A Generic Data-Driven Framework via Physics-Informed Deep Learning. App. F-21b-Yuan

Program: Computational and Information Science and Technology Office (CISTO) - App. F-21 Research Title: Computational and Technological Advances for Scientific Discovery Research Topic: High Performance Computing; Evolving Applications to Exascale High-fidelity physics-based computational models become increasingly important to bring predictive, optimized, and focused capabilities to NASA’s strategic plan. A complex system, e.g. Earth/Planetary Atmosphere, can consume dramatic computational resources on the current High- Performance Computational facility. Replacing physics-based model components with artificial intelligence evolves as a viable path to enhance performance for HPC applications. In this project, the team at the University of South Carolina proposes to explore and establish a physics-informed deep learning driven computational framework, relying on data from both physics-based numerical models and experimental data, to generate a scalable data-driven model constrained by physical limits to replace computationally intensive high-fidelity models. This work will start to answer whether we can replicate complex interactions in data-driven model and conceptualized them with extensive domain knowledge to reestablish a physically meaningful model that can recognize or react to new circumstances they have not been trained for. With well-established modeling capabilities in Additive Manufacturing (AM) and availability of experimental data, we propose to start such framework from studying the melt pool dynamics during laser-powder melting process. This system also largely shares the governing equations as in General Circulation Models of a planetary atmosphere, where Navier–Stokes equations are solved with thermodynamic terms for various energy sources. In this project, physics-based highfidelity model for melt pool dynamics that solves fluid dynamics, heat transfer, solidification and vaporization will be performed to generate dataset for the learning process of machine learning (ML) algorithms. Informed by ML, multistage virtual and physical experiments will be carried out to reduce the uncertainly of the predictive model and validate the physics-based model. Physicsinformed deep learning, in turn, will be developed by approximating the unknown solution with a deep neural network. In the meantime, transfer learning will be leveraged to accelerate the datadriven model establishment from both experimental and computational data, and structural causal models will be developed to incorporate physics recovery. The proposed framework will generate a new paradigm that creates and applies both novel physicsbased and data-driven techniques to dramatically enhance traditional computational, theoretical tools for scientific discovery and application. Interfaces between computational framework, theoretical models with experimental observations through advanced machine learning will be expected to extend to complex problems to generate physically sound mathematical model to evolve applications to exascale.

This is a mature technology (TRL 7+) — the realistic path in is usually NASA's Technology Transfer Program: licensing an existing NASA patent, or a Space Act Agreement to use NASA facilities/expertise directly. NASA also runs a startup licensing program with no upfront fee for companies formed to commercialize a specific NASA technology.

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.