🔬 physics

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

This paper proposes a Vision Transformer-based approach that predicts the full temporal evolution of viscoelastic droplet impact dynamics using only the initial 10–20% of simulation data, thereby reducing computational costs by 80–90% while maintaining physical consistency and accurately capturing complex spreading and bouncing regimes.

Original authors: Diego A. de Aguiar, Cassio M. Oishi

Published 2026-06-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Diego A. de Aguiar, Cassio M. Oishi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a video of a water balloon hitting a table. Sometimes it splatters flat (spreading), and sometimes it bounces back up like a rubber ball (bouncing). Predicting exactly how that balloon will behave is tricky, especially if the liquid is thick and stretchy, like honey mixed with rubber bands.

Scientists usually use powerful computers to simulate these events step-by-step. However, running these simulations is like trying to watch a movie in slow motion frame-by-frame using a supercomputer; it takes a long time and costs a lot of "computing energy." If you want to test thousands of different scenarios (changing the speed, the stickiness of the liquid, or the size of the drop), the computer gets overwhelmed.

The Big Idea: The "Crystal Ball" for Droplets
The authors of this paper asked: Can we teach a computer to watch just the beginning of the video and then guess the rest of the story without running the full simulation?

They built a digital "crystal ball" using a type of Artificial Intelligence called a Video Vision Transformer (ViViT). Think of this AI as a very smart movie critic. Instead of just looking at one frame, it watches the first 10% to 20% of the droplet's journey (the setup) and then predicts the remaining 80% to 90% of the action.

How They Did It

The Training: They fed the AI thousands of computer-generated movies of droplets hitting surfaces. These droplets weren't just water; they were "viscoelastic," meaning they had both liquid and elastic properties (like silly putty).
The Shortcut: Instead of making the computer calculate every single tiny movement of the liquid for the whole video, they let the AI take a shortcut. It only needed to see the first few seconds to figure out the rest.
The Comparison: They tested this "smart critic" (ViViT) against a much simpler, older AI model (called an MLP).
- The Simple AI (MLP): This model was like someone trying to predict a movie by guessing the average of all movies they've ever seen. For short clips, it was okay. But as the movie got longer, it started to hallucinate. The droplet would just turn into a blurry blob, losing all its shape and bouncing behavior.
- The Smart AI (ViViT): This model was like a director who understands the whole story. It kept track of how the droplet stretched, squished, and bounced over time. Even when predicting a long sequence, it kept the droplet looking real and behaving according to the laws of physics.

The Results

Speed: By using this AI shortcut, they saved about 80% to 90% of the computing time. It's like finishing a 10-hour movie in just 1 hour because the AI filled in the gaps.
Accuracy: The smart AI didn't just guess; it preserved the physics. It correctly predicted whether the droplet would spread out like a pancake or bounce back like a ball, even for long time periods. The simple AI failed at this, turning bouncing droplets into spreading ones and vice versa.
Real-World Potential: The authors noted that because this system looks at "volume fraction fields" (which are basically maps showing where the liquid is), it could theoretically be trained on real videos from a camera, not just computer simulations. This means it could eventually learn from real-world experiments.

In a Nutshell
The paper demonstrates that a sophisticated AI (ViViT) can act as a highly efficient "time machine" for fluid dynamics. It watches the start of a droplet's impact and accurately predicts the future, saving massive amounts of computer power while keeping the physics correct. It's a tool that turns a slow, expensive calculation into a fast, smart prediction, specifically for complex, stretchy liquids hitting solid surfaces.

Note: The paper focuses strictly on computer simulations and the AI model itself. It does not claim to have tested this on real medical procedures, industrial manufacturing lines, or specific clinical applications yet, though it suggests the method could be adapted for real experimental data in the future.

