Choose timezone
Your profile timezone:
CERN openlab Summer Student Lightning Talks (1/2)
→
Europe/Zurich
503/1-001 - Council Chamber (CERN)
Description
On Monday 17th and Tuesday 18th of August, the 2026 CERN openlab summer students will present their work at two dedicated public Lighting Talk sessions.
In a 5-minute presentation, each student will introduce the audience to their project, explain the technical challenges they have faced and describe the results of what they have been working on for the past two months.
It will be a great opportunity for the students to showcase the progress they have made so far and for the audience to be informed about various information-technology projects, the solutions that the students have come up with and the potential future challenges they have identified.
Please note
- The event will be accessible via webcast for an external audience (Please invite your university professors and other students)
Day 2 information: https://indico.cern.ch/event/1707614/
Please note that pictures and videos might be taken during the event. The pictures and videos might be used for communication about the event. By joining the lecture, you are agreeing to being featured in these communication actions.
Webcast
There is a live webcast for this event
-
-
13:30
→
13:35
Welcome by the CERN openlab team 5m 503/1-001 - Council Chamber
-
13:35
→
13:42
Exploring MCP Tooling for Debugging and Observability in Kubernetes Clusters 7m 503/1-001 - Council Chamber
Project Description:
Kubernetes clusters are becoming increasingly complex, making debugging and fixing issues difficult. We would like to explore the emerging capabilities of agentic AI tooling for interacting with and diagnosing issues in Kubernetes clusters. The project aims to show how MCP-based tooling can make complex systems easier to work with: by giving guidance on how clusters are set up, providing suggestions on how clusters could be optimised (e.g. how best to bin pack nodes, security reconfigurations etc) and suggesting preventative measures to make clusters more resilient to future incidents. The outcome of this project will be an evaluation of the current landscape of MCP tooling for working with Kubernetes broken down by their use cases whilst assessing the feasibility to integrate with the existing MCP based tooling within the section.
Supervisor: Jack Charlie Munday, Diana Gaponcic
Speaker: Elisa Marzioli -
13:42
→
13:49
Securing ORDS on Kubernetes: Transparent mTLS and GitOps worklofws with CNCF Service Mesh 7m 503/1-001 - Council Chamber
Project Description:
Supervisor: Antonio Nappi, Artur Wiecek
Speaker: Filip Piotr Poplewski -
13:49
→
13:56
Apache Airflow POC for Data Reconciliation 7m 503/1-001 - Council Chamber
Project Description: CERN maintains file metadata across Rucio, EOS, and the CERN Tape Archive (CTA), making periodic reconciliation necessary to detect inconsistencies such as dark or missing data. This project investigated Apache Airflow as an alternative to Rundeck for orchestrating the reconciliation workflow. A proof-of-concept Airflow deployment was implemented on CERN’s Kubernetes infrastructure using Argo CD, Git-synced DAGs, CERN S3, and temporary PostgreSQL storage. Overall, a representative EOS–CTA reconciliation DAG was developed based on the Rundeck implementations. The results show that Airflow can represent the workflow’s dependencies, parallel processing, validation, and cleanup stages clearly while providing centralized execution monitoring, although this comes with additional deployment and operational complexity that must be considered before production adoption.
Supervisor: Idriss LarbiSpeaker: Hung Manh Nguyen -
13:56
→
14:03
Applying PQuantML to Large Architectures: Transformers for HEP and Time-Series Data 7m 503/1-001 - Council Chamber
Project Description:
The project studies how a jet tagging model with a transformer architecture trained on synthetic top quark pair production events can be compressed with PQuantML, a library for hardware-aware compression of neural networks during training, for inference on compute-constrained devices such as FPGAs, which are used for ultra-low-latency classification tasks such as jet flavour tagging.
