2nd GSI/FAIR AI Workshop

This talk gives a brief overview of the IT infrastructure supporting AI activities at GSI/FAIR, with a particular focus on emerging inference and LLM services. It will also present the GSI strategy for the professional use of AI-related tools, including considerations around data protection, trade secrets and information security, and outline how infrastructure and institutional guidance can support the growing use of AI across the campus.

We present a versatile production-grade LLM inference and document intelligence server designed for GSI/FAIR computing infrastructure.

The system employs a modern containerized architecture built on vLLM for GPU-accelerated inference, OpenWebUI for user interaction, and Milvus vector database for Retrieval-Augmented Generation (RAG) capabilities.
The deployment utilizes several specialized vLLM instances running multiple models with support for thinking, tool calling and vision capabilities. A dedicated embedding model for document vectorization and retrieval is available as well.
This multi-model approach, deployed on an 8xH200 GPU server, balances computational cost with task-specific performance requirements.
The RAG implementation enables users to upload and query documentation, code repositories, and technical papers using natural language.
API access allows users to explore and deploy their own solutions while leveraging high performance hardware.
The architecture supports production deployment on a server with enterprise authentication and development configurations for testing on resource-constrained hardware.

Performance characteristics, deployment patterns, and lessons learned from production usage are presented, along with token consumption metrics across different workflows.
This work demonstrates that modern LLM technologies can be successfully integrated into GSI/FAIR computing infrastructure while maintaining security, auditability and performance.

PowerSched presents a proof of concept for deep reinforcement learning-based, price-aware capacity control in high-performance computing clusters, dynamically adjusting powered resources in response to fluctuating electricity prices and workload demand. Evaluations using realistic price and workload traces demonstrate that measurable cost savings can be achieved while maintaining complete job execution and introducing only small increases in waiting time.

The configuration parameters of the SLURM guide, the working of the scheduler. The efficiency of the cluster where SLURM is run can be optimised with the help of SLURM simulator adapted and tailored to the work load of the HPC cluster.

We investigate fast methods for evaluating particle hypotheses by combining signals from multiple detectors during the early readout stage of the data-acquisition (DAQ) chain. As a first use case, we consider straight particle trajectories reconstructed from straw-tube drift-time measurements and time-of-flight (TOF) wall responses, initially neglecting energy loss. A compact multilayer perceptron implemented in PyTorch is trained on reconstructed spatial and timing observables to produce a muon-compatibility score. Its performance is compared with the classical chi2 selection in terms of muon efficiency, background rejection, and momentum dependence. This study provides a proof of concept for combining detector information with lightweight machine-learning models, with the longer-term goal of deployment in low-latency, real-time data-processing systems.

Artificial intelligence methods developed for one scientific domain often struggle to generalize to other fields that produce irregular and high dimensional detector data, creating a need for domain agnostic reconstruction approaches.
BRAID is a research consortium developing machine learning models particularly graph neural network and attention-based architectures, that can work across this kind of irregular, high-dimensional data from experiments in high energy physics, astroparticle physics, hadron and nuclear physics.
Here, we give an overview of this recently started research initiative, with a focus on its first use cases in experiments in hadron physics at GSI/FAIR.

We present HEIDI, a deep learning-based conditional diffusion model for the ultra-fast generation of event-by-event heavy-ion collision output. Trained on UrQMD simulations, HEIDI learns to generate point clouds of final-state particles and accurately reproduces the multiplicity and momentum distributions of 26 different hadron species produced in UrQMD. Compared to standard UrQMD cascade simulations, HEIDI achieves a speedup of roughly three orders of magnitude, while preserving the physical correlations between particles within an event. We demonstrate that HEIDi can be applied to other particle-shower problems, using the example of cosmic ray air showers. These results highlight HEIDI's potential as a flexible and efficient AI tool for both theoretical modelling and experimental data analysis, where fast and reliable event generation is essential.

We present an unsupervised machine learning approach to automate quality 
monitoring in the HADES experiment, combining a Variational Autoencoder 
with HDBSCAN clustering (VAE-HDBSCAN) to detect anomalies in detector 
performance plots without needing to hand-label data. Our method 
outperforms non-latent clustering and approaches the performance of 
supervised CNN baselines, and is being integrated into Jefferson Lab's 
HYDRA web platform for operator-friendly deployment during beamtime.

The CBM experiment is a complex detector system composed of numerous subdetectors operating simultaneously and generating thousands of monitoring plots to ensure stable and reliable performance. Given the massive data volumes expected during data taking, continuous verification of data quality and the prompt identification of technical faults are essential. This project proposes the adoption of an AI-based automatic anomaly detection framework, HYDRA, originally developed for the GlueX experiment at Jefferson Lab. HYDRA employs computer-vision models for near–real-time image classification to recognize irregular patterns in detector monitoring outputs. The project focuses on adapting and deploying HYDRA for CBM, including the training and maintenance of dedicated models for individual subdetectors such as STS and MUCH, as well as improving HYDRA capabilities to facilitate unsupervised learning using Autoencoders.

A unified framework based on machine learning for the identification of hadrons by combining responses from different sub-detectors is developed for the CBM experiment. In the first iteration, gradient-boosted decision trees (xGBOOST) are used as base models. This contribution focuses on the implementation of the models and the recent results achieved through their application. Furthermore, a comparison to the previously used conventional cut-based method is presented.

