BubbleSH: A Dataset of Rising Bubbles with Deformable Interfaces

Bubbly flows are characterized by complex, multiscale interactions where deformable bubbles influence each other's trajectories and the surrounding fluid. While Direct Numerical Simulations (DNS) provide high-fidelity data, they are computationally expensive and rarely shared in reusable formats. Consequently, researchers often rely on simplified Euler-Lagrange models that sacrifice physical detail for efficiency. The authors introduce BubbleSH to bridge this gap, providing a lightweight, structured dataset that captures the transient, three-dimensional dynamics of bubble swarms. By representing bubble shapes using spherical harmonics, the dataset enables machine learning models to learn both trajectory evolution and surface deformation in a compact, computationally efficient manner.

The dataset is generated using a Front-Tracking (FT) method for incompressible two-phase flow, which explicitly resolves the gas-liquid interface without the artificial coalescence often found in other methods. The researchers simulated air-water systems across 24 configurations, varying bubble diameters (4, 5, and 6 mm) and gas volume fractions (5% to 40%). For each bubble, the dataset records centroid positions, velocities, and a set of 225 spherical harmonic coefficients. This approach achieves a compression ratio of approximately 173x compared to the raw mesh data, making it highly suitable for training geometric deep learning models and probabilistic emulators.

BubbleSH serves as a critical benchmark for the development of data-driven emulators in multiphase fluid dynamics. By providing standardized metrics for both trajectory and shape prediction, the dataset allows researchers to evaluate how well models capture the stochastic, chaotic nature of bubble interactions. It is particularly well-suited for testing generative models that must predict distributions of future states rather than single deterministic outcomes, offering a path toward scalable, high-fidelity simulations that retain the physical accuracy of DNS.