Supervisor: Roope Niemi
Speaker: Dino Liwen Cheng -
14:03
→
14:10
Git-Based Declarative Configuration for the CTA Catalogue 7m 503/1-001 - Council Chamber
Project description:
CERN stores a lot of physics data on tape. Which data lands on which tapes, among other things, comes from a central configuration that operators edit using a command-line tool. This project lets them define it declaratively in Git instead, so changes are easy to review, audit and roll back.Supervisor: Niels Alexander Buegel
Speaker: Daniel Galvan Cancio -
14:10
→
14:17
Benchmarking and Optimisation of High-Performance Data Access and Storage Systems 7m 503/1-001 - Council Chamber
Project Description: This project aims to benchmark and compare different storage solutions used or considered for CERN's computing infrastructure. The main goal is to evaluate a Pure Storage appliance, HDD/SSD baselines, and CERN's production storage systems, including AFS, CephFS, XRootD and EOS. The benchmarking methodology will rely on Flexible I/O (FIO) scripts designed to reasonably reproduce realistic High Energy Physics (HEP) analysis patterns and common user activities. The results of this study will provide a comprehensive performance characterisation of the investigated storage solutions, helping to identify suitable candidates for the next deployments and improvements to CERN's computing services.
Supervisor: Guilherme Amadio, Luca Mascetti
Speaker: Gustavo Mattos Lopes -
14:17
→
14:24
Container Checkpointing for interactive and Batch/ML workloads 7m 503/1-001 - Council Chamber
Project Description:
Investigate and integrate container checkpointing solutions for CPU/GPU HPC Workloads to enable seamless workload migration during routine cluster maintenance and intelligent suspension/resumption of idle sessions.Supervisor:
Raulian-Ionut Chiorescu
Hannes Jakob HansenSpeaker: James Bhattarai -
14:24
→
14:31
Composable Infrastructure and Green Scheduling for Large-Scale Scientific Computing 7m 503/1-001 - Council Chamber
Project Description:This project evaluates two complementary areas aiming at improving energy efficiency in HPC or Cloud based deployments: composable infrastructure, enabling dynamic runtime composition and assignment of GPUs and other specialized resources; and optimized scheduling and suspension of non-critical workloads when possible. The student will: Study the available options in both areas, considering tools such as CoHDI for composable infrastructure or EAR and kube-green for energy aware scheduling, deploy them in test clusters, with the goal of measuring operational impact on overall utilization, scheduling efficiency and savings, identify limitations and integration challenges, highlighting synergies as well as incompatibilities between HPC and cloud environments.
Supervisors: Laura Eve Sarah Llinares, Matteo Bunino.
Speaker: Sebastian Andres Uribe Ruiz -
14:31
→
14:38
Energy based training of quantum circuit Born machines 7m 503/1-001 - Council Chamber
Project Description:
Supervisor: Michele Grossi, Cenk Tuysuz
Speaker: Hassan El Bouz -
14:38
→
14:45
Integrating LLM proxy with CERN infrastructure 7m 503/1-001 - Council Chamber
Project Description: CERN is consolidating access to its AI resources behind a single gateway: models served on its own GPUs, models bought from commercial providers, and increasingly MCP servers and agents. Placing them on one path allows access to be granted once from CERN identity and group definitions rather than arranged separately for each provider, and makes usage, cost and reliability visible for the organisation as a whole instead of scattered across teams and invoices. This project covers the integration of that gateway, LiteLLM, with the services CERN already runs: a dedicated development instance was deployed, the gateway's Prometheus metrics were onboarded into MONIT and a Grafana dashboard was built over them, and alerting and budget controls were configured.
Supervisors: Juan Manuel Guijarro, Ricardo Rocha
Speaker: Nour Guermazi -
14:45
→
14:52
Advancing NeutrinoReview: Creating an End-to-End Platform for Systematic Reviews with Large Language Models and Automated Full-Text Retrieval 7m 503/1-001 - Council Chamber
Project Description:
Systematic reviews (SRs) are widely regarded as the most rigorous method for scientific evidence synthesis, but conducting them is highly time-consuming. The NeutrinoReview platform is a proof-of-concept application to automate the labor-intensive steps of systematic reviews by integrating LLM screening techniques like the CAL-X algorithm, with a strong emphasis on reliability and high sensitivity.
