Prediction of Viscoelastic Droplet Impact Dynamics Using a Vision Transformer-Based Approach

This study addresses the high computational cost of simulating viscoelastic droplet impacts on solid surfaces. Traditional numerical methods, such as the Volume of Fluid (VOF) method, require significant resources to resolve the complex interplay between inertial, viscous, surface tension, and elastic forces. The authors propose a surrogate modeling approach using a Video Vision Transformer (ViViT) to predict the droplet's morphology over time.

By feeding the model the initial 10% to 20% of a simulation, the ViViT architecture learns to forecast the remaining 80% to 90% of the droplet's evolution. This autoregressive strategy allows for rapid inference compared to full-scale computational fluid dynamics (CFD) simulations. The model was trained on a dataset of 180 simulations covering various spreading and bouncing regimes, characterized by dimensionless parameters including the Reynolds number (Re), Weber number (We), solvent viscosity ratio (β), and Weissenberg number (Wi).

The ViViT architecture successfully captures the complex dynamics of droplet impact, including the transition between spreading and bouncing regimes. The researchers found that the transformer-based model preserves geometric features and structural similarity significantly better than simpler baseline models like the Multilayer Perceptron (MLP). By leveraging the attention mechanism, the ViViT effectively captures long-range temporal dependencies, which is critical for maintaining physical consistency over longer prediction horizons. The framework is designed to be flexible; since it operates on volume fraction fields, it can theoretically be adapted to incorporate experimental video data, potentially enhancing the fidelity of the predictions.

Computational fluid dynamics is essential for industries like inkjet printing, spray cooling, and pharmaceutical processing, but the cost of exploring large parametric spaces is often prohibitive. This work provides a viable path to accelerate design and optimization cycles by replacing expensive numerical solvers with lightweight machine learning inference. The ability to accurately predict droplet behavior from a small initial context window offers a scalable solution for researchers needing to perform rapid parametric studies without sacrificing physical accuracy.