Talks

**Project Description**: This project develops a reproducible, hardware-aware MLOps pipeline for optimizing neural-network models for LHC hardware triggers. It combines automated architecture and quantization search, experiment tracking, physics-level validation, and FPGA deployment, with a reusable workflow demonstrated on an MHA-based minimum-bias anomaly-detection use case. **Supervisor**: Ioannis Xiotidis

00:00:00 Slide 1
00:00:40 Slide 2
00:02:16 Slide 3
00:02:39 Slide 4
00:03:31 Slide 5
00:04:03 Slide 6
00:04:34 Slide 7

**Project Description**: The CMS experiment relies on rapid event reconstruction at the High-Level Trigger to select interesting proton–proton collisions in real time. For the upcoming Run 4, the Next Generation Trigger (NGT) scouting stream aims to record events at exceptionally high rates by utilizing dedicated fast reconstruction algorithms and reduced data formats. This presentation evaluates the feasibility of the Jet-Plus-Tracks (JPT) algorithm as a lightweight alternative to the computationally intensive Particle Flow (PF) reconstruction. Using simulated Phase-2 CMS data, the study compares both approaches in terms of basic physics quantities, such as jet energies and event activity, as well as their computing footprint, specifically execution runtime and data size. The results highlight the trade-offs between physics performance and computing costs in the context of the Phase-2 scouting ecosystem. **Supervisor**: Marco Musich, Marco Rovere

00:00:00 Slide 1
00:00:05 Slide 2
00:00:21 Slide 3
00:01:09 Slide 4
00:01:31 Slide 5
00:01:48 Slide 6
00:02:06 Slide 7
00:02:53 Slide 8
00:03:18 Slide 9
00:03:48 Slide 10
00:04:07 Slide 11
00:04:56 Slide 12
00:05:23 Slide 13
00:06:14 Slide 14
00:07:18 Slide 15

**Project Description**: Machine learning models are widely used at CERN for a variety of tasks, including data analysis, pattern recognition, and anomaly detection. These models are often highly sophisticated in order to handle the complexity of the data they process. In addition, when deployed in real-time systems such as trigger pipelines, they must deliver extremely fast inference performance. With the upcoming upgrades of the LHC experiments for the High-Luminosity phase, machine learning models are expected to become even more complex while simultaneously needing to process significantly higher data rates. This places stringent requirements on inference latency, memory usage, and hardware efficiency. SOFIE (System for Optimized Fast Inference code Emit) is a tool that translates trained machine learning models into optimized C++ code with BLAS as its only dependency. In its current implementation, SOFIE-generated code has demonstrated up to 2× faster inference and approximately 50% lower memory consumption on CPUs compared to ONNX Runtime for commonly used models in high-energy physics research. SOFIE is also being extended to support inference on heterogeneous architectures through alpaka, an abstraction library for portable heterogeneous programming. In its next iteration, SOFIE will have integrations with PQuant, a library for end-to-end hardware-aware model compression using pruning and quantization techniques. This integration will enable the generation of C++ inference code from quantized machine learning models. **Supervisor**: Sanjiban Sengupta, Lorenzo Moneta

00:00:00 Slide 1
00:00:12 Slide 2
00:00:34 Slide 3
00:01:01 Slide 4
00:01:37 Slide 5
00:03:07 Slide 6
00:03:41 Slide 7
00:03:53 Slide 8
00:04:23 Slide 9
00:04:45 Slide 10
00:05:07 Slide 11
00:06:06 Slide 12

**Project Description:** This project focuses on optimising the ATLAS HL-LHC trigger menu for Higgs-pair production, targeting the HH -> bbτ(had)τ(lep) final state. The goal is to maximise the signal efficiency while maintaining the background rate within the available trigger budget. To achieve this, the muon and electron channels are optimised separately across the trigger chain. At Level-1, different object definitions and preselections are studied, combining large grid scans with BDTs to identify efficient and interpretable selections. The analysis is then extended to the Event Filter, where b-tagging and additional identification requirements provide further background rejection. **Supervisors:** Christopher Young and Tamara Vázquez Schröder

00:00:00 Slide 1
00:00:24 Slide 2
00:01:06 Slide 3
00:02:08 Slide 4
00:02:48 Slide 5
00:03:47 Slide 6
00:04:52 Slide 7
00:05:14 Slide 8
00:06:00 Slide 9
00:06:37 Slide 10
00:08:26 Slide 11