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

Problem Statement
The impact of droplets on solid surfaces is a complex fluid dynamics phenomenon critical to applications such as spray cooling, inkjet printing, and pharmaceutical processing. While Newtonian fluid impacts are characterized primarily by the Reynolds ($Re$) and Weber ($We$) numbers, viscoelastic droplet dynamics introduce additional complexity due to elastic effects, requiring the solvent viscosity ratio ( $\beta$ ) and the Weissenberg number ($Wi$). Investigating these dynamics across a broad parametric space using traditional numerical simulations (specifically Computational Fluid Dynamics or CFD) is computationally expensive. Full simulations often require hours to days per run, making extensive parametric studies prohibitive. Furthermore, existing machine learning approaches for droplet dynamics often rely on single-step predictions or simple architectures that fail to capture long-range temporal dependencies, leading to error accumulation and loss of physical consistency over extended prediction horizons.

Methodology
The authors propose a deep learning framework utilizing a Video Vision Transformer (ViViT) to predict the temporal evolution of viscoelastic droplet impacts. The approach replaces the need for full numerical simulations by using only the initial 10% to 20% of a simulation trajectory as a temporal context to predict the remaining 80% to 90%.

Data Generation: The dataset consists of 180 numerical simulations of viscoelastic droplets impacting a solid surface, generated using the open-source flow solver Basilisk. The simulations solve the two-phase incompressible axisymmetric Navier-Stokes equations using the Volume of Fluid (VOF) method. The dataset covers variations in $Re $(2.5–150),$ We $(0.5–40),$ \beta$ (0.1–0.9), and $Wi$ (1.0–5.0), resulting in both spreading and bouncing regimes. A mesh convergence study determined that a $128 \times 128$ resolution was sufficient to preserve physical behavior while minimizing cost.
Model Architectures: Two architectures were compared:
1. Multilayer Perceptron (MLP): A baseline model treating the flattened volume fraction fields as independent inputs to predict the next time step.
2. Video Vision Transformer (ViViT): An architecture adapted from [ADH+21] that processes spatio-temporal data. It treats image patches from a sequence of time steps as tokens, utilizing self-attention mechanisms to capture global spatial and temporal dependencies. The model recursively predicts future volume fraction fields.
Training Strategy: The dataset was split into training, validation, and test sets, with specific $We$ values reserved for validation and testing to evaluate extrapolation capabilities. Three input-output configurations were tested: predicting 1 step from 50 ( $50 \to 1$ ), 50 steps from 50 ( $50 \to 50$ ), and 100 steps from 100 ( $100 \to 100$ ).
Evaluation Metrics: Performance was assessed via predictive accuracy ( $R^2$ , RMSE), image quality (SSIM), preservation of geometric properties (Mean Absolute Error of horizontal/vertical diameters, center of mass, and contact time), and computational cost.

Key Results

Short-term Prediction ( $50 \to 1$ ): Both MLP and ViViT performed reasonably well, though ViViT consistently outperformed the MLP in structural similarity and physical metrics (e.g., contact time), albeit with a higher training time.
Long-term Prediction ( $50 \to 50$ and $100 \to 100$ ): The MLP architecture failed significantly as the prediction horizon increased. The $R^2$ score dropped near zero, and geometric errors exploded, as the model converged toward an average behavior rather than capturing specific regime dynamics. In contrast, the ViViT maintained stable performance across all horizons.
Physical Consistency: The ViViT successfully preserved the distinction between spreading and bouncing regimes, accurately capturing geometric properties and contact times even for long prediction horizons. The MLP failed to maintain physical consistency, producing morphologies that deviated significantly from expected dynamics.
Computational Efficiency: By using the ViViT to predict the majority of the trajectory based on a small initial context, the framework reduces computational costs by approximately 80% to 90% compared to running full numerical simulations.

Significance and Claims
The paper claims that the proposed ViViT-based approach offers a robust alternative to full numerical simulations for predicting viscoelastic droplet impact dynamics. The primary significance lies in the model's ability to capture long-range spatio-temporal dependencies, which is essential for maintaining physical consistency over time—a limitation observed in simpler architectures like MLPs. The authors assert that the framework achieves high predictive accuracy and preserves geometric properties while drastically reducing computational time.

Furthermore, the authors note that because the input data (volume fraction fields) can be extracted from segmented frames of real experimental videos, the framework has the potential to be adapted for training on experimental data, thereby improving the physical fidelity of learned dynamics without relying solely on numerical simulations. The work includes a publicly available Python implementation with a graphical user interface (GUI) to facilitate reproducibility and further exploration by the research community.

More like this