<scp>3DBubbles</scp> : An experimental dataset for model training and benchmarking

Abstract Bubbly flow plays an important role across various industries, and the introduction of artificial intelligence (AI) into this field is expected to advance hydrodynamic study. However, the lack of high‐quality experimental datasets tailored for AI model training and benchmarking presents a key challenge. To address this question, an experimental method for three‐dimensional (3D) bubble structure dataset construction is proposed. This method commences with the preparation of a static bubbly flow phantom and 3D structures digitization using an in‐house developed x‐ray computed tomography system. Subsequently, 3D structural parameters and corresponding two‐dimensional (2D) image labels are characterized by spherical harmonic analysis and rendering algorithms. With this method, an experimental dataset named 3DBubbles is constructed, which contains over 10,000 3D bubble structures and 300,000 corresponding 2D projection images. The dataset has been open sourced and can be continuously expanded and improved to serve the models training and benchmarking for the bubbly flow study.