A Hybrid Partitioned Deep Learning Methodology For Moving Interface and Fluid-Structure Interaction

A hybrid partitioned deep learning methodology for moving interface
and fluid-structure interaction
Rachit Guptaa , Rajeev Jaimana,∗

a Department of Mechanical Engineering, University of British Columbia, Vancouver, BC Canada V6T 1Z4
Abstract
arXiv:2102.09095v2 [physics.flu-dyn] 30 Oct 2021
In this work, we present a hybrid partitioned deep learning framework for the reduced-order modeling
of moving interfaces and predicting fluid-structure interaction. Using the discretized Navier-Stokes in the
arbitrary Lagrangian-Eulerian reference frame, we generate the full-order flow snapshots and point cloud
displacements as target physical data for the learning and inference of coupled fluid-structure dynamics.
The hybrid operation of this methodology comes by combining two separate data-driven models for fluid
and solid subdomains via deep learning-based reduced-order models (DL-ROMs). The proposed multi-
level framework comprises the partitioned data-driven drivers for unsteady flow and the moving point
cloud displacements. At the fluid-structure interface, the force information is exchanged synchronously
between the two partitioned subdomain solvers. The first component of our proposed framework relies on
the proper orthogonal decomposition-based recurrent neural network (POD-RNN) as a DL-ROM proce-
dure to infer the point cloud with a moving interface. This model utilizes the POD basis modes to reduce
dimensionality and evolve them in time via long short-term memory-based recurrent neural networks
(LSTM-RNNs). The second component employs the convolution-based recurrent autoencoder network
(CRAN) as a self-supervised DL-ROM procedure to infer the nonlinear flow dynamics at static Eulerian
probes. We introduce these probes as spatially structured query nodes in the moving point cloud to treat
the Lagrangian-to-Eulerian conflict together with convenience in training the CRAN driver. To deter-
mine these Eulerian probes, we construct a novel snapshot-field transfer and load recovery algorithm.
They are chosen in such a way that the two components (i.e., POD-RNN and CRAN) are constrained
at the interface to recover the bulk force quantities. These DL-ROM-based data-driven drivers rely on
the LSTM-RNNs to evolve the low-dimensional states. A popular prototypical fluid-structure interaction
problem of flow past a freely oscillating cylinder is considered to assess the efficacy of the proposed
methodology for a different set of reduced velocities that lead to vortex-induced vibrations. The pro-
posed framework tracks the interface description with acceptable accuracy and predicts the nonlinear
wake dynamics over the chosen test data range. The proposed framework aligns with the development
of partitioned digital twin of engineering systems, especially those involving moving boundaries and
fluid-structure interactions.
Keywords: Fluid-structure interaction, Deep learning-based reduced-order model, Proper orthogonal
decomposition, Convolutional autoencoder, Long short-term memory network, Digital twin
1. Introduction
Fluid-structure interaction (FSI) is a coupled physical phenomenon that involves a mutual interplay
and bidirectional interaction of fluid flow with structural dynamics. This phenomenon is ubiquitous in
∗ Correspondingauthor
Email addresses: [email protected] (Rachit Gupta), [email protected] (Rajeev Jaiman)
Preprint submitted to Journal of Computers & Fluids November 2, 2021

nature and engineering systems such as fluttering flags [26], flying bats [36], offshore platforms and
pipelines [33, 34], oscillating hydrofoils with cavitation [38], two-phase flow in a flexible pipeline [35]
and among others. For example, the two-way coupling between the fluid and solid exhibits rich flow
dynamics such as the wake-body interaction and vortex-induced vibrations [43], which are important to
understand from an engineering design or a decision-making standpoint. Due to the complex character-
istics of the fluid-structure coupling, frequent but reliable practice is to model these complex interplay
and underlying dynamics by solving numerically a coupled set of partial differential equations (PDEs)
describing the physical laws. Of particular interest in these numerical techniques for unsteady fluid-
structure interaction is to accurately simulate the wake-body interaction in terms of vortex-induced loads
and structural displacements, which represent the prominent coupled dynamical effects.
