Data-driven modeling of a settling sphere in a quiescent medium

TL;DR

This paper presents data-driven neural network models, including NODEs and NSDEs, to predict the settling behavior of a sphere in a fluid, capturing both trajectory details and statistical features from experimental data.

Contribution

The study introduces neural ODE and SDE models trained on experimental trajectories to predict sphere settling dynamics without explicit fluid simulation.

Findings

NODEs accurately reconstruct trajectories and generalize across initial conditions.

NSDEs effectively model long-term statistical behavior.

Both models capture long-time dynamics, but short-time transients remain challenging.

Abstract

We develop data-driven models to predict the dynamics of a freely settling sphere in a quiescent Newtonian fluid using experimentally obtained trajectories. Particle tracking velocimetry was used to obtain a comprehensive dataset of settling motions, which we use to train neural networks that model the spatial evolution of a spherical particle without explicitly resolving the surrounding fluid dynamics. We employ deterministic neural ordinary differential equations (NODEs) and stochastic neural stochastic differential equations (NSDEs) to reconstruct the sphere's trajectory and capture key statistical features of the settling process. The models are evaluated based on short- and long-time dynamics, including ensemble-averaged velocity evolution, settling time distributions, and probability density functions of the final settling positions. We also examine the correlation between lateral displacement and streamwise velocity and assess the impact of dataset size on predictive accuracy. While NODEs excel in trajectory reconstruction and generalization across different initial conditions, NSDEs effectively capture statistical trends in the long-time behavior but are more sensitive to data availability. Acceleration profiles computed via second-order finite difference schemes confirm that both approaches accurately capture long-time dynamics, though short-time transients pose challenges.