Physics-informed constrained bubble generative adversarial network for realistic bubble image generation

Deep generative models provide an efficient way to expand datasets for training and benchmarking bubble image recognition software. Yet, most existing methods generate bubbles that lack physical realism as they often neglect the critical relationship between bubble size and morphology. To bridge this gap, we propose a Bubble Generative Adversarial Network constrained by the known size-morphology relationship, building on recent advances in image recognition and morphology classification. Trained on 12,147 inline-acquired bubble images from bubble column, the model incorporates bubble size as a conditional input and leverages deep convolutional architectures to ensure stable training and high-fidelity texture reproduction. The generated bubbles exhibit both high morphological plausibility and diversity, as demonstrated by a Fréchet inception distance (FID) of 44.51 and a size controllability accuracy of 94.5%. This work establishes a robust framework for generating physically consistent bubble images, thereby providing valuable data support for advanced hydrodynamic analysis and industrial image-processing algorithms. © 2026 American Institute of Chemical Engineers.