Fluid-structure systems can undergo highly nonlinear interactions involving complicated interface
dynamics and a wide range of spatial and temporal scales. Moreover, these complex spatial-temporal
characteristics are very sensitive to physical parameters and geometric variations. Numerical techniques
such as the arbitrary Lagrangian-Eulerian [19, 33], level-set [70], immersed boundary [59], fictitious do-
main [78] and phase-field modeling [54, 35] can provide high-fidelity PDE-based numerical solutions
and physical insights of the underlying FSI phenomena. Using the state-of-the-art discretization tech-
niques such the finite element method, accurate solutions have been possible by solving millions of fluid
and structural variables using the full-order models (FOMs) and incorporating proper treatment of the
fluid-structure interface [36]. As the complexity of an FSI system elevates, the accurate treatment of
the interface is directly linked with increasing fidelity near the fluid-structure interface, which implies
solving more and more unknown variables numerically [35]. At higher resolutions of space and time,
the equation-based forward simulations can come at the expense of prohibitively large computing time
and high-dimensionality, rendering them almost ineffective in real-time predictions or control required
by digital twin development [72].
The applicability of deep learning has emerged as a promising alternative for constructing data-driven
prediction models for fluid flow [48, 50] and nonlinear dynamical systems e.g., fluid-structure interaction
[49]. Deep learning is a subset of machine learning that refers to the use of multilayered neural net-
works to classify spatial-temporal datasets and make inference from a set of training data [23, 41, 68].
By learning the underlying data-driven input-output model via deep neural networks, the goal is to make
predictions to unseen input data. The construction of heavily over-parametrized functions by deep neural
networks rely on the foundations of the Kolmogorov–Arnold representation theorem [69] and the univer-
sal approximation of functions via neural networks [17, 15, 32]. Using deep neural networks along with
a collection of algorithms, one can find useful features or embedding functions in a low-dimensional
space from the input-output relation in datasets. Deep neural networks have the ability to separate
spatial-temporal scales and automatically extract functional relations from high-dimensional data with
hierarchical importance. Bestowed by the state-of-the-art back-propagation and stochastic gradient de-
scent techniques for estimating weights and biases adaptively, deep neural networks can provide efficient
low-dimensional representation in a flexible way while learning multiple levels of hierarchy in data [23].
While deep neural networks are heavily overparametrized, they have an inherent bias to induce and make
inferences to unseen data which is termed as inductive bias [6]. For example, convolutional neural nets
possess an implicit inductive bias via convolutional filters with shared weights (i.e., translational symme-
try) and pooling to exploit scale separation [6]. Nevertheless, these black-box deep learning techniques
do not account for prior domain knowledge that can be important for interpretability, data efficiency and
generalization.
In the last few years, there is a growing interest to exploit the inherent inductive bias in deep learning
and to infuse explicit bias or domain knowledge in the network architectures for efficient predictions
and interpretability [24]. In that direction, there have been many promising approaches established in
the research community for a synergistic coupling of the deep learning and physical-based models [8,
9]. These models are often trained to represent a full or partial parametrization of a forward physical
2
process while emulating the governing equations. These coarse-grained inference models and surrogate
representations aim to reduce the high computational costs in forecasting dynamics and to model the
quantities of interest [9]. For example, one can use the trained parameters to carry out the state-to-
state time advancements [27, 7, 67] and inverse modeling [14, 45, 58]. Here, we refer to the state-
to-state time advancements as inferring dependent physical variables from the previous states. At the
same time, inverse modeling identifies the physical system parameters from the output state information.
One can also exploit the approximation properties and inductive bias of neural calculus to solve the
underlying differential equations within a deep learning framework [28, 40, 65, 44]. Without appropriate
modifications of the state-of-the-art DL techniques, the approximate solution of the PDEs via black-box
DL approaches may lead to slow training, reduced accuracy, and a lack of generality to multiphysics and
multiscale systems involving complex geometries and boundary conditions [77, 75, 60, 50, 76].
For increasing generality, we can broadly classify the physics-based machine learning into three cate-
gories: (a) the modification of the objective or loss function by adding a regularizer, (b) the adjustment of
neural architecture designs, and (c) the hybridization of deep learning with projection-based model reduc-
tion. In the first category, one can infuse the physical information by applying a regularizer to the standard
loss function. Such regularizers are generally based on physical conservation principles or the govern-
ing equations of the underlying problem. Many researchers have explicitly enforced such regularizers
to boost generalizability; for instance Karpatne et al. [37], Raissi et al. [62], Zhu et al. [79], Erichson
et al. [20], Geneva et al. [21], Mallik et al. [47], and among others. For physically constrained ML
architectures, as the second category, Daw et al. [18] and Chang et al. [12] modified the RNN cell struc-
ture by intermediate constraints while Muralidhar et al. [56] and Ruthotto and Haber [66] took a similar
route for the convolutional neural networks (CNNs). Recently, Li et al. [42] developed the Fourier neural
operators by applying a Fourier transform and linear weight operator on the input before the activation
function in a neural network. The third category involves the development of the hybrid DL-ROMs that
take into account the spatial and temporal domain knowledge in neural networks. The spatial knowledge
is incorporated by effectively building the low-dimensional states using the projection-based techniques
which inherit physical interpretation as discussed in Miyanawala & Jaiman [52]. The temporal knowl-
edge comes from evolving these low-dimensional states in time. We refer to such DL-ROM frameworks
as physics-based because they incorporate physical interpretability via proper orthogonal decomposition
and its variants. These DL-ROMs can operate in synchronization with the FOMs to boost predictive
abilities.
