Large-Scale Computing and Computational AI

High performance computing (HPC) remains a critical part of NASA Science Mission Directorate's (SMD) infrastructure needed to address fundamental questions of science and seek answers in order to understand the Sun, Earth, Solar System, and beyond. HPC serves a specific need in numerical simulation as high fidelity and efficient computing capabilities are needed to tackle large-scale research and studies to investigate complex dynamic processes up to the scale of entire science ecosystems. One of the largest challenges facing the HPC community today is the tremendous amount of refactoring that is typically required of existing large scale applications in order to address the hardware paradigm shift that has taken place over the past 5 to 10 years to usher in the exascale era, which is now upon us and this shift is expected to continue and become even more heterogeneous in the coming years. There is an urgent need for application refactoring and performance portability in this environment. A second challenge is the emergence of the field of quantum computing and assessment of its potential to drive breakthroughs needed to analyze and solve large scale science problems currently beyond the reach of classical computing methods. The National Aeronautics and Science Administration (NASA) scientists and engineers are increasingly turning to large scale numerical simulation on supercomputers to advance understanding of complex systems and to conduct high fidelity science and engineering analyses. To address these challenges, novel software technologies are sought such as artificial intelligence (AI)/machine learning (ML) that will increase the mission impact of NASA's investments in supercomputing systems and associated operations and services. Large Language Models (LLMs) are increasingly capable for use in HPC for tasks like code generation and analysis and have potential for further use with some data privacy and security concerns addressed for secure adoption. Small Language Models (SLMs) are emerging to address applying mechanistic interpretability to smaller tasks like OpenMP (OMP) code completion, data race detection, and OMP code generation. SLMs are emerging for HPC physics by offering efficient, resource-light AI for tasks like code generation, data analysis, and model acceleration, fine-tuned on domain-specific data for physics code principles, enabling faster, localized AI within complex scientific workflows, using like quantization techniques to run on less hardware, making them ideal for specific physics problems. Proposals should demonstrate a relation to fields of study relevant to SMD or demonstrate the portability to SMD disciplines. Innovative computation technology directly targeting a stated NASA need outside of an SMD application will still be considered, but offerors are strongly encouraged to identify at least one relevant NASA subject matter expert and relevant project. Proposals should also demonstrate awareness of the state of the art of their proposed technology and should leverage existing commercial capabilities and research efforts where appropriate, including open source software and open standards.