Bayesian neural networks for Arrhenius-based turbulent combustion closures in large eddy simulations
par Piu, Lorenzo
;Pequin, Arthur
;Iavarone, Salvatore
;Freitas, Rodolfo R.S.M.;Jung, Ki Sung;Chu, Hongchao;Pitsch, Heinz
Référence 12th European Combustion Meeting (ECM) (7th - 10th April 2025: Edinburgh)
Publication Non publié, s.d.
;Pequin, Arthur
;Iavarone, Salvatore
;Freitas, Rodolfo R.S.M.;Jung, Ki Sung;Chu, Hongchao;Pitsch, HeinzRéférence 12th European Combustion Meeting (ECM) (7th - 10th April 2025: Edinburgh)
Publication Non publié, s.d.
Communication à un colloque
| Résumé : | Accurate turbulent combustion simulations are essential to design modern combustion systems, enhancing efficiency, and reducing emissions.Direct Numerical Simulations (DNS) of industrial systems are currently out of reach in terms of computational costs. As a more feasible alternative, Large Eddy Simulations (LES) provide a high-fidelity approach that resolves the largest turbulent structures and filters out the smallest ones, relying on sub-filter models to represent the effects of these unresolved scales. The recent advances in deep learning offer a promising solution for the sub-filter modeling of turbulent flows in LES, leveraging the information available from DNS simulations—given that the latter is representative of the physics to be modeled. Deep learning techniques suffer from a lack of interpretability and produce overconfident predictions when faced with out-of-sample data. This prediction uncertainty poses challenges in applications where reliability is paramount, emphasizing the need for effective uncertainty quantification methods. In this study, we introduce a synthetic one-dimensional dataset that mirrors the training conditions of DNS of turbulent flows and provides an enhanced visualization. We compare three approximate uncertainty quantification techniques, i.e., Bayes by backprop, Monte Carlo dropout, and deep ensembles, to evaluate the quality of the uncertainty estimated in each approach and better understand which model is more suited to the prediction of epistemic and aleatoric uncertainty. In addition, a Bayesian neural network is trained on high-fidelity DNS data of a slot hydrogen-air lifted non-premixed flame—as hydrogen emerges as a carbon-free fuel that produces no CO2 emissions—to improve the laminar finite rate approach in LES. The network is then tested on filter sizes outside of the training data; the results show a progressively increasing uncertainty of the model when tested on out-of-sample filter sizes, making Bayesian neural networks a promising tool for identifying uncertainty in deep learning-based sub-filter LES models. |