In this paper, we are interested in developing a hybrid DL-ROM framework to address the curse of
spatial dimensionality while utilizing the desired properties of projection-based model reduction and deep
neural network. Our intent is to construct the state-to-state time advancements of the flow fields while
handling the moving interface description in a coupled data-driven system of fluid-structure interaction.
In particular, we address the issue of high-dimensionality by combining the projection-based ROMs with
deep learning techniques such as the convolutional autoencoder and recurrent neural networks based on
long short-term memory. These models can project the high-dimensional data on the low-dimensional
spaces linearly or nonlinearly for data compression [5] and feature extraction [48]. The projected spaces
are formed such that the loss of information is minimum or recoverable. More recently, these spaces have
been combined with deep learning models to enhance the predictive abilities and are termed as projection-
based DL models. This combination of projection and deep learning helps in reducing dimensionality
and boosting real-time predictive abilities. There exist many variants of such hybrid projection-based
DL models in the literature such as the POD-CNN by Miyanawala & Jaiman [52], the POD-RNN from
Bukka et al. [10], the CNN-RNN by Gonzalez et al. [22]. Notably, for the first time, Miyanawala &
Jaiman [52] proposed a hybrid partitioned DL-ROM technique for fluid-structure interaction to combine
the POD and CNN to take the advantage of the optimal low-dimensional representation given by POD
and the multiscale feature extraction by the CNN. Besides the prediction of unsteady flow and wake-body
interaction, such hybrid DL-ROM models have also been explored for other problems including turbulent
3
flow control application [53], bifurcating flow phenomenon [61], combustion [74] and parametric steady-
state solutions of the PDEs [30, 16]. These works reported substantial speed-ups using the projection-
based DL models during the online predictions compared with their full-order counterpart.
The primary objective of this paper is to develop a partitioned and modular data-driven framework that
can simulate a coupled fluid-structure interaction system with efficient predictive abilities. The present
work builds upon the previous works [9, 25], wherein the authors presented an assessment of two DL-
ROMs, the so-called POD-RNN and the CRAN, for the flow past single and side-by-side stationary
cylinder configurations. It was demonstrated that the CRAN framework achieves a longer time series
prediction due to its nonlinear projection space compared with the POD-RNN, but the former relies on a
uniform grid. This uniform grid can lose the interface description when projecting and interpolating the
fluid variables from the full-order grid. Although the POD-RNN has comparatively lesser flow predictive
abilities for complex flows, this model recovers a high-dimensional interface during the reconstruction
process. Therefore the question arises is there a way to learn and predict structural motion and fluid
dynamics in a modular data-driven framework? How can we develop an efficient data-driven model that
predicts the flow fields and the moving interface description simultaneously?
In this paper, we develop a deep learning based reduced-order modeling framework to address the
aforementioned questions. We consider a partitioned data-driven approach to couple the fluid and solid
subdomains via deep neural networks. Specifically, the present work unifies the attributes of the POD-
RNN and the CRAN frameworks for reducing the coupled dynamics and inferring the fluid flow inter-
acting with a freely vibrating structure. Of particular interest is to predict the flow fields via the CRAN
framework, while handling the moving interfaces via the POD-RNN. Using high-fidelity time series data,
we first project the dataset to a low-dimensional subspace using POD and then the time-dependent coeffi-
cients of the POD are iteratively trained using the recurrent neural network. We model the flow variables
in a uniform with a static pixeled grid via convolutional architecture. This simplified procedure addresses
the Lagrangian-to-Eulerian conflict by embedding the full-order fluid-structure data on the uniform grid
while providing convenience in the training of the CRAN framework. Based on an unstructured irregular
grid for the full-order simulation, we propose a strategy termed as the snapshot-field transfer and load
recovery (snapshot-FTLR) to select the structured grid for the fluid domain. Together with the CRAN
framework, this strategy maintains the accuracy of the full-order interface forces and provides a reduced
learning model for fluid-structure interaction. The proposed end-to-end partitioned DL-ROM framework
is modular and fully data-driven; hence, it aligns with the development of a digital twin involving FSI
effects.
