Data‐Driven Mapping and Classification of Combustion Regimes in a Stainless Steel–Platinum Segmented Microreactor Using a CNN–XGBoost Hybrid Framework

In previous studies, flames could be stably maintained within the designed platinum–stainless steel segmented microreactor. However, difficulties remained in identifying combustion modes. Therefore, this study proposes a method that combines convolutional neural network (CNN) with the XGBoost model to classify and predict different combustion modes. In this work, the combustion characteristics of the burner were analyzed by varying four parameters that affect combustion modes: the flow rates of methane/air and hydrogen/air, as well as the equivalence ratios of methane/air and hydrogen/air. Flame images under different operating conditions were captured and used for CNN and XGBoost modeling to identify and predict combustion modes. Five CNN architectures—AlexNet, VGG16‐BN, ResNet18, ResNet50, and ResNet101—were compared in this study. The results show that, under condition‐grouped fivefold cross‐validation, four of the five CNN architectures classified the flame images with near‐ceiling validation accuracy (99.5%–100%), whereas VGG16‐BN exhibited fold‐dependent training instability (90.6% ± 20.2%), while ResNet18 required the least training time. After classification of the flame images using the trained CNN models, XGBoost further utilized the operating conditions and classification results for training. The trained model was then able to predict the combustion modes across the entire investigated operating range and determine the boundaries between different combustion modes. The results indicate that the hydrogen equivalence ratio emerges as the dominant organizing parameter of the predicted combustion‐mode maps. The manually assigned labels of Set 2 were withheld from the CNN‐assisted workflow and used as reference labels, so that the downstream effect of replacing manual labels with CNN‐generated labels could be quantified; the two workflows produced closely matching regime maps (whole‐map agreement 99.6% ± 0.5% and Class 1 intersection‐over‐union [IoU] 99.5% ± 0.7% over 100 paired condition‐level resampling runs). This study proposes and validates a method that combines CNNs with XGBoost for analyzing flame combustion characteristics, providing a new concept for applying image recognition in the combustion field.