This project focuses on extending NeutrinoReview with an interface for manual fulltext screening along with a legally compliant and resilient automated retrieval pipeline with multiple fallback mechanisms. By these changes, NeutrinoReview becomes an end-to-end prototype for efficient systematic reviews, combining the potential of LLMs and human-in-the-loop screening.
Supervisor: Elias Sandner, Andreas Wagner
Speaker: Thomas Edwin Steinberger -
14:52
→
14:59
Natural-Language Querying of Large-Scale Industrial Control Systems Data 7m 503/1-001 - Council Chamber
Project Description:
Supervisor: Rafal Kulaga
Speaker: Melchior Vos -
14:59
→
15:14
Coffee Break 15m 61/1-201 - Pas perdus - Not a meeting room -
-
15:14
→
15:21
Agentic AI automation of UNICOS application specification processes 7m 503/1-001 - Council Chamber
Project Description:
This project explores the use of agentic AI for PLC/UNICOS engineering workflows at CERN. The aim is to support engineers in two main tasks: Functional Analysis quality and compliance checking and UNICOS specification generation. The system processes Functional Analysis documents and checks them against engineering guidelines to identify missing, incomplete, or non-compliant requirements. It also builds knowledge bases from available engineering information, including Functional Analyses, IO lists File, UNICOS documentation, and specification templates to support specification generation and validation. The solution combines multimodal document processing, Retrieval-Augmented Generation (RAG), and Model Context Protocol (MCP) to create an engineering copilot that can interact with existing tools while keeping the engineer in control of the final decisions.Supervisor: Filippo Berto
Speaker: Daniya Niazi -
15:21
→
15:28
Evaluation of Cerabyte: Archival Data Storage Technology using Ceramic Nanolayers 7m 503/1-001 - Council Chamber
Project Description:
Supervisor: Vladimir Bahyl
Speaker: Andreas Bräuer -
15:28
→
15:35
Anomaly detection with transformer models on AMD Versal 7m 503/1-001 - Council Chamber
Project Description:
This project targets the deployment of real-time anomaly detection algorithms on FPGA devices to identify physics phenomena that may evade standard, model-driven trigger selections in the CMS experiment. The central objective is the implementation of a ML architecture based on a transformer model, adapted for ultra-low-latency inference in hardware.
This project investigates whether modern heterogeneous hardware platforms, specifically AMD Adaptive Compute Acceleration Platforms (ACAPs), can meet our strict latency, throughput, and determinism requirements. These devices integrate programmable logic, processors, and Adaptive Intelligence (AI) Engines, offering a promising path toward scalable, real-time ML-based triggering and scouting solutions for CMS and future experiments.
Supervisor:
Elias Leutgeb
Thomas Owen JamesSpeaker: Eliot Mario Abramo -
15:35
→
15:42
Low latency anomaly detection for the ATLAS trigger system 7m 503/1-001 - Council Chamber
Project Description:
The project develops a low-latency machine-learning anomaly detector for the ATLAS Level-1 trigger, where collision events must be processed in real time under strict latency and FPGA resource constraints. A normalizing flow is trained on background collision data to learn their probability distribution and assign high anomaly scores to unusual events.
Because the normalizing flow is too complex to deploy directly on the trigger hardware, knowledge distillation is used to transfer its anomaly score to a much smaller neural network. The student model is then quantized using HGQ2 and converted with hls4ml for FPGA synthesis, allowing its latency, timing, and resource usage to be evaluated for deployment in the ATLAS trigger system.
Supervisors: Paula Martinez Suarez, Stefano Veneziano
Speaker: Jorgen Bergh (CERN) -
15:42
→
15:49
CMS physics reconstruction running on NextSilicon Maverick accelerators 7m 503/1-001 - Council Chamber
Project Description:
This project evaluates the suitability of NextSilicon’s Maverick-2 dataflow accelerator for HPC workloads at CERN using the CLUE clustering algorithm as a benchmark. The CLUE implementation is adapted and optimized for execution on Maverick-2 and compared against a CPU OpenMP baseline. The study analyzes kernel-level performance, end-to-end execution time, memory movement, and accelerator offload overheads. The results provide insight into the benefits and current limitations of dataflow architectures for CMS reconstruction workloads.
