Surrogate Models to Accelerate High-Fidelity Physics Based Simulation

The MDA has identified a need for decreasing the time required for simulations and rendering by orders of magnitude. To address this need, Stellar Science and the University of New Mexico (UNM) propose the development and use of physics-informed neural networks (PINNs), which have been shown to reduce execution times dramatically while still maintaining small errors compared to traditional physics simulations. PINNs can be used to solve a wide variety of problems, and the tool chain resulting from this effort will not be limited to a specific problem type. At the same time, the tool chain needs to be applied to solve a specific problem to demonstrate that it can solve realistic problems of interest. We suggest using the radiative transfer equation, which is complex enough to demonstrate the effectiveness of the approach, yet small enough to allow a solution without resorting to high performance computers (HPCs). The problem is also of interest for realistic scene generation in the electro-optical / infrared (EO/IR) region. We provide existing physics-based tools that can help solve the problem. The proposed framework makes use of existing software with full government rights to speed its development. This includes the Varied Interface & Phenomenology Engineering Relationship Suite (VIPERS), a surrogate modeling tool set that uses a common interface layer to simplify training and make use of a surrogate completely transparent to the underlying surrogate type; Galaxy, which allows scalable surrogate model selection from laptops to multiple HPCs; and Dakota, which provides dozens of standard optimizers that can be used to quickly tune surrogate model parameters during model selection. All the tools in the tool chain are either open source or have full government use rights. Our Phase I effort will add a PINN to this tool chain and validate it against published results using public data sets. It then applies the software to the radiative transfer equation to demonstrate the technique on a reasonably complex problem of interest. This includes the development of metrics to evaluate the model and perform uncertainty qualification on the solution. Time permitting, additional unclassified problems of interest to MDA will be examined, and additional techniques like physics-informed Gaussian processes (PIGP) will be tried. Work in Phase II and Phase III will solve specific problems of interest to MDA, some of which may be classified. It will also add a user interface and documentation suitable for use by analysts from a variety of backgrounds. The basic tool can also be integrated into an existing simulation or rendering software as directed by MDA. Approved for Public Release | 22-MDA-11339 (13 Dec 22)