The article is organized as follows: Section 2 describes the full-order governing physical model and
a reduced-order representation of the fluid-structure interaction. Section 3 presents the hybrid partitioned
DL methodology and the DL-ROM components for the solid and fluid dynamics. Section 4 introduces
the snapshot-FTLR technique and the DL-ROM grid search. The article ends with an application to the
VIV motion of a freely oscillating cylinder for different set of reduced velocities in section 5 and the main
conclusions are provided in section 6.
2. Full-order vs reduced-order modeling
This section starts by describing the full-order equations of a coupled fluid-structure interaction, fol-
lowed by a brief description of reduced-order modeling.
4
2.1. Full-order modeling
A coupled fluid-solid system in an arbitrary Lagrangian-Eulerian (ALE) reference frame is modeled
by the isothermal fluid flow interacting with a rigid body structure:
∂ uf
ρf + ρ f (uf − w) · ∇uf = ∇ · σ f + bf on Ωf (t), (1)
∂t
∂ ρf
+ ∇ · (ρ f uf ) = 0 on Ωf (t), (2)
∂t
∂ us
ms + cs us + ks (ϕ s (z0 ,t) − z0 ) = Fs + bs on Ωs , (3)
∂t
where superscripts f and s denotes the fluid and structural variables, respectively. In the fluid domain
Ωf , uf and w represent the fluid and mesh velocities, respectively and bf is the body force, f
and σf is the
f f f f f T
Cauchy stress tensor for a Newtonian fluid written as σ = −p I + µ ∇u + (∇u ) . Here, p is the
pressure in the fluid, I is the identity tensor and µ f is the fluid viscosity. Any arbitrary submerged rigid
body Ωs experiences transient vortex-induced loads and, as a result, may undergo large structural motion
if mounted elastically. The rigid body motion along the two Cartesian axes is modeled by Eq. (3), where
ms , cs and ks denote the mass, damping and stiffness matrices, respectively. us represents the rigid body
motion at time t with Fs and bs as the fluid traction and body forces acting on it, respectively. Here, ϕ s
denotes the position vector that transforms the initial position z0 of the rigid body to time t. This coupled
system must satisfy the no-slip and traction continuity conditions at the fluid-body interface Γfs (t) as
follows:
uf (t) = us (t) , (4)

Z
σ f · ndΓ + Fs = 0, (5)
Γfs (t)
where n and dΓ denote the outward normal and the differential surface area of the fluid-solid interface,
respectively. While Eq. (4) enforces the velocity continuity on the moving interface, the balance of the
net force exerted by the fluid on the rigid body is given by Eq. (5). Due to the body-conforming Eulerian-
Lagrangian treatment, the interface conditions provide accurate modeling of the boundary layer and the
vorticity generation over a moving body. The coupled differential equations in Eqs. (1)-(3) are numer-
ically solved using the Petrov–Galerkin finite element and the semi-discrete time stepping [33]. The
weak form of the incompressible Navier-Stokes equations is solved in space using equal-order isopara-
metric finite elements for the fluid velocity and pressure. We employ the nonlinear partitioned staggered
procedure for the stable and robust coupling of the fluid-structure interaction [33]. This completes the
description of the coupled full-order FSI system.
In relation to the model-order reduction, the coupled equations (Eqs. (1)-(3)) can be written in an
abstract state-space form as
dz
= F(z), (6)
dt
where z ∈ RM is the state vector for a coupled FSI domain with a total of M variables in the system. For
our fluid-body system in the current study, the state vector involves the fluid velocity and the pressure as
z = {uf , pf } and the structural velocity includes the three translational degrees-of-freedom. Note that the
pressure p f can be written as some function of density ρ f via the state law of a fluid flow. The right hand
side term F represents a dynamic model and can be associated with a vector-valued differential operator
describing the spatially discretized PDEs in Eqs. (1)-(3). The temporal term dz/dt is the system dynamics
which determines the instantaneous physics of a fluid-structure system and is useful for creating a low-
5
order representation. The resultant spatial-temporal dynamics of the fluid-structure interaction are driven
by the inputs such as the modeling parameters and boundary conditions. Next, we briefly review a data-
driven reduced-order modeling based on traditional projection and convolutional autoencoder techniques.
