Surrogate Models to Accelerate High-Fidelity Physics Based Simulation

OUSD (R&E) MODERNIZATION PRIORITY: Artificial Intelligence/Machine Learning TECHNOLOGY AREA(S): Information Systems OBJECTIVE: Creation of a simulation emulator that accelerates execution time by orders of magnitude and includes uncertainty quantification without requiring large amounts of training data. DESCRIPTION: End-to-end simulations have significantly improved the ability to massively parallelize simulations in cloud based HPC environments. However, cloud based environments can be expensive and can be infeasible to acquire the number of resources needed in an acceptable time span. Creating simulation emulators is a popular approach to speeding up simulation execution; however, traditional methods, such as data driven models or traditional machine learning models, have several limitations including requiring large amounts of data, limited accuracy, narrow applicability, and uncharacterized uncertainty. Recent advances in physics-informed machine learning (PIML) and reduced order modeling have demonstrated incredible promise at accelerating traditional simulators (sometimes up to 1,000x faster) while sacrificing minimal amounts of fidelity. Replacing these bottlenecks enables the ability to scale these simulations to collect sufficient data in resource-constrained environments. This topic seeks to further develop and utilize these advances in surrogate modeling to accelerate a specific bottleneck in simulation execution (such as debris or EO-IR scene generation) using a limited amount of data. This acceleration of simulation will remove barriers to real-time and faster-than-real-time execution and enable rapid testing and prototyping. The surrogate model will characterize its uncertainty and explicitly define its bounds of applicability. Beyond demonstration of a specific use case, the approach will be generalizable to other simulation components. PHASE I: Develop proof-of-concept algorithms, tools, software and analyses that demonstrate a potential for achieving the topic objectives: - Identify candidate physics models at unclassified level - Develop evaluation metrics for comparison to existing model baseline - Incorporate physics based simulation execution time enhancements - Provide increased knowledge and confidence in surrogate models derived from high fidelity physics based data and simulations - Demonstrate understanding of uncertainty and bounds of applicability for surrogate models PHASE II: Develop a full prototype capability demonstrating initial capabilities per topic objectives (per Phase I) with the intent of testing the capability for experimentation in government Modeling and Simulation labs using government provided physics models. This should include prototype level user and design documentation. Development should facilitate cyber security approval for loading the prototype software on government computer systems through cyber aware design decisions and development of cyber security artifacts. PHASE III DUAL USE APPLICATIONS: Develop operational capability for use in government simulations, including user and design documentation. Maintain and improve capabilities based on Phase I and Phase II use experience. Continue to support cyber assurance. REFERENCES: 1) Kasim, M., Watson-Parris, D., Deaconu, L., Oliver, S., Hatfield, P., Froula, D. & Vinko, S. (2020). Up to two billion times acceleration of scientific simulations with deep neural architecture search. In APS Division of Plasma Physics Meeting Abstracts (Vol. 2020, pp. BO05-001) 2) Kadupitiya, J. C. S., Sun, F., Fox, G., & Jadhao, V. (2020). Machine learning surrogates for molecular dynamics simulations of soft materials. Journal of Computational Science, 42, 101107 3) Moseley, B., Markham, A., & Nissen-Meyer, T. (2021). Finite Basis Physics-Informed Neural Networks (FBPINNs): a scalable domain decomposition approach for solving differential equations. arXiv preprint arXiv:2107.07871 KEYWORDS: Surrogate Models; Hybrid Models; Data Driven Models; Physics Based Models; physics-informed machine learning; machine learning; reduced order modeling