Argonne Hackathon Helps Researchers Accelerate Scientific Discovery on Aurora

retained text

Off the Wire Press Releases
July 29, 2026 — The Argonne Leadership Computing Facility (ALCF) recently hosted its annual hackathon, bringing together researchers and developers to prepare scientific applications for large-scale computing on the facility’s Aurora and AI Testbeds. Held at the U.S. Department of Energy’s (DOE) Argonne National Laboratory, the event connected participants with ALCF staff mentors who provided hands-on support for optimizing and scaling high performance computing (HPC) and artificial intelligence (AI) workloads.
The teams worked closely with ALCF experts to improve application performance, address technical challenges, and demonstrate readiness for leadership-class computing resources. The hackathon supports researchers preparing proposals for DOE allocation programs, including INCITE (Innovative and Novel Computational Impact on Theory and Experiment) and ALCC (ASCR Leadership Computing Challenge), while also helping current users advance their projects on ALCF systems. The ALCF is a DOE Office of Science user facility at Argonne.
“The hackathon gives research teams an opportunity to work directly with our staff and gain valuable experience on Aurora and Polaris before launching large-scale production runs,” said ALCF computational scientist Yasaman Ghadar, who directed the event. “By the end of the program, participants have not only improved the performance of their applications, but also developed a deeper understanding of how to make the most effective use of leadership computing resources. That preparation can have a significant impact on the success of their future research efforts.”
The ALCF’s annual hackathon continues to serve as a key entry point for researchers seeking to leverage advanced computing and AI technologies to accelerate scientific discovery.
This year’s event covered a wide range of topics, including AI/ML Frameworks on Aurora, TAU performance analysis toolkit, Intel debugging tools, an Linaro Forge tools.
Building Intelligent Workflows for the Next Generation of Synchrotron Science
Tim Dunn, an engineering physicist at the Stanford Synchrotron Radiation Lightsource (SSRL) at DOE’s SLAC National Accelerator Laboratory, came to the ALCF hackathon with a goal of making scientific computing a more active part of live experiments. Working as part of the Light Sources SYNAPS team alongside collaborators from SSRL and Lawrence Berkeley National Laboratory’s Advanced Light Source (ALS), Dunn sought to explore how AI and leadership-class computing could help researchers process and interpret tomography data more quickly and efficiently.
“The research I am pursuing is about making synchrotron experiments much more automated, and eventually autonomous,” says Dunn. “At ALCF, I wanted to take the next step. Instead of treating HPC as a place where we simply submit a fixed batch job and wait for the result, I wanted to explore whether we could run long-lived scientific agents on HPC that communicate with a beamline agent, evaluate their own progress, and make the workflow observable while it is running.”
Before the hackathon, Dunn’s team had already developed an automated workflow that transfers tomography data from SSRL to ALCF’s Polaris supercomputer for reconstruction and AI-assisted analysis. During the event, the team worked to transform that workflow into an agent-based system capable of coordinating across facilities and monitoring its own progress. Among the milestones achieved was the successful launch of software agents within Polaris jobs that could perform GPU-accelerated tomography reconstruction and use large language models to analyze reconstruction logs and return structured assessments of the results.
The project was also designed with collaboration and portability in mind. Dunn worked alongside researchers developing similar tomography workflows at ALS, with the goal of creating a reusable approach that could extend beyond a single facility. Rather than building a workflow tailored to one beamline, the team explored how advanced computing, data movement, and AI-driven orchestration could support scientific workflows across multiple DOE user facilities.
“A major highlight was the moment when the reconstruction agent running inside a Polaris PBS job completed a real reconstruction, read its own log, called an LLM, and returned a structured verdict that the reconstruction had passed, with evidence,” Dunn says. “That was a small technical step, but it felt like a big conceptual shift. Instead of a human tailing logs after the fact, the workflow itself was beginning to interpret its own execution state.”