2.2. Reduced-order modeling

From the perspective of a data-driven approach, the idea is to build the differential operator F by pro-
jecting onto a set of low-dimensional trial subspace. In that sense, one can decompose F(z) to encapsulate
the constant term C, a linear Bz and nonlinear F0 (z) dynamical components as
F(z) = C + Bz + F0 (z). (7)
Using the Galerkin-based ROMs, we represent the state vector z ∈ RM via a subspace spanned by the
column vectors of the low-dimensional modes V ∈ RM×K , with K << M. The matrix V is referred
to as the reduced basis spanning the subspace onto which the dynamics is projected. This allows to
approximate the state vector z as V z̃ with z̃ ∈ RK and thereby reducing the system dynamics as
d z̃
= V T C + V T BV z̃ + V T F0 (V z̃). (8)
dt
Here, the reduced-order space is constructed by defining a choice of modes V ∈ RM×K using the snapshot
matrix Z = {z1 z2 . . . zS } ∈ RM×S , with M number of variables and S denotes the number of snapshots.
The columns of the matrix V = {V 1 V 2 . . . V K } form an orthonormal basis of Z̃ = {z̃1 z̃2 . . . z̃S } ∈
RK×S , a K dimensional subspace of RM . The linear expansion of the state vector snapshot matrix Z = V Z̃
allows the reduction to take place through special choices of subspace Z̃ for any dataset. For example,
the well known POD derives this subspace to be such that the manifold variance is preserved as much as
possible when projected to Z̃, given a fixed dimension constraint K. Using the singular value decomposi-
tion (SVD) decomposition, the snapshot matrix can be expressed as Z = V ΣW T = ∑KJ=1 σJ vJ wTJ , where
vJ are the POD modes of the snapshot matrix Z and Σ ∈ RK×K is a diagonal matrix with diagonal entries
σ1 > σ2 > . . . > σK > 0. The total energy contained in each POD mode vJ is given by σJ2 . V and W are
the matrices with orthonormal columns that represent the eigenvectors of ZZT and ZT Z, respectively.
It is well known that the above POD-Galerkin process is very effective for dimensionality reduction
of the linear term while the nonlinear term cannot be reconstructed properly for the unsteady wake flow
dynamics. Therefore, the linear POD method may result in a similar order of computational expense to the
full-order simulation. For a low-order representation of such nonlinear terms, hyperreduction techniques
such as the discrete empirical interpolation method [13] and energy-conserving sampling and weighting
[3] method can provide an additional level of approximation for the dimensionality reduction. These
hyperreduction strategies can reduce the required number of modes, hence decreasing the computational
cost while capturing the nonlinear regions properly [51]. All these techniques primarily aim to resolve
the nonlinear term, i.e, V T F0 (V z̃), where z̃ is the reduced-order state of full-order state z. We note
that these decomposition techniques are dependent on the choice of dataset, parameters and the initial
condition of a nonlinear system. For instance, it has been shown that for the flow problems with small
Kolmogorov n-width, the POD modes optimally reduce the dimension of the nonlinear unsteady flows
[52, 9, 8]. However, in general, these empirical projection-based reduced-order models can come at the
cost of large subspace dimensions for turbulence or convection-dominated problems characterized by
large Kolmogorov n-width.
To address the drawbacks of the projection-based ROMs, the use of autoencoders provides a promis-
ing alternative for constructing the low-order projections of the state vector snapshot matrix Z. In an
autoencoder, F is trained to output the same input data Z such that Z ≈ F (Z; θ AE ), where θ AE are the
parameters of the end-to-end autoencoder model. The process is to train the parameters θ AE using an
6
iterative minimization of an error function E
θ AE = argminθ AE [E(Z, F (Z; θ AE ))]. (9)
For the use of the autoencoder as a dimension compressor, the dimension of the low-order space called
the latent space, say H̃, is smaller than that of the input or output data Z. When we obtain the output
F (Z) similar to the input such that Z ≈ F (Z), the latent space is a low-dimensional representation of its
input which provides a low-rank embedding. In an autoencoder, the dimension compressor is called the
encoder Fe and the counterpart is the decoder Fd . Using the encoder-decoder architecture, the internal
procedure of the autoencoder can be expressed as H̃ = Fe (Z), Z = Fd (H̃).
For a general nonlinear system, one can construct the subspace projection as a self-supervised DL
parametrization of the autoencoders without any SVD process of the state Z. Instead, the decomposi-
tion can be achieved by generating the trainable layers of the encoder and decoder space such that the
error function is minimized. By using a series of convolutional and nonlinear mapping process, the au-
toencoder can lead to an efficient construction of a compressed latent space to characterize the reduced
dynamics of a given nonlinear system. The autoencoders can be interpreted as a flexible and nonlinear
generalization of POD [63]. Both projection-based POD and convolutional autoencoder help in reducing
the dimensionality of the incoming data Z by encoding dominant patterns from the high-dimensional
data. We next turn our attention to the proposed hybrid partitioned DL-ROM for solving fluid-structure
interaction problems.
