GSI/FAIR AI Workshop

The evolution of neural networks from their early conceptual
roots in the 1940s to their modern-day applications are presented.
Starting with McCulloch and Pitts' model of neurons, we explore key
milestones like Rosenblatt’s perceptron, the multiple rediscovery of backpropagation,
the Hopfield networks and boltzmann machines,
and the advent of deep neural networks. The discussion will
highlight challenges, breakthroughs, and future directions, offering a
comprehensive overview of how neural networks have shaped, and continue to
shape, artificial intelligence.

EuCAIF is an European initiative for advancing the use of Artificial Intelligence (AI) in Fundamental Physics. Members are working on particle physics, astroparticle physics, nuclear physics, gravitational wave physics, cosmology, theoretical physics as well as simulation and computational infrastructure. The objective of the coalition is to federate on various AI-related aspects (knowledge, infrastructure, sustainability, education, etc.) among these communities and to create funding opportunities towards the EU. GSI/FAIR, mandated via NuPECC, is representing the nuclear physics aspects within EuCAIF.

Ion beam therapy is the most advanced form of radiotherapy, requiring fast, precise, and accurate treatment planning. Speed is a critical factor for cancer patients, as timely interventions can significantly impact outcomes. While Monte Carlo simulations offer high-quality dose calculations, they are too slow for routine clinical workflows. In contrast, analytical pencil beam algorithms deliver faster results, but at the cost of some accuracy. We leverage the Dose Transformer Algorithm (DoTA) [DOI:10.1088/1361-6560/ac692e] to predict input data for the biological effects of carbon ion beams, aiming to achieve Monte Carlo-level quality with a speed comparable to the analytical approach.

Particle therapy provides a significantly more precise dose distribution within the patient compared to conventional photon radiotherapy, reducing radiation in healthy tissue but increasing the need for dose verification. A mixed beam of helium and carbon ions, recently first produced at GSI, can enable online range monitoring and simultaneous imaging. In the case of lung cancer, the measured helium range however does not directly correlate to the carbon range in the patient due to the strong density differences which is highly patient specific. We developed a deep-learning model to predict the carbon range from the measured helium range.

The precision of carbon ion therapy offers great potential for more targeted lung cancer therapy. To use the steep dose gradients to target moving tumors requires motion mitigation strategies that rely on real-time motion information acquired, e.g., with optical tracking systems. To utilize the information most effectively requires to predict motion variations. We developed an LSTM model to be used in conjunction with a real-time dose engine, for detecting issues in the treatment in real time and possibly correcting the dose delivery. The model was verified with experimental data from the Italian carbon ion therapy center CNAO.

Using the recent advances in machine learning (ML) algorithms, especially in the area of image recognition, we plan to develop new ML particle identification algorithms for the PANDA Barrel DIRC and compare the results to conventional reconstruction methods.
First network implementations show a performance comparable to conventional methods on a limited phase space. As a next step, we are investigating ways to extend the phase space, while also experimenting with different data input structures and network types to optimize the training process and increase PID performance.

The calibration of detectors in heavy-ion experiments is key to the measurement of several physics observables. Corrections must be applied to the expected detector signal in order to account for changes in environmental or hardware conditions over time and in different spatial regions. The use of neural network-based corrections trained on clean input signals allows this to be performed with a multidimensional approach. In this presentation I will introduce the neural network techniques used to parametrise the expected energy loss of charged particles in the ALICE Time Projection Chamber, which is crucial for reconstructing particle signals from their decay products.

LISA project aims to develop an array of five 5 × 5 grids of thin single-crystal diamond detectors, to measure otherwise inaccessible lifetimes in exotic nuclei. Determining reaction positions in the target is a pivotal component of the project and we are developing ML algorithms for characterization.We used ANN and Random Forest Algorithm on the simulated data. RF algorithm has shown the most promising results so far in detecting reaction layers for simulated data and next step will be to apply this to experimental data. In future, to improve the accuracy we plan to use DNN with better feature selections.

Non-destructive Schottky detectors can be used to accurately determine the masses and lifetimes of exotic nuclear species and/or their isomeric states in storage rings. The analysis requires time-resolved spectra to undergo particle identification and determination of frequency spread ∆f and decay time. Often times corrections for rigidity, drift, etc. need to be applied, or spectra need to be generated with different conditions. As detector efficiency increases, and with it the amount of data, manual analysis becomes increasingly error-prone and time-consuming. DNNs can be used to analyze 1D and 2D spectra. In the future, data from several detectors can be combined.

Real-time calibrations for future detectors at FAIR

The online data processing of the next generation of experiments conducted at FAIR requires a reliable reconstruction of event topologies and, therefore, will depend heavily on in-situ calibration procedures. In this study we present a neural network-based tool designed to provide real-time predictions of calibration constants, which rely on continuously available environmental data. A proof-of-principle of this approach has been demonstrated using data from the Drift Chambers of the HADES detector, for which our method demonstrated the ability to provide fast and stable calibration predictions.

