Abstract:The dynamics of flexible filaments entrained in flow, important for understanding many biological and industrial processes, are computationally expensive to model with full-physics simulations. This work describes a data-driven technique to create high-fidelity low-dimensional models of flexible fiber dynamics using machine learning; the technique is applied to sedimentation in a quiescent, viscous Newtonian fluid, using results from detailed simulations as the data set. The approach combines an autoencoder neural network architecture to learn a low-dimensional latent representation of the filament shape, with a neural ODE that learns the evolution of the particle in the latent state. The model was designed to model filaments of varying flexibility, characterized by an elasto-gravitational number $\mathcal{B}$, and was trained on a data set containing the evolution of fibers beginning at set angles of inclination. For the range of $\mathcal{B}$ considered here (100-10000), the filament shape dynamics can be represented with high accuracy with only four degrees of freedom, in contrast to the 93 present in the original bead-spring model used to generate the dynamic trajectories. We predict the evolution of fibers set at arbitrary angles and demonstrate that our data-driven model can accurately forecast the evolution of a fiber at both trained and untrained elasto-gravitational numbers.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to use a data - driven method to create a low - dimensional dynamic model of the sedimentation of flexible fibers in a fluid. Specifically, the researchers focus on the sedimentation dynamics of flexible fibers in a stationary viscous Newtonian fluid, which is a very important but computationally expensive problem in both biological and industrial processes. The traditional full - physical simulation method has a very large amount of computation because it needs to simulate the movement of each particle in detail. To solve this problem, the paper proposes a data - driven technique that combines an auto - encoder neural network and neural ordinary differential equations (Neural ODE) to create a low - dimensional model of flexible fiber dynamics with high fidelity. ### Main contributions of the paper 1. **Creation of low - dimensional model**: Through machine - learning techniques, especially the auto - encoder neural network, learn the low - dimensional latent representation of the fiber shape. This enables the fiber shape that originally required 93 degrees of freedom to be described to be represented by only 4 degrees of freedom, greatly reducing the computational complexity. 2. **Time - evolution prediction**: Use neural ordinary differential equations (Neural ODE) to learn the time - evolution of the low - dimensional latent state, thereby predicting the dynamic behavior of the fiber under different initial angles and elastic - gravity numbers. 3. **Generalization ability of the model**: The model can not only accurately predict the fiber dynamics in the training data, but also make accurate predictions under unseen initial angles and elastic - gravity numbers, showing good generalization ability. ### Specific problem description - **Background**: The dynamic behavior of flexible fibers in a fluid is crucial for understanding many biological and industrial processes. For example, the behavior of DNA in a fluid is similar to that of elastic fibers, and the processing of pulp fibers is crucial to the paper - making industry. - **Challenge**: The traditional full - physical simulation method has a high computational cost and is difficult to be applied on a large scale. - **Goal**: Develop a data - driven low - dimensional model that can significantly reduce the computational cost while maintaining high accuracy. ### Method overview 1. **Auto - encoder**: Used to learn the low - dimensional latent representation of the fiber shape. 2. **Neural ordinary differential equations**: Used to learn the time - evolution of the low - dimensional latent state. 3. **Dataset**: Use detailed simulation results as a dataset to train the model to predict the dynamic behavior of the fiber under different conditions. ### Results and discussion - **Model performance**: The model performs well on the test dataset and can accurately predict the shape and position evolution of the fiber under different initial angles and elastic - gravity numbers. - **Generalization ability**: The model not only performs well on the training data, but also can make accurate predictions under unseen conditions, showing strong generalization ability. In conclusion, this paper successfully creates a data - driven low - dimensional model by combining an auto - encoder and neural ordinary differential equations, which can efficiently and accurately predict the sedimentation dynamics of flexible fibers in a fluid. This provides new tools and methods for research in related fields.

Data-driven low-dimensional model of a sedimenting flexible fiber

Data-driven low-dimensional model of a sedimenting flexible fiber

A simulation platform for slender, semiflexible, and inextensible fibers with Brownian hydrodynamics and steric repulsion

Bending fluctuations in semiflexible, inextensible, slender filaments in Stokes flow: Toward a spectral discretization

Coarse-graining the dynamics of immersed and driven fiber assemblies

Discrete element simulation studies of angles of repose and shear flow of wet, flexible fibers.

The sedimentation of flexible filaments

An integral-based spectral method for inextensible slender fibers in Stokes flow

A general formulation of Bead Models applied to flexible fibers and active filaments at low Reynolds number

Conformations, correlations, and instabilities of a flexible fiber in an active fluid

Scattering-Informed Microstructure Prediction during Lagrangian Evolution (SIMPLE) -- A data-driven framework for modeling complex fluids in flow

A Bonded Sphero-Cylinder Model for the Discrete Element Simulation of Elasto-Plastic Fibers

Collective Langevin Dynamics of Flexible Cytoskeletal Fibers

Machine learning based non-Newtonian fluid model with molecular fidelity

Prediction of submicron particle dynamics in fibrous filter using deep convolutional neural networks

A fast platform for simulating flexible fiber suspensions applied to cell mechanics

Elastoinertial turbulence: Data-driven reduced-order model based on manifold dynamics

Dynamics of a Data-Driven Low-Dimensional Model of Turbulent Minimal Pipe Flow

Dynamics of flexible fibers in viscous flows and fluids

Brownian microhydrodynamics of active filaments