3. Hybrid partitioned DL-ROM framework
Herein, we present a hybrid partitioned DL-ROM which is trained using the high-dimensional sim-
ulation generated by solving the full-order fluid-structure interaction model. Consistent with the high-
dimensional FSI solver, we consider two separate data-driven solvers for fluid and solid subdomains with
their own independent neural network in a partitioned manner. Once the high-dimensional data is effec-
tively reduced via POD and convolutional autoencoder, the low-dimensional states can be used to infer
the coupled behavior of a fluid-structure system. Together with POD and convolutional autoencoder for
the dimensionality reduction, we use the well-known LSTM-RNNs for evolving the low-dimensional
states in time. The mathematical structure of the LSTM-RNN represents a nonlinear state-space form.
This particular attribute makes them suitable for a nonlinear dynamical system of fluid-structure inter-
action. When the dimensionality of the data is reduced using POD and evolved in time using RNN, the
hybrid DL-ROM is referred to as the POD-RNN. On a similar note, when the low-dimensional states,
obtained using a convolutional autoencoder, are similarly evolved in time, the hybrid framework is called
the CRAN.
For predicting the coupled dynamics of the fluid flow interacting with a freely vibrating structure, we
unify the combined effects of the POD-RNN and the CRAN frameworks for the partitioned fluid and solid
sub-domains. The illustration of this hybrid partitioned DL-ROM approach is provided in Fig. 1. The first
partition or level of our framework is the POD-RNN for learning the solid displacements, while the second
partition is the CRAN for learning the flow fields in a uniform pixel grid. On a given training data, the
POD-RNN model is employed to track the structural displacements. This tracking helps in constructing
a level-set function. A level-set function is a distance function based on structural geometry. It is used
for setting the level set boundaries on the computational grid based on the position of the structure with a
binary condition as the input function. Miyanawala & Jaiman [48, 50] presented a CNN model combined
with the level set to predict the unsteady fluid forces over stationary geometries. Further details about
the coupling of a level-set function with convolutional neural network can be found in [48, 50, 52]. An
interface load recovery is constructed to select the pixel grid for the CRAN framework while extracting
the interface information using the level-set and forces.
7
𝛀𝐟 : CRAN
𝛀𝐟𝐬
𝐒𝑛 (𝐲ො 𝑛+𝑝 , 𝐒෠ 𝑛+𝑝 )
𝐒 𝑛−1 (𝐲ො 𝑛+𝑝−1 , 𝐒෠ 𝑛+𝑝−1 )
𝐒1
𝐅ത𝑏 INTERFACE LOAD RECOVERY 𝛟
(𝐲ො 𝑛+1 , 𝐒෠ 𝑛+1 )
𝛀𝐬 : POD-RNN
Flow field, rigid body motion & force

predictions
𝐲𝑛
𝐲 𝑛−1
𝐲1 1
Figure 1: Illustration of a hybrid partitioned DL-ROM framework for fluid-structure interaction. The field variables Ωf (t) are
learned and predicted on the uniform grid using the CRAN (blue box), while the interface information Ωs with moving point cloud
is predicted via the POD-RNN (red box). These boxes exchange the interface information (grey box) that couples the system via
level-set Φ and force signals F̄b . The yellow box demonstrates the synchronized predictions. As the CRAN learns the flow variables
on the uniform grid (blue box), the level-set Φ is utilized in calculating the pixelated the force signals F̄b on the interface.
The solid and fluid equations are solved sequentially in a partitioned manner and constrained at the
interface with traction and velocity continuity in a partitioned full-order solver for fluid-structure in-
teraction. Analogous to the full-order partitioned solver, we apply the DL-ROMs sequentially for the
solid and fluid variables and couple them via an interface recovery process. The choice of the so-called
physics-based DL drivers can be different and dependent on the nature of the problem at hand. How-
ever, the interface load recovery is general data coupling for learning and modeling of fluid-structure
interaction with pixel-based flow field information. We first elaborate on the integration of POD-based
low-dimensional approximation with the RNN process for predicting structural motion.
