Driving Design with Cognitive Simulation

Project Overview

Many contemporary science and engineering problems rely on the use of expensive simulations to design new technologies. However, since these simulations are both expensive and have several design parameters, using them for optimal design (which automatically seeks out and finds the best solution) is a challenge. Through the invention of novel machine learning and optimization algorithms and software, this project developed the capability to automatically search over and optimize dozens of design parameters simultaneously. We developed key general technologies to enable optimal design on exascale-class high performance computing systems and demonstrated their particular use for inertial confinement fusion (ICF) design for the National Ignition Facility (NIF). In this case, we successfully optimized 17 ICF design parameters simultaneously in only a few hundred simulations. This is a significant improvement over the prior state-of-the-art, which requires tens of thousands of simulations and is limited to fewer than ten parameters. Additionally, the project developed techniques for multifidelity optimization and optimization under uncertainty, which become important for real-world applications, such as design for manufacturability and robustness. Many of the technologies developed by this project have been open source released and successfully transferred to several DOE and non-DOE mission applications, including fusion energy, exascale computing, stockpile stewardship and bio and space security.

Mission Impact

Several core missions of DOE/NNSA rely on the use of expensive computer codes, whether to design a new experimental configuration, or assess the uncertainty and confidence in an existing design. Many of these problems can be constructed as optimization problems in many parameters (for instance, minimizing the difference between an experiment and simulation, or maximizing the likelihood of a certain experimental result). This project has built novel technology to significantly accelerate these missions: using machine learning and high-performance computing, combined with advanced optimization algorithms, we are now able to automate many of these difficult problems, helping us explore a larger number of scenarios and design possibilities than previously possible. This new technology significantly improves NNSA's responsiveness to new national security challenges, helps us more quickly find and develop new science and technology solutions, and empowers an agile and flexible workforce that makes efficient use of resources. Several DOE and non-DOE projects have begun to use these technologies, including the stockpile stewardship, hypersonics, fusion energy, bio-resilience and space science and security programs.

Publications, Presentations, and Patents

R. Anirudh, J.J. Thiagarajan, "Out of Distribution Detection with Neural Network Anchoring" (Presentation, IEEE International Conference on Computer Vision and Pattern Recognition, New Orleans, LA, 2022).

A. Gillette, "Barycentric Coordinates in General Dimensions for Scientific Machine Learning" (Presentation, CSF Workshop on Generalized Barycentric Coordinates in Computer Graphics and Computational Mechanics, Ascona, Switzerland, June 2022).

—. 2021. "Delaunay interpolation diagnostics for model assessment," (Presentation, Center for Math and AI Colloquium, Washington, DC, virtual, 2021).

Liu, Shusen. 2023. "Topological Data Analysis Guided Domain Partition for Bayesian Optimization." OSTI Technical Report.

Y. Mubarka, N.-Y. Chiang, A. Gillette, I. Kim, E. Ku,r J.M., Koning, R. Lee, J. L. Peterson, R. Tran, J. Wang, "Resilient Workflow Management for Interfacing Among Black Box Algorithms" (Presentation, SIAM Conference on Computational Science and Engineering, Amsterdam, The Netherlands, Feb-Mar 2023).

J.L, Peterson, "Towards Digital Design at the Exascale" (Presentation, 4th International Conference on Data Driven Plasma Science, Okinawa, Japan, April 2023). 

Peterson, J. L., Bay, B., Koning, J., Robinson, P., Semler, J., Anirudh, R., Athey, K., Bremer, P. T., Castillo, V., Di Natale, F., Fox, D., Gaffney, J. A., Hysom, D., Jacobs, S. A., Kailkhura, B., Kustowski, B., Langer, S., Spears, B. K., Thiagarajan, J.J.  2022. "Enabling Machine Learning-Ready HPC Ensembles with Merlin." Future Generation Computer Systems, 131.

J. L. Peterson, B. Kustowski, L. Masse, J. Koning, B. Bay,  J. Gaffney, K. Humbird, M. Kruse, R. Nora, B. Spears, "Engineering Robustness into Inertial Confinement Fusion Designs" (Presentation, 62nd Annual Meeting of the APS Division of Plasma Physics, Nov 2020). 

J. L. Peterson, J.J. Thiagarajan, R. Anirudh, Y. Mubarka, I. Kim, P.T. Bremer,  B. K. Spears, V. Narayanaswamy, "Towards Digital Design at the Exascale: Advances in Bayesian Optimization with Neural Networks" (Presentation, 64th Annual Meeting of the APS Division of Plasma Physics, Spokane, WA, Oct. 2022).

J. J. Thiagarajan, Anirudh, R., Narayanaswamy, V. , Bremer, P.T. 2022. "Single Model Uncertainty Estimation via Stochastic Data Centering" (Presentation, Neural Information and Processing Systems, New Orleans, LA, Nov. 2022). 

J. J. Thiagarajan, Anirudh, R., Narayanaswamy, V., Mubarka, Y., Kim, I, Bremer, P. T., Peterson, J. L., Spears, B., "Data-Efficient Scientific Design Optimization with Neural Network Surrogates" (Presentation, Efficient Scientific Design Optimization with Neural Network Surrogates ReALML Workshop at ICML 2022, Baltimore, MD, July 2022).