Abstract:Stable partitioned techniques for simulating unsteady fluid-structure interaction (FSI) are known to be computationally expensive when high added-mass is involved. Multiple coupling strategies have been developed to accelerate these simulations, but often use predictors in the form of simple finite-difference extrapolations. In this work, we propose a non-intrusive data-driven predictor that couples reduced-order models of both the solid and fluid subproblems, providing an initial guess for the nonlinear problem of the next time step calculation. Each reduced order model is composed of a nonlinear encoder-regressor-decoder architecture and is equipped with an adaptive update strategy that adds robustness for extrapolation. In doing so, the proposed methodology leverages physics-based insights from high-fidelity solvers, thus establishing a physics-aware machine learning predictor. Using three strongly coupled FSI examples, this study demonstrates the improved convergence obtained with the new predictor and the overall computational speedup realized compared to classical approaches.

What problem does this paper attempt to address?

This paper attempts to solve the problem of slow convergence in fluid - structure interaction (FSI) simulations. Specifically, the author proposes a machine - learning - based predictor to accelerate the convergence of partitioned fluid - structure interaction simulations. ### Problem Background When simulating the interaction between fluids and structures, especially in cases of high added mass, traditional partitioning methods are usually computationally expensive and converge slowly. To accelerate these simulations, existing coupling strategies usually use simple finite - difference extrapolation as a predictor, but with limited effectiveness. ### Core Problem of the Paper The paper aims to accelerate the convergence of fluid - structure interaction problems by introducing a new, non - invasive data - driven predictor. This predictor combines reduced - order models (ROMs) of the solid and fluid sub - problems to provide an initial guess for the nonlinear problem in the next time step. Each reduced - order model consists of a nonlinear encoder - regressor - decoder architecture and is equipped with an adaptive update strategy to improve the robustness of extrapolation. ### Main Contributions 1. **New Predictor**: A machine - learning - based predictor is proposed to accelerate the convergence of fluid - structure interaction problems. 2. **Coupled Reduced - Order Models**: The reduced - order models of the solid and fluid are coupled as an initial guess for the next calculation. 3. **Adaptive Update**: The reduced - order models are updated online to enhance the robustness and accuracy of extrapolation. 4. **Physics - Aware Machine Learning**: Using the physical insights of high - fidelity solvers, a physics - aware machine - learning predictor is established. 5. **Significant Acceleration**: Compared with classical predictors, a faster convergence rate and up to 3.2 - fold acceleration are achieved. ### Mathematical Formula Representation - **Iterative Update Formula**: \[ f^{k + 1}=F(u^k) \] \[ u^{k + 1}=S(f^k) \] - **Aitken Acceleration Formula**: \[ f^k = w_k\tilde{f}^k+(1 - w_k)f^{k - 1} \] where, \[ w_k =-\frac{r_{k - 1}^T(r_k - r_{k - 1})}{\|r_k - r_{k - 1}\|_2^2} \] \[ r_k=\tilde{f}^k - f^{k - 1} \] - **Quasi - Newton Method**: \[ x^k = x^{k - 1}-J^{-1}r_k \] where, \[ r_k=\tilde{x}^k - x^k \] Through these improvements, the paper shows that in three strongly - coupled FSI examples, the new predictor can significantly improve the convergence rate and achieve overall computational acceleration.

Machine-Learning Enhanced Predictors for Accelerated Convergence of Partitioned Fluid-Structure Interaction Simulations

Partitioned Deep Learning of Fluid-Structure Interaction

A Temporal Consistent Monolithic Approach to Fluid-Structure Interaction Enabling Single Field Predictors

Coupling Machine Learning Local Predictions with a Computational Fluid Dynamics Solver to Accelerate Transient Buoyant Plume Simulations

Machine learning accelerated computational fluid dynamics

Accuracy and Stability Enhancements in the Incompressible Finite-Volume-particle Method for Multiphase Flow Simulations.

A stable partitioned FSI algorithm for rigid bodies and incompressible flow in three dimensions

Towards development of enhanced fully-Lagrangian mesh-free computational methods for fluid-structure interaction

A sharp MLS penalty immersed finite element method for fluid-structure interaction of highly deformable slender body in turbulent flow

Efficient split-step schemes for fluid-structure interaction involving incompressible generalised Newtonian flows

A machine learning-oriented pseudo-field approach to accelerate runtime of molecular dynamics simulation of liquids

Computationally Efficient CFD Prediction of Bubbly Flow using Physics-Guided Deep Learning

Learning Mesh Motion Techniques with Application to Fluid-Structure Interaction

Enhancing computational fluid dynamics with machine learning

Adaptive Immersed Mesh Method (AIMM) for Fluid Structure Interaction

A stable added-mass partitioned (AMP) algorithm for elastic solids and incompressible flow

Residual-based physics-informed transfer learning: A hybrid method for accelerating long-term CFD simulations via deep learning

Scaled-up prediction of steady Navier-Stokes equation with component reduced order modeling

Machine learning acceleration of simulations of Stokesian suspensions