MS5H – Co-Design of Workflow Orchestration Systems for AI/ML-Enhanced Scientific Workloads on HPC Infrastructure
Session Chair
Event Type
Minisymposium
Engineering
Physics
Computational Methods and Applied Mathematics
TimeWednesday, July 114:00 – 16:00 CEST
LocationBldg. 8 – B 102
DescriptionOrganizer(s): Ozgur Ozan Kilic (Brookhaven National Laboratory), and Charles Leggett (Lawrence Berkeley National Laboratory)
Scientific workflows increasingly combine traditional simulations with AI/ML methods, creating complex pipelines that must operate across heterogeneous, distributed infrastructure. This minisymposium examines how co-design approaches—bringing together domain scientists, HPC architects, and AI researchers—can address the resulting orchestration challenges. We explore three interconnected themes. First, we examine intelligent workflow generation, including how agentic AI and large language models can assist scientists in developing and optimizing workflows for heterogeneous architectures. Second, we address federated infrastructure services, discussing architectural patterns for unified APIs that enable workflows to seamlessly traverse institutional boundaries while respecting security and data governance constraints. Third, we consider AI-ready data and model services, exploring how federated catalogs, automated curation, and model registries can accelerate discovery. While drawing examples from High Energy Physics—where petabyte-scale data and international collaborations have driven workflow innovation—the patterns we discuss apply broadly to climate science, materials research, genomics, and other data-intensive domains. Our concluding panel examines future directions for co-designed scientific computing systems.
Scientific workflows increasingly combine traditional simulations with AI/ML methods, creating complex pipelines that must operate across heterogeneous, distributed infrastructure. This minisymposium examines how co-design approaches—bringing together domain scientists, HPC architects, and AI researchers—can address the resulting orchestration challenges. We explore three interconnected themes. First, we examine intelligent workflow generation, including how agentic AI and large language models can assist scientists in developing and optimizing workflows for heterogeneous architectures. Second, we address federated infrastructure services, discussing architectural patterns for unified APIs that enable workflows to seamlessly traverse institutional boundaries while respecting security and data governance constraints. Third, we consider AI-ready data and model services, exploring how federated catalogs, automated curation, and model registries can accelerate discovery. While drawing examples from High Energy Physics—where petabyte-scale data and international collaborations have driven workflow innovation—the patterns we discuss apply broadly to climate science, materials research, genomics, and other data-intensive domains. Our concluding panel examines future directions for co-designed scientific computing systems.
Presentations