**Project Description**: Accurate physics reconstruction of collision events at the LHC relies on precise knowledge of detector conditions at the moment an event is acquired. As these conditions fluctuate during data-taking, calibrations must be derived to account for them. Within the Next Generation Triggers (NGT) project, a novel approach is being tested to derive calibrations for online reconstruction at the High-Level Trigger (HLT) with much higher frequency than current operations. A demonstrator has been commissioned to test the feasibility of this new paradigm for the Phase-2 conditions framework, with the aim of deriving calibrations continuously within a strict 8-hour window while using limited computing resources and maintaining physics quality. This summer student project will focus on a retrospective, comprehensive physics validation of this demonstrator using results collected during the final Run 3 data-taking period. The student will perform a comparative analysis between the conditions derived by the NGT Demonstrator, the standard HLT conditions, and the high-precision Offline conditions which serve as the reference truth. The goal is to quantify improvements in physics object reconstruction and assess triggering efficiencies, to determine if the continuous calibration approach is feasible for Phase-2. **Supervisor**: Jessica Prendi, Marco Musich

00:00:00 Slide 1
00:00:14 Slide 2
00:00:43 Slide 3
00:01:00 Slide 4
00:02:21 Slide 5
00:04:07 Slide 6
00:05:06 Slide 7
00:05:35 Slide 8
00:06:14 Slide 9

**Project Description**: Machine learning is rapidly replacing heuristic-based algorithms across many domains, and ATLAS plans to make extensive use of ML in both its hardware and software triggers during the HL-LHC era. While mature optimization platforms exist for deploying ML in software, very few MLOps pipelines targeting custom FPGA-based electronics have been tested at scale. Within Next-Generation Triggers (NGT) WP2.1, we are developing a dedicated MLOps pipeline that automatically optimizes and deploys ML algorithms for the ATLAS L0-Global hardware trigger. A prototype version already exists and is in use, while a more generalized approach, one that can potentially extend beyond the specific scope of WP2.1, is being developed in parallel. This work covers three areas: generating open data samples for validating pipeline developments; developing a command-line interface that allows users to create standalone ML algorithms, integrate them into the pipeline, and deploy custom pipeline stages; and integrating a software stack dedicated to the joint optimization of multiple trigger algorithms. The goal is a service-level design through which ML developers for hardware triggers can easily obtain deployable, optimized trigger algorithms that efficiently utilize the underlying hardware. Such a service significantly lowers the barrier to integrating ML into the ATLAS hardware trigger and ensures that common, well-defined metrics are used to evaluate whether deploying an ML model there provides a real advantage. **Supervisor**: Ioannis Xiotidis, Noah Clarke Hall

00:00:00 Slide 1
00:00:01 Slide 2
00:00:13 Slide 3
00:02:24 Slide 4
00:03:58 Slide 5
00:04:21 Slide 6

**Project Description**: With the start of the High Luminosity LHC era, or "phase-2," the upgraded machine will provide unprecedented volumes of data at extreme energies and intensities. Extracting meaningful physics requires advanced data-analysis techniques and algorithms. Within the CMS experiment, we are developing an innovative online trigger system to handle the enormous data rates expected in phase-2. This will be complemented by a novel Next-Generation Trigger (NGT) Scouting strategy, which records physics events at the full input rate while storing a reduced set of reconstructed physics objects. This approach aims to provide access to rare and unconventional signatures that would otherwise be inaccessible due to data-rate and storage limitations. In the CMS Tracker, we are exploring the use of phase-2 pixel tracks as seeds for the Gaussian Sum Filter (GSF) tracking algorithm. Currently, trajectory seeds are computed on-the-fly using tracker and calorimeter information, despite existing tracks being available. Historically, this was due to the low efficiency of pixel tracks as GSF seeds. Our goal is to assess the feasibility of using phase-2 pixel tracks, particularly NGT's recently developed extended pixel tracks, as GSF seeds under different track-quality criteria. If GSF physics performance is maintained, eliminating trajectory-seed computation could significantly accelerate the CMS HLT, making this a high-impact contribution. **Supervisor**: Bruno Alves, Wahid Redjeb

00:00:00 Slide 1
00:00:08 Slide 2
00:00:37 Slide 3
00:01:15 Slide 4
00:01:45 Slide 5
00:02:15 Slide 6
00:02:51 Slide 7
00:03:36 Slide 8
00:03:55 Slide 9
00:04:34 Slide 10
00:05:05 Slide 11
00:05:34 Slide 12
00:06:39 Slide 13
00:06:59 Slide 14
00:07:13 Slide 15
00:07:56 Slide 16
00:08:08 Slide 17
00:08:46 Slide 18
00:09:28 Slide 19
00:10:18 Slide 20
00:10:44 Slide 21
00:11:13 Slide 22
00:11:40 Slide 23
00:11:59 Slide 24
00:12:23 Slide 25
00:13:02 Slide 26

