Convolutional Neural Networks for modeling intrinsic flame instabilities in lean hydrogen flames

This thesis investigates the application of machine learning (ML), particularly Convolutional Neural Networks (CNNs), for the development of reduced-order models (ROMs) aimed at predicting key thermochemical quantities in laminar premixed hydrogen-air combustion, characterized by intrinsic flame instabilities. The primary objective is to assess whether data-driven methods can serve as accurate and efficient surrogates for high-fidelity numerical simulations, supporting ongoing efforts in clean energy research and hydrogen utilization. The methodology is based on a two-stage approach: an a priori phase and an a posteriori phase. In the a priori stage, CNNs are trained on Direct Numerical Simulation (DNS) data to learn mappings from input fields that can be the progress variable alone or in combination with the mixture fraction $Z_{mix}$, to target thermochemical quantities, including source terms and transport properties. Sensitivity analyses examine the influence of key training hyperparameters on learning capabilities, showing that while increasing convolutional kernel size has a marginal impact on accuracy, the training subdomain size (subbox) plays a critical role. In particular, the ratio between subbox size and flame thickness emerges as a key factor: subboxes must span at least twice the local flame thickness ($SB/\delta_L > 2$) to ensure accurate gradient resolution and to prevent boundary-related errors. In the a posteriori stage, the trained CNNs are integrated into an in-house CFD solver and tested under iterative simulation conditions. Results demonstrate that although CNNs can reproduce complex flame dynamics with high initial accuracy, they are prone to error accumulation at each time step due to the absence of physical dissipation mechanisms. Alternative strategies, such as adversarial training using Generative Adversarial Networks (GANs) and retraining on CNN-generated fields—were also explored to improve robustness but provided only limited benefits. These results underline both the promise and current limitations of ML-based ROMs for combustion modeling and suggest a promising path toward future application in turbulent and three-dimensional reactive flow scenarios.