.9+; NumPy, Gymnasium; the GUI application depends on CERN-internal components C8 If available Link to developer documentation/manual https://cernml-coi.docs.cern.ch/ C9 Support email for questions [email protected] Table 1: Code metadata 1 Motivation and significance The field of accelerator controls is concerned with the development of hardware and software for the operation of a particle accelerator or a complex of accelerators. It presents a number of challenges: 1. The hardware is operated around the clock and subjected to considerable levels of radioactivity. 2. Strong uptime requirements mean that the software must be efficient, reliable and deterministic even in case of failures. 3. Manpower, funds and upgrade efforts are limited, even though accelerators operate for years and decades. Faced with this, software engineers responsible for building and maintaining controls systems used to eschew the latest technologies or dynamic programming languages, and favored more traditional and battle-tested technologies instead. In practice, this means that middleware and control-room applications used the Java language and ecosystem. At the same time, the field of ML ( ML ) can no longer be ignored to satisfy the performance and uptime requirements with increasing accelerator complexity. Early attempts to use ML for accelerator controls [ 1 ] were successful in individual test cases, but didn’t gain traction due to the limited ML ecosystem in Java. Most contemporary ML research is based on the Python language and frameworks like TensorFlow and PyTorch. While individual accelerator scientists have used these technologies on a case-by-case basis [ 2 ] , this has led to a lot of duplicated work , lack of compatibility and disregard for usability . Geoff [ 3 ] is a Python framework developed at CERN and open-sourced as part of the EURO-LABS [ 4 ] project. It provides: • a standardized API for numerical optimization and RL ( RL ), making it easier to share and compare algorithms; • a library of utility functions to avoid duplicate work; • and a GUI application to easily to deploy solutions in the accelerator control room. Because Geoff is written in Python [ 5 ] , the full ecosystem of the latest ML is available. In this way, it bridges the gaps between ML experts, control-room operators, and accelerator physicists. In a typical case, Geoff is deployed at a research institute by a group of software and IT infrastructure experts, e.g. the institute’s controls department. Operators—engineers managing the day-to-day operations of the accelerator from a control room—use the Geoff application to solve certain well-defined optimization problems . These problems are Python plugins written by domain experts with knowledge of specific accelerator subsystems—often physicists or engineers. Geoff’s open nature enables these groups to communicate and empowers operators to improve existing and create new plugins. At the same time, its emphasis on versioning ensures that any erroneous updates can be easily rolled back. Geoff is preceded by a number of similar frameworks [ 6 , 7 , 8 ] . It improves upon them with its radical approach to extensibility, making many problems solvable that would otherwise be impossible to express. It competes [ 9 ] with Xopt [ 10 ] and Badger [ 11 ] . Both have distinct advantages, but lack support for RL or versioning and dependency management. Geoff’s API builds on that of the Gymnasium [ 12 ] library. By default, its application is bundled with state-of-the-art algorithms, such as provided by Stable-Baselines3 [ 13 ] , SciPy [ 14 ] and Py-BOBYQA [ 15 , 16 ] . It uses NumPy [ 17 ] for data transfer and PyQtGraph [ 18 ] for visualization. Plugins can visualize their own data via Matplotlib [ 19 ] . Geoff is hosted at CERN and its DevOps automation is facilitated by CERN’s Acc-Py project [ 20 ] . 2 Software description Geoff consists of a number of Python packages, some more central to its concepts than others. The most important ones are as follows: cernml-coi defines and standardizes the C ommon O ptimization I nterfaces interfaces for optimization problems. The interfaces are divided by capabilities so that an optimization problem can support numerical optimization, RL , or both. Interfaces for RL are provided by the Gymnasium package, which this package builds upon. Various mix-in interfaces can express additional capabilities, e.g. configurability via the GUI application. The package also provides routines that verify whether an implementation satisfies the invariants of these interfaces. cernml-coi-optimizers standardizes the interface specifically for numerical optimization algorithms. It also provides adapters to this interface for many popular libraries such as SciPy. The adapters are written as optional dependencies, so that they need not be installed if not required. cernml-coi-utils provides a number of utility functions and classes for optimization problems. These allow Geoff plugins to remain short and focused on the task at hand. Examples include a queue for simplified communications with an accelerator device and a method decorator to manage the state of a custom Matplotlib visualization. geoff-app is a reference implementation of a GUI application based on PyQt [ 21 ] that loads a set of plugins and communicates with them via the above interfaces. Research institutes outside of CERN are encouraged to fork this project to take local concerns into account and provide accelerator-specific information. See section 3 for an example of how this is realized at CERN. All of these packages follow Semantic Versioning [ 22 ] independently. This allows them to evolve at different speeds and keeps version numbers meaningful; e.g. a backwards-incompatible change to the utilities library (incrementing its MAJOR version number) has less impact than a similar change to the interfaces. 2.1 Software architecture Figure 1: Design architecture of Geoff The general architecture of the framework is shown in Fig. 1 . The user controls the system via a host application . This may be a GUI application like geoff-app , or a console script, or even a server endpoint. The application loads available algorithms and optimization problems, either by direct import or via Python’s entry points API [ 23 ] . In addition to

Read the complete text