Implementing Design Optimization with CREATE Physics Codes Using SAGE+PAGODA
Evolutionary algorithms (EAs) can investigate large parameter spaces efficiently during design optimization. However, EAs typically require on the order of millions of function evaluations to converge; this is infeasible when the objective being optimized is computed using high-fidelity physics-based codes. The Create Applied Surrogates Institute (CASI) has developed a tool (SAGE) that can produce efficient AI/ML surrogate models using data from high fidelity physics-based simulations. Inferencing surrogate models is several orders of magnitude faster than performing the full simulations, and can often execute in seconds on a single GPU. By driving surrogate models with EAs, we can bridge the current gap between high-fidelity simulations and design optimization. A proof-of-concept parameter evolution was performed with a surrogate model developed by CASI using SAGE to optimize flapper deflections for a simple open-source missile problem. This was made possible by a new backend for the EA code (PAGODA) that can run arbitrary Python workflows using a model server. PAGODA employs a highly refined self-adaptive EA with a wide variety of options for mutation and recombination, and can optimize real, integer or combinatorial variables.
IMPACT
Accomplishment: We have integrated an example of AI/ML surrogate model inference (SAGE) into a massively parallel Evolutionary Algorithm (PAGODA) enabling design optimization in otherwise computationally infeasible parameter spaces. Additionally, the Python-based model-server approach is completely general, and can be used to rapidly employ additional AI/ML models or traditional black-box codes into parameter evolutions.  Result – Mission Impact: The high-precision physics-based codes developed by CREATE have been widely employed for analysis but can now be applied to design optimization tasks by first generating a surrogate model with SAGE and then running an evolution with PAGODA. This will allow for design optimization to be performed across all branches of CREATE using a common code base in support of diverse Digital Engineering objectives.
PRESENTER
Lill, James
James.Lill@GDIT.com
937-255-1461
HPCMP/PET
CO-AUTHOR(S)
Anderson, Calvin
Calvin.Anderson@GDIT.com
CATEGORY
Surrogate Modeling for HPC
SECONDARY CATEGORY
GPU usage for HPC
SYSTEM(S) USED
Raider, Nautilus