3.1. The POD-RNN framework: structural motion

This DL-ROM technique proposed by Bukka et al. [10] utilizes high-fidelity time series of snapshot
data obtained from the full-order simulations or experimental measurements. These high-dimensional
data are decomposed into the spatial modes (dominant POD basis), the mean field, and the temporal
coefficients as the reduced-order dynamics using the Galerkin projection of POD. The spatial modes and
the mean field form the offline database, while the time coefficients are propagated and learned using
variants of RNNs: closed-loop or encoder-decoder type. This projection and propagation technique is
hence called the POD-RNN. In this work, we learn and infer the full-order ALE mesh displacements
using the POD-RNN. The key assumptions adopted here are that (a) the complete position and shape
8
of the exact fluid-solid interface are inherent in the point cloud displacements, and (b) the POD energy
spectrum for these mesh displacements are fast decaying due to the finite degrees of freedom of the rigid
body motion. This can allow us to compress the displacements using the POD modes and then propagate
with RNNs.
Let Y = {y1 y2 ... yn } ∈ Rm×n be the ALE displacement snapshot matrix along a given cartesian di-
rection from the initial position y0 . Here, yi ∈ Rm is the displacement snapshot at time t i and n represents
the number of such snapshots. m n are the number of data probes, for instance, here the number mesh
nodes in the full-order simulation. The target is to predict the future mesh position: ŷn+1 , ŷn+2 , ... using
the training dataset Y. The POD-RNN framework can be constructed by the following three step process:
Step 1: Construct the POD modes ν for the dataset Y

Given the n snapshots of the ALE displacements along any co-ordinate axis, determine the offline
database: the temporal mean vector (y ∈ Rm ) and the POD basis modes (ν ∈ Rm×k ) using the snapshot
matrix of the ALE displacement Y. Here k < n m and k is selected based on the energy hierarchy of the
POD modes. This projection aims to reduce the order of the high-dimensional moving point cloud from
O(m) to O(k). The low-dimensional states are obtained linearly using the POD reconstruction technique
yi ≈ y + νAiν , i = 1, 2, ..., n, (10)
where Aiν = [ai1 i2 ik T k

ν aν ... aν ] ∈ R are the time coefficients of the k most energetic modes for the time
instant t . The low-dimensional states Aν = {A1ν A2ν ... Anν } ∈ Rk×n are hence generated for the dataset
i
n+1 n+2
Y. Once generated, the problem boils down to predict Âν , Âν , ... using the reduced-order state Anν in
an iterative manner. For a detailed analysis on linear or nonlinear POD reconstruction, readers can refer
to [51, 11].
Step 2: Supervised learning and prediction of the low-dimensional states using LSTM-RNN
Training consists of learning a dynamical operator gν that allows to recurrently predict finite time
series of the low-dimensional states. We employ LSTM-RNN, which exploits both the long-term depen-
dencies in the data and prevents the vanishing gradient problem during training [31]. This is employed as
a one-to-one dynamical mapping between the low-dimensional states Ai−1 ν and Aν .
i
i
Aiν = fν (yi ), Âν = gν (Ai−1
ν ; θν,evolver ), i = 1, 2, ..., n, (11)
where fν is the known function relating the low-dimensional state and the full-order ALE data. The alge-
braic relation in Eq. (10) can be re-written as Aiν ≈ ν T (yi − y) ≈ fν (yi ). Hence, fν is the known mapping
function utilized to obtain the POD time coefficients from a full-order snapshot. The term gν is the
i
trainable operator parametrized by θν,evolver , which outputs a time advanced state Âν with an input Ai−1ν .
To find θν,evolver , one can set the LSTM-RNN such that the low-dimensional states {A1ν A2ν ... An−1 ν }
2 3 n
are mapped to the time advanced low-dimensional states {Âν Âν ... Âν } in a way that gν mapping is
one-to-one
i
Âν = gν (Aνi−1 ; θν,evolver ), i = 2, 3, ..., n, (12)
where θν,evolver are the weight operators of the LSTM cell. Once trained, θν,evolver forms an offline pa-
rameter recurrent space which generates a finite amount of the predicted low-dimensional states. This
closed-loop network predicts the output iteratively with the same trained weights, meaning that the gen-
erated output is fed as an input for the next prediction.
