EAGER: NAIRR Pilot: Enabling parametric sweeps on exascale AI-integrated simulations through federated learning

This NAIRR-Pilot project introduces an innovative artificial intelligence (AI) integrated framework to dramatically enhance the efficiency and accessibility of exascale multiphysics simulations, addressing the critical challenge of enabling comprehensive parameter space exploration at unprecedented scales. While exascale simulations represent the pinnacle of supercomputing power, their significant computational demands have hindered comprehensive parameter testing, crucial for scientific discovery and engineering optimization. The approach integrates advanced AI techniques with traditional numerical methods, democratizing access to exascale-level insights without sacrificing accuracy. Key innovations include a novel spatial coupling mechanism between AI and numerical solvers, efficient communication techniques for distributed computing, and adaptive learning methods that quickly adjust to emerging behaviors at the exascale level. The coupling of pre-trained AI models with numerical solvers allows for rapid solution generation for large portions of the domain. The numerical solver is selectively deployed in critical sections where complex physical processes occur, ensuring high accuracy while significantly reducing computational costs. This NAIRR-Pilot project democratizes access to advanced AI research capabilities in computational science, enabling efficient parametric sweeps of exascale simulations. Serving as a crucial testbed for integrating AI resources with exascale scientific computing applications, it contributes to NAIRR's mission of broadening access to cutting-edge AI research tools to solve global challenges like climate change and future manufacturing. The project aligns with national priorities in maintaining leadership in high-performance computing and AI via educational initiatives to cultivate the next generation of diverse STEM talent through coding clubs for children and teens, inspiring future scientists, and fostering community engagement with open-source computational tools. The project develops a hybrid spatial coupling framework by integrating graph neural network-based neural operators with traditional numerical solvers. This integration maintains the accuracy of simulations while greatly improving computational speed, enabling efficient parametric sweeps of exascale multiphysics simulations. The approach includes probabilistic sampling-based message passing to optimize communication in distributed machine learning and hierarchical federated learning to enhance reduced-order model (ROM) predictions through efficient in situ learning and model adaptation in exascale environments. Additionally, one-shot learning techniques enable ROMs to adapt to new dynamics quickly using limited high-fidelity data. The project demonstrates this transformative approach by benchmarking against carbon capture processes and additive manufacturing problems as a proof of concept. The methodology significantly reduces computational expenses while maintaining high accuracy, potentially enabling comprehensive parametric studies of complex multiphysics problems at the exascale level. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.