Accelerating Design of ML and AI Experiments in Scientific Simulation

Artificial intelligence (AI) and machine learning (ML) have demonstrated incredible success across industry and scientific fields. This impact is felt across scientific simulation domains, where AI and ML techniques are being used to explore complex patterns, improve accuracy of physics solvers, and accelerate time to insight. However, “black box” integration of AI and ML tools into simulations has yet to demonstrate significant impact, leading practitioners to develop their own implementations and integrations to satisfy their workflow needs. In this proposal, we seek to reduce the barriers to adoption of AI and ML tools for mature scientific simulation codes. In Phase I, we will focus on demonstrating an in situ ML toolbox for coupling mature simulation codes to AI and ML tools. We will leverage our team’s expertise, applying our toolbox to two approaches to computational fluid dynamics relevant to NASA missions, large eddy simulation and Reynolds averaged Navier Stokes, which each have unique requirements and methodologies for incorporating AI and ML for improved model accuracy. By exploring these two regimes of CFD, we seek to highlight the flexibility of our approach to empowering mature simulation codes with cutting edge AI and ML tools, aiming toward other physics domains of interest to NASA in future work. Phase I funding will support this research and development effort, laying the groundwork for improved access to new AI and ML tools, and improved infrastructure for simulation users to experiment with these cutting edge tools.

NASA leverages numerical simulation for many of its missions and has been at the forefront of computational science. Examples of such simulations include the CFD solvers Overflow, FUN3D, and LAVA. All of these tools support NASA’s Aeronautics Research Science and Space Technology, and Space Exploration missions. NASA has been advancing its simulation capabilities to leverage the new exascale hardware, especially GPU accelerators. Already, FUN3D has been demonstrated utilizing over 33k GPUs on the Frontier supercomputer at OLCF to simulate retropropulsion of a Martian lander at a scale and fidelity not possible on previous generations of supercomputers. The next step in this path is the use of new and emerging AI capabilities, especially deep learning, to further accelerate scientific computation. Unlike GPU acceleration, deep learning has not been developed with scientific computing in mind and as such, it is very challenging to exploit it to accelerate computations. Our proposal focuses on making the process of integrating simulation codes with AI infrastructure (especially deep learning) significantly easier, liberating the NASA researchers from the nitty gritty details and enabling them to focus on finding innovative ways of leveraging deep learning for advanced numerical modeling. We also propose research and development in the area of AI-assisted turbulence modeling bringing it closer to production use and available to NASA researchers for further exploration of this very important area. We also propose to apply this technology to other NASA relevant modeling codes such as the GEOS Earth System Model and VPIC which is used for space weather modeling. Over the last decade or so, AI has been at the forefront of advances in modeling through significant advancements in deep learning. Especially in the areas of natural language processing, computer vision, and data science, breakthrough advancement has been at a neck breaking speed. These advancements are also impacting computational sciences focusing on modeling physical phenomena. All major simulation vendors (Ansys (Synopsis), Altair, Hexagon etc.) have announced AI-assisted simulation efforts. There have also been new startups solely focused on this domain, such as Navier AI. These products are all closed and although they benefit thousands of users, they are not usable by small- and medium-sized simulation developers as well as research organizations in government and academia to advance more domain specific codes. The work proposed here will make it easier for the entire community to integrate with deep learning frameworks. As a consulting company, Kitware will significantly benefit from this growth by offering custom development services and support. In addition, through the development of the AI integration toolbox and advancement in the area of turbulence modeling, we expect to further grow our business in the CFD market. The CFD market has the potential to grow by $606.76 million during 2021-2025, and the market’s growth momentum will accelerate at a CAGR of 5.33%. The aerospace and defense sector, specifically, is expected to have a higher-than-average growth rate.

This is early/mid-stage (TRL 6) — the most realistic path in is NASA SBIR/STTR, which funds small businesses and research institutions to develop technology aligned with NASA's needs (equity-free, phased funding). Check whether a current SBIR/STTR solicitation topic overlaps with this project's technology area, or contact the project directly (above) to ask.

None of these are guaranteed paths for this specific project — TechPort itself doesn't have an "apply" button. Reaching out to the contact(s) above with a specific question is usually the fastest way to find out what's actually open.