• Citer

Evaluating the performance of dimensionality-reduction and rotation techniques in a latent variable framework

par Muñoz Salamanca, Eva ;Malik, Mohammad Rafi;Cuoci, Alberto ;Im, Hong G.;Parente, Alessandro
Référence Combustion theory and modelling, page (1-17)
Publication Publié, 2026-03

Article révisé par les pairs

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

This work provides a systematic evaluation of how different low-dimensional representations of the chemical state space affect the accuracy and computational performance of a latent variable (LV) framework for reactive flows. These latent representations are constructed using Principal Component Analysis (PCA) and Partial Least Squares (PLS) combined with orthogonal and oblique rotation strategies, trained using zero-dimensional reactors with the detailed AramcoMech 1.3 mechanism for CH4/air combustion, and tested on unseen conditions. Performance is assessed in terms of solution accuracy, computational speed-up and metrics describing the latent space and system dynamics. Results show that while all latent-basis configurations achieve computational savings, their numerical behaviour strongly depends on the structure of the latent space. PCA-derived bases provide the lowest reconstruction errors but only a moderate speed-up. Unrotated PLS bases exhibit strong sensitivity to the latent dimension, whereas orthogonal rotations restore stability across compression levels. Among all techniques, the PLS–Quartimax combination provides the best overall performance, yielding smooth source-term manifolds and latent variables explicitly aligned with the fast chemical time scales due to the PLS formulation. This alignment enables larger stable time steps and results in the highest computational speed-up. These findings demonstrate that appropriate combinations of dimensionality-reduction techniques and rotation strategies significantly improve the numerical behaviour of the LV solver, enabling faster chemistry integration while preserving the full thermochemical state.