Extrapolation Performance of Convolutional Neural Network-Based Combustion Models for Large-Eddy Simulation:...

Area of Science:

• * Computational fluid dynamics
• * Combustion modeling
• * Machine learning applications

Background:

• * Convolutional Neural Networks (CNNs) are increasingly used for subgrid-scale modeling in Large-Eddy Simulations (LES).
• * Evaluating the extrapolation performance of these CNN-based models is crucial for their practical application in turbulent premixed combustion.

Purpose of the Study:

• * To investigate the extrapolation capabilities of CNN models for turbulent premixed combustion.
• * To analyze the influence of training Reynolds numbers, filter sizes, and filter kernels on model performance.
• * To assess the generalization of CNN models to unseen simulation conditions.

Main Methods:

• * Training CNN models on Direct Numerical Simulation (DNS) datasets of methane/air and hydrogen/air jet flames.
• * Testing model performance across varying Reynolds numbers, filter sizes (Gaussian and box kernels), and filter types.
• * Evaluating extrapolation performance on out-of-sample conditions.

Main Results:

• * CNN models exhibit good extrapolation performance when trained at sufficiently high Reynolds numbers.
• * Performance degrades significantly when models trained on low-Reynolds number data are applied to higher Reynolds numbers.
• * Models show satisfactory interpolation and reasonable extrapolation across different filter sizes and kernel types, especially at smaller sizes.
• * Strategic weighting of training data towards larger filter sizes improves generalization.

Conclusions:

• * CNN-based combustion models can generalize effectively across different filter conditions if trained appropriately.
• * High Reynolds number training is key for robust extrapolation performance in turbulent combustion LES.
• * The selection and weighting of training data significantly impact the generalization ability of CNN models.