Neural Network Based Particle Identification for HADES

The HADES experiment measures a number of parameters that can be used for particle identification (PID). Traditionally, this is done using simple graphical cuts, individually selected for each analysis. To improve upon this method, we have developed a neural network-based PID system for hadrons that can be applied universally. This approach leverages all available particle information simultaneously and operates with probabilities, allowing for more flexible classification. In this talk, we will discuss the details of the method and compare its performance with traditional graphical cuts.

The $\textbf{H}$igh $\textbf{A}$cceptance $\textbf{D}$i$\textbf{E}$lectron $\textbf{S}$pectrometer (HADES) is a versatile magnetic spectrometer aimed at studying dielectron production in pion, proton and heavy-ion induced collisions. The conventional reconstruction algorithm for the HADES RICH is based upon the Hough Transform (HT) method for ring finding. This method fails to efficiently identify dileptons with small opening angles. The complexity of HT scales exponentially with number of parameters involved and given the expected 10 MHz trigger rate in CBM/FAIR, a faster algorithm will be necessary to accelerate online analysis. A preliminary U-Net architecture for ring detection has been developed and first results will be discussed.

In this contribution we present weak decay topology recognition utilizing an artifial Neural Network. This approach significantly improved the precision of the analysis of Λ hyperons and K0s mesons and enabled us to measure more rare probes like the Ξ hyperon or Hypernuclei for the first time in heavy-ion collisions with HADES. We highlight the choice of input parameters as well as the applied training proceedure and finally present the obtained results.

LISE quick safety setting simulation using ML

The LISE tool from MSU is widely use to prepare settings for nuclear physics spectrometers around the world. LISE is a very versatile and potent tool including all nuclear physics models needed for precise calculation. The main drawback is a standard setting calculation for a Super-FRS setting, used to check the total particle rate at each focal plane would take up to 2 hours. This is fine for preparing experiments, but a bit problematic to use as a “check is the setting is safe” for an operator. Way to increase the speed of the simulation are investigated.

Machine learning assisted Super-FRS operation

The Super-FRS is built on the same principle than the FRS, but the number of detectors, focal plane and magnets is increased by 6 to 10 folds. This means the standard operation, to be performed in a reasonable time, will require more computer assistance than the FRS operation. There is two solutions to this problem, the first one is to rely more on predictive setting, which would be included in the machine model using LSA, the second is to support the operator in automatizing/guiding some of the task In this paper we explore few ideas where Machine Learning would support Super-FRS operation in commissioning and standard operation phase. The requirement (interface) to permit such work will be presented.

The ARTIFACT project leverages AI to enhance tuning, calibration, beam focusing, and anomaly detection in accelerator systems. By developing an advanced AI framework, the project aims to improve the efficiency and precision of accelerator operations across various scientific applications. A key feature is the creation of a unified metadata schema to structure and separate metadata from different accelerators, enabling cross-system analysis and optimization. This will allow for real-time monitoring and enhanced model performance, driving innovation in fields like high-energy physics, materials science, and medical research.

Accelerator laboratories across the globe are investigating numerous techniques to achieve this goal, including classical optimization, Bayesian optimization (BO), and reinforcement learning. This presentation will provide an overview of recent activities in these domains at GSI. The implementation of the Generic Optimization Framework & Frontend (GOFF) at GSI, supported by the EURO-Labs project, has significantly enhanced workflow, requiring only a few hours to adapt to new accelerators and optimization tasks. The beam loss during injection into the SIS18 was reduced in 15 minutes. GOFF has also been effectively utilized at the GSI Fragment Separator (FRS) for beam steering and focusing.

CRYRNG@ESR is a low-energy ion storage ring and Swedish in-kind contributon to FAIR. It is operational since 2019 serving atomic, nuclear and material physics experiments. Besides this purpose, it serves as a test bench for new FAIR concepts such as novel detectors and control system features. In this context, the "device-automator" application provides an end-user application for implementing various optimizers, in particular a genetic optimizer. It is integrated into the FAIR Control System stack and available for operating in the FAIR CS launcher. We will report on the experiences made using genetics optimization, challenges and future prospects.

To improve beam quality and beam intensity at the Cooler Synchrotron COSY several machine learning and artificial intelligence projects were carried out. Highlights include:
- automated beam line control with Reinforcement Learning
- automated intensity optimisation with Bayesian Optimization for Arbitrary Targets
- improved agreement between machine and model with Genetic Optimization
A digital instance of one of the beamlines was created to test the algorithms independently from the availability of the machine. This basic digital twin offers the same interfaces as the real machine.