Abstract: AI-assisted operation of the Super-FRS will be tested during the commissioning activities in 2027. The approach treats the language model as a reasoning engine rather than a knowledge base: it retrieves from the facility's own data, computes with tested tools, and cites what it used. The data itself already exists — S-FRSRoot delivers physics-grade output from the DAQ streams, and the EPICS layer delivers the machine state. Everything runs on local hardware, so no data leaves GSI.
Two agents are in operation today. The first answers questions from the PID detector documentation and gives the page it came from. The second investigates EPICS slow-control parameters through a fixed set of read-only analysis tools, so every number it reports comes from a real measurement.
Commissioning will put both under real conditions, and it will also shape the next generation: predictive maintenance on an archived machine state, a shift assistant that carries context across hand-overs, AI-assisted beam tuning, and an agent that writes and compiles S-FRSRoot analysis classes on request, so that a new observable costs a sentence instead of a week of coding.
The same recipe applies to any experiment that has documentation, a control system, and an analysis framework.

An AI chat assistant based on open-source large language models (LLMs) to provide context-aware guidance and answer questions related to the GSI/FAIR accelerators and their operation. The model runs locally, using prompt engineering and retrieval-augmented generation (RAG) to gain access to shift data from operations and domain knowledge spread across the various GSI wikis.

A fully on-premise hybrid RAG system that makes electronic logbook entries searchable in natural language, surfacing expert knowledge that keyword search misses.

SIS-SYS develops beam-based feedback, automated control and expert-diagnostic systems for SIS18 and SIS100, combining classical DSP and control methods with numerical models and, increasingly, ML/AI-based approaches.

A central focus is how beam-data-driven algorithms are qualified and maintained over the long term through physics-informed digital twins that provide a reproducible environment for replay, stress testing and direct benchmarking ML/AI-based against established classic algorithms and measurable ground truth, as well as online/offline data analysis and offline debugging.

ML/AI-based methods are not privileged, but only one of many solutions; all must demonstrably match or outperform existing classical approaches while satisfying requirements for reproducibility, robustness, diagnosability, bounded failure behaviour and long-term maintainability.

This talk presents selected examples from beam-based diagnostics and feedback systems and GREMLIN, illustrating this approach towards a dependable 24/7 FAIR accelerator operation.

Since 10 years we are maintaining a Java application "DeviceAutomator" ready for use in the control room. Step by step this application is enhanced by AI-powered functionality - classical optimizers until now. We want to report on the latest developments, application scenarios and how we use DeviceAutomator in our daily work.

This contribution presents the ideas for an AI-assisted beamline optimisation framework for the FRS/Super-FRS Ion Catcher. The project will explore the use of machine-learning algorithms together with ion-optical simulations and offline measurements to automate tuning of the low-energy beamline and reduce reliance on manual optimisation.

The GSI/FAIR facility’s complexity demands a high degree of automation for future operations. As part of this effort, we developed Geoff, a framework for real-time optimization of beam parameters and experimental setups designed for fast deployment and control room integration. Geoff significantly improves operational performance, having reduced SIS18 synchrotron injection losses from 45% to 12% and accelerated fragment separator setup using a classification algorithm. Furthermore, using dedicated ion-source setups of the PUMA experiment at TU Darmstadt, we demonstrated the feasibility of real-time tuning for ion sources and particle traps by optimizing a hot-cathode electron source and a multi-reflection time-of-flight mass spectrometer. Geoff’s modular design facilitates seamless integration of both classical and machine-learning-based algorithms, bridging traditional accelerator operations with modern data-driven optimization. Geoff was supported by EURO-LABS and will receive further EU funding through the TwinRise, iRIS, and HADRON-2030 initiatives.

At the GSI Helmholtz Centre, civil engineering, research, plant engineering and operations work in separate data and tool environments. FedTwin does not replace these systems but places an integration layer on top of them. The existing building and asset data remain in their respective domain systems and are made accessible across disciplines through an access layer based on the Model Context Protocol (MCP), because what is unified is the access, not the tool. This takes Building Information Modelling (BIM) a step further. Where BIM structures building information in a model-based way, FedTwin opens that space to disciplines without BIM: AI agents translate between the domain languages and ground their answers in verified as-built data via Retrieval-Augmented Generation (RAG). Manual searches across system boundaries are thereby eliminated, and the Federated Digital Twin evolves from a passive data model into a proactive, source-referenced basis for decision-making.

The Helmholtz Information & Data Science Academy (HIDA) is dedicated to the development of information and data science skills within the Helmholtz Association. HIDA offers a comprehensive range of training opportunities and promotes the exchange of knowledge between scientists through funding, mobility programs, and networking events. HIDA also serves as the umbrella organization for seven Helmholtz Data Science Schools. Together, HIDA and the Helmholtz Data Science Schools form Germany's largest postgraduate training network in the field of information and data science.

AT A GLANCE
- Data science courses
- AI literacy training & LLM courses
- Course funding opportunities
- Lectures, events & summer schools
- Mobility programs
- Data science job board

AI agents are getting good at doing pieces of research, such as searching literature, writing code, working with data, and making plots. But a useful research agent needs more: it needs to learn how a community actually works.

In this talk, I’ll introduce Ori (Open Research Intern), our AI research assistant at AIP, and the Skill Commons, a database where researchers can share practical know-how as skills. Ori can search the Commons, bring in new skills, and contribute back the knowledge it learned. The aim is not one agent that knows everything, but an ecosystem in which researchers teach their agents, and those agents learn from one another. It starts at AIP, but doesn't stop there.