Step 3: Reconstruction to the point cloud ŷn+1

n+1
In Step 2, the reduced-order dynamics is predicted using the LSTM-RNN. Once Âν is generated, it
9
𝐲ത
𝐲𝑛 𝐀𝑛𝜈 ෡𝑛+1
𝐀 𝜈 𝐲ො 𝑛+1
𝐲ො 𝑛+1 ෡𝑛+1
𝐀 𝜈
෡𝑛+2
𝐀 𝜈 𝐲ො 𝑛+2
𝒱T 𝐑𝐍𝐍 𝒱
+
…
…
…
…
𝐲ො 𝑛+𝑝−1 ෡𝑛+𝑝−1
𝐀 𝜈 ෡𝑛+𝑝
𝐀 𝐲ො 𝑛+𝑝
𝜈
Figure 2: Schematic of the POD-RNN framework with a block diagram for the iterative prediction. With one input driver yn ,
the predictions ŷn+1 , ŷn+2 , ..., ŷn+p are achieved autonomously. The diagram illustrates the prediction of p number of time steps
from ŷn+1 to ŷn+p . The dashed lines imply an iterative process of generating all the predictions between ŷn+2 and ŷn+p using the
POD-RNN method.
is straightforward to reconstruct the full-order ALE displacement at t n+1 as follows:

n+1
ŷn+1 ≈ y + ν Âν . (13)
Hence, to track the moving point cloud, the predicted low-dimensional states are linearly combined with
the temporal mean vector (y ∈ Rm ) and the POD basis modes (ν ∈ Rm×k ). An illustration of the entire
predictive process is shown in Fig. 2. This technique of the POD-RNN on the ALE displacements can be
capable of predicting the FSI motion for finite time steps ŷn+1 , ŷn+2 , ..., ŷn+p autonomously provided that
the error is within the acceptable range of accuracy.
Normalisation of dataset: In the POD-RNN, we train and predict the time coefficients obtained from
the POD analysis of the full-order dataset. Let Aν = {A1ν A2ν ... Anν } ∈ Rk×n be the low-dimensional states
obtained from the POD analysis of the ALE snapshot matrix Y. The scaling of Aν is conducted as
Aiν − Aν,min
Aiν,s = , i = 1, 2, ..., n, (14)
Aν,max − Aν,min
where Aν,max and Aν,min are the maximum and minimum values in Aν , respectively, and arranged as
a vector. This ensures Aiν,s ∈ [0, 1]k . The obtained dataset matrix Aν,s = A1ν,s A2ν,s . . . Anν,s ∈ Rk×n

consists of the normalised time coefficients.

Training: The training consists of finding the evolver parameters of the LSTM-RNN θν,evolver such that
the observable loss is minimized
"
i i−1 2#
1 n Aν,s − gν (Aν,s ; θν,evolver ) 2
θν,evolver = argmin(θν,evolver ) ∑ 2
. (15)
n − 1 i=2 Aiν,s 2
This loss function is iteratively minimised over the entire training sequence using the adaptive moment
optimization [39].
10
3.2. The CRAN framework: flow field
The POD-RNN method can provide an optimal low-dimensional space (encoding) in which the modes
are not only computationally inexpensive to obtain but can be physically interpretable as well [51]. How-
ever, there are several problems with the POD encoding for highly nonlinear flow physics: (a) the en-
coding scales linearly, which may cause significant loss of the flow physics characterized by a large
Kolmogorov n-width, (b) the POD reconstruction of the flow problems dominated by turbulence often
results in the slow decay of the POD energy spectrum, implying that the number of POD modes can
increase significantly and be difficult to learn, and (c) the POD basis can have additional orthogonality
constraint on the low-dimensional space which can limit its flexibility in general. So an alternative ap-
proach is to utilize a much more flexible encoding that can address the above challenges. One method
is to use CNNs instead of POD, which leads to a non-intrusive type of DL-ROM framework called the
CNN-RNN or simply the CRAN. Furthermore, as depicted in our previous work [9], CRAN outperforms
the POD-RNN in complex flow predictions by nearly 25 times. Hence, we rely on the CRAN technique
for learning nonlinear flow variables due to the aforementioned advantages.
While operating synchronously in a nonlinear neural space, this projection and propagation technique
extracts the low-dimensional embedding whereby the flow variables are extracted via CNNs and the
encoding evolved via LSTM-RNN. Since there is no knowledge of the mean or basis vectors here, we
construct a decoding space of transpose convolution that up-samples the low-dimensional encoding back
to the high-dimensional space. It is assumed that the solution space of the unsteady flow attracts a low-
dimensional subspace, which allows the embedding in the high-dimensional space. This end-to-end
convolutional autoencoder architecture is illustrated in Fig. 3. For the sake of completeness, the CRAN
architecture is further elaborated in this section. For more details about the CRAN architecture, readers
can refer to our previous work [9].
The CRAN framework needs a high-fidelity series of snapshots of the flow field obtained by full-
order simulations. Let S = {S1 S2 ... Sn } ∈ RNx × