Dimensionality reduction and unsupervised classification for high-fidelity reacting flow simulations

The development of reduced-order combustion models able to accurately reproduce the physics reactive systems has been a highly challenging aspect numerical research in recent past. complexity problem can be reduced by identifying and using low-dimensional manifolds account for turbulence-chemistry interactions. Recently, Principal Components Analysis (PCA) shown its potential reducing dimensionality chemically system while minimizing reconstruction error. present work demonstrates application Manifold Generated Local PCA (MG-L-PCA) approach direct simulation (DNS) turbulent flames. is enhanced with an unsupervised clustering based on Vector Quantization (VQPCA) on-the-flyPCA-based classification technique. model then applied three-dimensional (3D) premixed NH3/air flame transporting only subset original state-space variables computational grid basis reconstruct non-transported variables. Results are compared both detailed reaction mechanism Computational Singular Perturbation (CSP) skeletal mechanism. A comparison between training one-dimensional (1D) 3D data sets also included. Overall, MG-L-PCA allows not reduction number transport equations, but significant stiffness system, providing accurate results.