• Citer

Sparse surrogate modeling for combustion chemistry via time-lag autoencoders and gradient-based clustering

par Castellanos, Luisa;Freitas, Rodolfo R.S.M.;Galassi, Riccardo Malpica;Jiang, Xi;Parente, Alessandro
Référence Chemical engineering science, 324, 123323
Publication Publié, 2026-04-01

Article révisé par les pairs

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

This work presents a data-driven framework for efficient and interpretable surrogate modeling of combustion chemistry, combining time-lag autoencoders (TAEs) with gradient-based clustering. The methodology targets applications where high-fidelity simulations are computationally prohibitive, enabling accurate prediction of thermochemical states with minimal input variables. Ignition trajectories are first encoded into low-dimensional, temporally informed latent spaces, from which physically meaningful chemical carriers are identified via statistical correlation. Cases are then grouped by progress-variable gradient similarity, and cluster-specific regression models reconstruct the full thermochemical state from the carriers and equivalence ratio. The proposed framework is demonstrated on methane/air combustion with GRI-Mech 2.11 over a broad range of equivalence ratios. The approach achieves a coefficient of determination greater than 0.9 in most cases and maintains robustness in interpolation and sparse-data scenarios. By integrating dynamical feature learning with gradient-informed clustering, the approach produces generalizable, interpretable, and computationally efficient surrogates for complex reaction systems. This facilitates their incorporation into computational fluid dynamics (CFD) solvers to enable accelerated reactor simulations and optimize chemical engineering designs.