Error bounds and stability of neural reduced-order models for combustion dynamics
Reduced-order modelling offers an effective means of approximating high-dimensional nonlinear dynamical systems at reduced computational cost. However, data-driven approaches often lack rigorous guarantees on approximation accuracy and long-time stability. This work develops a unified theoretical and computational framework for neural reduced-order models (Neural-ROMs) applied to nonlinear reaction–diffusion dynamics. Combining reduced-space projection, learned nonlinear closure, and nonlinear stability analysis, we derive explicit error bounds separating projection and learning errors. We also establish a sufficient stability condition linking the effective Lipschitz constant of the learned closure to the dissipativity of the reduced linear operator. The theoretical framework is assessed using a bistable reaction–diffusion benchmark exhibiting flame-like front propagation. Numerical experiments demonstrate improved approximation accuracy over classical and data-driven reduced-order baselines, including under increasing reaction intensity. Furthermore, controlled experiments confirm long-time decay of the reduced-state norm when the theoretical stability condition is satisfied. These results provide a mathematically controlled foundation for neural reduced-order modeling of reactive dynamics, with potential relevance to combustion-oriented applications.
Authors
- Kolade Matthew Owolabi (ORCID: https://orcid.org/0000-0001-9290-3458)
Institutions
- Federal University of Technology (NG)
- Sefako Makgatho Health Sciences University (ZA)
- Fenerbahçe University (TR)
Publication Details
- Journal
- Combustion Theory and Modelling
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1080/13647830.2026.2765511
- Primary Topic
- Model Reduction and Neural Networks
- Type
- article
- Field-Weighted Citation Impact
- 0.00