Looking ahead, Dunn hopes to continue advancing these capabilities across the DOE ecosystem. “The broader lesson is that facility portability has to be designed in from the start,” he says. “The goal is to create reusable patterns that can move across SSRL, ALS, APS, NSLS-II, ALCF, NERSC, and other parts of the DOE ecosystem.”
Scaling Agentic AI to Accelerate Scientific Machine Learning
Jiyong Kwon, a PhD student in mechanical engineering at Purdue University, attended the hackathon to explore how agentic AI workflows could be scaled on leadership-class computing systems. As part of the Agentic AI for Scientific Machine Learning team led by Guang Lin, Kwon focused on optimizing ATLAS, a self-evolving agentic framework that uses reinforcement learning and collaborative AI agents to tackle advanced scientific machine learning challenges.
Kwon’s goal was to move ATLAS beyond the team’s existing GPU environment and evaluate how it could leverage ALCF resources, including Polaris and Aurora. During the hackathon, he worked on porting and optimizing the framework, successfully migrating ATLAS execution to Polaris and improving performance by reducing the average training round duration from 163 seconds on an NVIDIA L40S system to 147 seconds on Polaris. He also explored the ALCF Inference Service and vLLM support to improve the integration of large language models within the workflow.
Through the process, Kwon gained firsthand experience with the challenges of bringing AI applications to large-scale computing environments. “Moving an AI workflow to a supercomputing environment is not just about having more compute,” he says. “Environment setup, hardware compatibility, model loading, memory behavior, and inference integration all matter.”
The experience also gave Kwon insight into the future of AI-driven scientific discovery. He plans to continue applying the workflows and techniques explored during the hackathon, including LLM inference services and Intel accelerator technologies, to future ATLAS development. His team is also continuing its collaboration with ALCF through future INCITE efforts and an ALCF Director’s Discretionary allocation focused on scaling agentic AI approaches for scientific machine learning.
Advancing Fluid Dynamics Simulations for Exascale Computing
Samuel Kemmler and Philipp Suffa, PhD students at Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), joined the hackathon to optimize high-performance computational fluid dynamics (CFD)applications for next-generation supercomputing architectures. Both researchers are members of the waLBerla development team, with Suffa also serving as a research associate at FAU and Kemmler affiliated with the German Federal Institute for Materials Research and Testing (BAM) as a research associate.
Their research focuses on CFD using the lattice Boltzmann method (LBM). As developers of the waLBerla framework, Kemmler and Suffa attended the hackathon to test and optimize their newly developed SYCL backend on Aurora, validate its correctness, and investigate its performance and scalability on Intel GPUs.
The team focused on performance engineering of a CFD application in waLBerla, analyzing single-GPU performance, developing performance models, and conducting scaling studies from a single GPU up to thousands of nodes on Aurora. Their work provided valuable insight into how the framework could be optimized for large-scale systems and future Intel-based HPC architectures.
A major milestone during the event was scaling their application to 8,192 nodes on Aurora. “A highlight was scaling our application to 8,192 nodes on Aurora,” Kemmler and Suffa said. “We learned a great deal about Intel GPU architecture, the Intel software ecosystem, and performance optimization techniques for large-scale systems.”
Through the hackathon, the researchers gained hands-on experience with Intel GPU technologies and optimization strategies that will directly support the continued development of waLBerla.
The team plans to apply what they learned about SYCL optimization, Intel GPUs, and large-scale performance analysis to improve the framework’s efficiency on future systems.
Following their experience at ALCF, Kemmler and Suffa hope to continue their collaboration with the facility and have applied for an INCITE allocation to further their research on Aurora.
Talks from the event are available for viewing on the ALCF’s YouTube channel.
Source: Logan Ludwig, ALCF

document fetched 2026-07-30 08:15 · canonical https://hpcwire.com/off-the-wire/argonne-hackathon-helps-researchers-accelerate-scientific-discovery-on-aurora