Center for optimized modern parallel adaptive system software (COMPASS) will devise, utilize, and validate performance models that leverage machine learning solutions to model and predict system &…

The Department of Energy National Nuclear Security Administration awarded a $5.240 million Cooperative Agreement to Massachusetts Institute of Technology on September 29, 2025, under the Predictive Science Academic Alliance Program (CFDA 81.124) to develop and validate machine learning-based performance models for high-performance computing systems and coupled multiphysics simulation.

The award funds two research centers. The Center for Optimized Modern Parallel Adaptive System Software will devise, utilize, and validate performance models that leverage machine learning to predict application and system scalability, capture prediction uncertainty, and inform system optimization decisions. The Center for the Exascale Simulation of Coupled High Enthalpy Fluid-Solid Interactions will develop physics-based and embedded machine learning surrogate models for aero-thermo-chemo-mechanical interaction of high-enthalpy flows and solid materials; advance computer science methodologies including reduced order modeling, uncertainty quantification, machine learning, domain-specific languages, and compiler technologies for heterogeneous architectures; develop inference techniques and uncertainty quantification methodologies supported by validation experiments; and integrate research and development efforts into a validated exascale simulation capability to advance predictive capability of thermal protection system operational and failure response.

Performance is conducted at MIT's Cambridge, Massachusetts location. The period of performance runs through September 28, 2030.

Center for optimized modern parallel adaptive system software (COMPASS) will devise, utilize, and validate performance models that leverage machine learning solutions to model and predict system & application performance. The primary research question addressed is as follows: how can HPC performance models be structured and trained using machine learning techniques to predict application and system scalability and performance, capture prediction uncertainty, and effectively inform system optimization decisions.