Parametric reduced order models with machine learning for spatial emulation of mixing and combustion problems

TL;DR

This paper develops parametric reduced order models combined with machine learning techniques to rapidly and accurately emulate spatial distributions in mixing and combustion processes, significantly reducing computational time.

Contribution

It introduces a novel integration of experimental design, data assimilation, POD-based reduction, and ML methods for efficient spatial emulation in combustion problems.

Findings

Kriging-based ROM outperforms other ML methods in accuracy.

ROM achieves up to eight orders of magnitude speedup over traditional simulations.

Deep neural network ROM requires large training data, limiting its applicability.

Abstract

High-fidelity simulations of mixing and combustion processes are generally computationally demanding and time-consuming, hindering their wide application in industrial design and optimization. The present study proposes parametric reduced order models (ROMs) to emulate spatial distributions of physical fields for multi-species mixing and combustion problems in a fast and accurate manner. The model integrates recent advances in experimental design, high-dimensional data assimilation, proper-orthogonal-decomposition (POD)-based model reduction, and machine learning (ML). The ML methods of concern include Gaussian process kriging, second-order polynomial regression, k-nearest neighbors, deep neural network (DNN), and support vector regression. Parametric ROMs with different ML methods are carefully examined through the emulation of mixing and combustion of steam-diluted fuel blend and oxygen issuing from a triple-coaxial nozzle. Two design parameters, fuel blending ratio (hydrogen/methane) and steam dilution ratio, are considered. Numerical simulations are performed and training data is assimilated at sampling points determined by the Latin-hypercube design method. The results show that ROM with kriging presents a superior performance in predicting almost all physical fields, such as temperature, velocity magnitude, and combustion products, at different validation points. The accuracy of ROM with DNN is not encouraging owing to the stringent requirement on the size of training database, which cannot be guaranteed as many engineering problems are specific and associated data availability is limited. For the emulation of spatial field, the parametric ROMs achieve a turnaround time of up to eight orders of magnitude faster than conventional numerical simulation, facilitating an efficient framework for design and optimization.