Supervisor: Andrea Bocci and Mario Gonzalez
Speaker: Fernando Antonio Marques Schettini -
15:49
→
15:56
CMS physics reconstruction running on Altera FPGAs 7m 503/1-001 - Council Chamber
Project Description:
Supervisor: Andrea Bocci and Mario Gonzalez Carpintero
Speaker: Farid Abi Doumit -
15:56
→
16:03
Exploration of state-of-the-art AI-acceleration platforms for HEP use cases 7m 503/1-001 - Council Chamber
Project Description:
Exotic AI-accelerator hardware, such as the CEREBRAS Wafer Scale Engine, and the Tenstorrent Blackhole processing board offer compelling alternatives to the standard CPU+GPU offerings for AI training and inference. Within CERN openlab we would like to investigate these solutions and understand which. The student will work to port our existing AI models to one or several of these exotic platforms (depending on hardware availability), documenting the trials and tribulations of the process, and eventually resulting in fair benchmark comparisons between the different hardware approaches. This experience will then help to inform the wider community at CERN of the potential pros and cons of developing for these alternative AI platforms.
Supervisor:
Thomas Owen James
Elias LeutgebSpeaker: Grace Margaret Rossiter -
16:03
→
16:10
Benchmarking, Evaluation & Automated Metrics 7m 503/1-001 - Council Chamber
Project Description:
BEAM is an internal platform for monitoring, managing, and reporting on the health and utilisation of compute nodes managed by the CERN openlab-systems group. It ingests telemetry from a fleet of machines via a lightweight Puppet agent, consolidates data with external authoritative sources (Puppet Git config, OpenDCIM, LanDB), and exposes everything through a REST API consumed by a web application.
Supervisor: Albane Carcenac and Jessy Sobreiro
Speaker: Suhani Bansal -
16:10
→
16:17
Improving AMD support for hardware-accelerated event generation 7m 503/1-001 - Council Chamber
Project Description: Monte-Carlo event generators are used in High-Energy Physics (HEP) to simulate collider events. They are a cornerstone in the physics programmers of Large Hadron Collider (LHC) experiments such as ATLAS and CMS. Given the very high experimental requirements on precision and the strong trend of increasingly relying on hardware accelerators such as GPU for high-performance computing, our team develops novel parallelized event generators that are accelerated by GPU devices and CPU vector instructions.
The student will study and improve AMD GPU support for the event generators Pepper and MadGraph5_aMC@NLO. This involves adding support for HIP in the widely used LHAPDF library, which both generators use to evaluate the quark and gluon content of the incoming protons. The student will further conduct a systematic profiling of the two generators on AMD hardware, to identify possibilities for performance improvements. The implementations and performance studies will prepare deployment of the hardware-accelerated event generators on AMD-based sites of the Worldwide LHC Computing Grid and on HPC clusters, to help alleviate the projected CPU budget limitations in the upcoming High-Luminosity era of the LHC, and to prepare event generation for future collider experiments.
Supervisor: Daniele Massaro and Enrico Bothmann
Speaker: Caitlin Buch (Niels Bohr Institute) -
16:17
→
16:24
Towards Fault Propagation Modeling in the WLCG Using Graph Neural Networks 7m 503/1-001 - Council Chamber
Project Description:
The Worldwide LHC Computing Grid (WLCG) is a large-scale distributed infrastructure supporting data processing and analysis for CERN experiments. Due to its complexity and the dependencies between its services, failures or degradation in individual components may affect other parts of the system. Understanding and predicting these effects could help improve its reliability and operational efficiency.
Graph Neural Networks are well suited to modelling interconnected systems. This project investigates whether the WLCG can be represented as a temporal graph and whether GNNs can use its topology to predict operational degradation. FTS monitoring data is used as an initial proxy, providing a first step towards data-driven modelling of the WLCG.
Supervisor: Maria Del Carmen Misa Moreira, Sofia Vallecorsa
Speaker: Pavel Khudov Yakovlev
-
13:30
→
13:35