About the Session

Scientific software has enabled numerical simulations and high-performance computations for decades. Much of today’s scientific software is based on past and enduring community development efforts, due to the complexity of both the science inputs and of developing performant applications for an evolving set of computing technologies. This talk starts by looking back at two such efforts, the Trilinos Project and the Exascale Computing Project, and at the current efforts in the PESO Project. We consider how the fundamental needs for discoverability, usability, and portability as well as performance are addressed through scientific software development practices by developers and communities, and what characterizes the communities for these successful projects.

Looking forward, we will explore how these same elements can be addressed in a future where AI agents become users of scientific software for discovery workflows. Scientific software is already becoming increasingly cloud-native and AI-oriented. This means making applications performant on heterogeneous architectures, improving metadata fidelity to support downstream AI inference, retaining rich time-stepped datasets for training, and more. Agentic engineering means ensuring that scientific applications, libraries, and tools are discoverable, usable, and portable in agentic workflows. We also ask the question: with agents as users and developers, what does community mean for scientific software development as we go forward? 

In this talk, we outline how these future requirements are not brand new but are evolutions of the same fundamental principles we have relied on for years and will shape how we engineer and deliver scientific computing solutions in the future.

Speaker