**Project Description**: Muon reconstruction is a central component of the ATLAS physics program and a key contributor to the overall computational cost of event reconstruction. The muon spectrometer provides precision tracking in the outermost region of the detector, and muon segment fitting is a core step in the outside-in reconstruction chain. Our group has implemented a highly optimized CUDA-based $\chi^2$ minimization fitter. However, approximately half of the current runtime is spent on host–device memory transfers, limiting suitability for latency-critical applications such as triggering. The proposed project aims to build on this foundation by developing a deeper GPU-resident reconstruction pipeline in which additional steps of the muon reconstruction chain are executed directly on the GPU. By moving hit preparation, region construction, and segment seeding onto the device, the project seeks to minimize data movement and amortize memory transfer costs. The reconstruction will be integrated more deeply with the A Common Tracking Software (ACTS) framework using a realistic ATLAS detector geometry and more complex physics processes, including jets and pileup. A full benchmarking campaign will be carried out to compare the GPU-based pipeline with the existing CPU implementation and to assess the feasibility of GPU acceleration for later-stage triggering and quasi-online reconstruction. **Supervisor**: Johannes Junggeburth, CERN

00:00:00 Slide 1
00:00:24 Slide 2
00:01:09 Slide 3
00:01:38 Slide 4
00:02:20 Slide 5
00:02:42 Slide 6
00:03:10 Slide 7
00:03:49 Slide 8
00:04:54 Slide 9
00:06:34 Slide 10

**Project Description**: Hls4ml is an open-source toolkit that converts trained machine-learning models into FPGA firmware using high-level synthesis, enabling ultra-low-latency inference for real-time applications such as LHC trigger systems. PandA/Bambu is an advanced open-source HLS framework that compiles C/C++ into optimized hardware designs and offers a rich set of optimization passes and backend targets. In this project, the student will develop a new hls4ml backend that generates PandA/Bambu-compatible C++ and integrates Bambu's HLS flow into the hls4ml build pipeline. The work includes implementing dataflow-friendly code generation, tuning Bambu optimizations for ML workloads, and benchmarking performance and resource usage on FPGA targets. This project provides hands-on experience with HLS, ML inference acceleration, and open-source hardware design tools. **Supervisor**: Nicolò Ghielmetti, George Flengas, Dimitrios Danopoulos

00:00:00 Slide 1
00:00:02 Slide 2
00:00:12 Slide 3
00:01:30 Slide 4
00:02:12 Slide 5
00:02:30 Slide 6
00:03:23 Slide 7
00:04:08 Slide 8
00:05:32 Slide 9
00:05:44 Slide 10
00:06:26 Slide 11
00:07:57 Slide 12
00:08:31 Slide 13

This project builds a live, bi-directional bridge between two robotics simulators, MuJoCo and NVIDIA Isaac Sim, to support robotic maintenance interventions inside a 3D digital twin of the LHC. MuJoCo acts as the single source of truth for physics, while Isaac Sim mirrors the simulation in real time, rendering it photorealistically inside CERN's actual accelerator geometry; control commands flow back the other way, so a robot can be driven through the tunnel live. The motivation is practical, maintenance in the LHC tunnels is slow, costly, and constrained by radiation limits on human access, so being able to test whether a robot can reach and operate on a target, without collisions before it ever enters the real tunnel is valuable. Alongside the bridge, the project implements collision detection against the tunnel's CAD geometry and demonstrates general applicability by driving a quadruped robot through the twin with a game controller. Furthermore there is VR-based control, allowing an operator to step inside the twin and rehearse an intervention before anyone, or any robot, enters the tunnel.: **Supervisor**: Drew Mason-Laurence

00:00:00 Slide 1
00:00:10 Slide 2
00:00:31 Slide 3
00:00:47 Slide 4
00:01:04 Slide 5
00:01:47 Slide 6
00:02:14 Slide 7
00:02:46 Slide 8
00:03:03 Slide 9
00:03:13 Slide 10
00:03:38 Slide 11
00:04:13 Slide 12
00:04:38 Slide 13
00:05:04 Slide 14
00:06:17 Slide 15

**Project Description**: The objective is to use Machine Learning to derive optimal representations of LHC collision events that naturally cluster Standard Model processes while maximizing sensitivity to anomalous signatures beyond the Standard Model (BSM). The project will focus on applying state-of-the-art representation learning methods, including unsupervised (Autoencoder) and self-supervised (BERT, JEPA, VICReg, DINO) learning, to compress high-dimensional LHC event data into descriptive latent spaces for trigger decisions. You will evaluate existing anomaly-detection algorithms and their latent-space representations, design custom training strategies to produce optimal representations, and analyze them in terms of their underlying statistics and correlations with physical observables. Finally, you will leverage NGT resources to study scaling laws and understand how representations evolve with training sample size and model capacity. **Supervisor**: Sabrina Giorgetti, Maciej Mikolaj Glowacki

00:00:00 Slide 1
00:00:36 Slide 2
00:01:42 Slide 3
00:02:59 Slide 4
00:04:26 Slide 5
00:05:16 Slide 6
00:05:35 Slide 7
00:07:43 Slide 8