(108c) Smart Proxy Modeling for Cfd; Application of Artificial Intelligence & Machine Learning to Numerical Simulation | AIChE

(108c) Smart Proxy Modeling for Cfd; Application of Artificial Intelligence & Machine Learning to Numerical Simulation

Authors 

Mohaghegh, S. D. - Presenter, West Virginia University
Shahnam, M., National Energy Technology Laboratory
Aboaba, A., West Virginia University
Martinez, Y., West Virginia University
Guenther, C., National Energy Technology Laboratory
Liu, Y., National Energy Technology Laboratory
Computational Fluid Dynamics (CFD) is a computationally expensive numerical simulation modeling technology. In the past few decades traditional proxy models (statistical response surfaces, or reduced order models) have been developed in order to expand the utilization and practicality of CFD models for applications such as optimization and uncertainty quantification.

Smart Proxy is a new technology that is defined as the engineering application of Artificial Intelligence and Machine Learning. Smart proxy modeling replicates the results of complex CFD models with high accuracy without reducing the physics of the models or its resolution in space and time. Unlike response surface approach, Smart Proxy modeling replicates the physics of the CFD model simultaneously for all the cells and all the required output variables. Once developed, Smart Proxy modeling accomplishes this objective at an incredibly fast pace when it is deployed on a laptop or a desktop. Development of this technology requires only a handful of CFD simulation runs.

The CFD model that is the subject of this particular study, simulates the combustion of natural gas within more than 9.3 million cells in three segments: Inlet-Combustion-Exhaust as shown in the figure below. A single run of this CDF model using a commercial CFD software application (FLUENT) takes 24 hours on an HPC that includes 40 cores. The output of interest in this model was the detail and comprehensive distribution of Pressure and Temperature, as well as Nitrogen, Oxygen, and CO2 volume fractions throughout the 9.3 million cells. The Smart Proxy model for this application was developed using only 8 CFD simulation runs.

Without compromising the complex physics that has been modeled or the original space-time resolution, the data-driven Smart Proxy replicates the CFD with high accuracy and at very high speed. Smart Proxy takes advantage of pattern recognition capabilities of Artificial Intelligence and Machine Learning to achieve its objectives.