AI4HPC: An Iterative Framework for AI-Assisted Scientific Software Development and Optimization

Computer simulations allow scientists to explore systems that are difficult or impossible to study through experiments alone, from designing new materials to understanding complex energy systems. Yet turning a new scientific idea into a working simulation can require substantial changes to software built over decades. The effort needed to implement new models, explore alternative algorithms, or take advantage of powerful new computers can limit both the pace of discovery and the questions researchers can pursue.

AI4HPC will develop a framework that uses artificial intelligence to help scientists overcome these barriers. It will combine the growing capabilities of AI coding agents with the execution, testing, and scientific validation needed to determine whether software changes produce trustworthy results and improve performance. Researchers will remain central to evaluating and guiding those changes, with a clear record of how they were developed and assessed.

Working with production scientific software across multiple research areas, the project will develop and evaluate methods for adapting and extending complex simulation codes. Scientific ideas often advance faster than the software needed to explore them. AI4HPC aims to change that relationship, developing the methods needed to turn emerging AI capabilities into reliable tools for transforming complex scientific software. If successful, this approach could give researchers substantially greater freedom to pursue new models, algorithms, and computational experiments—making software a more responsive partner in scientific discovery.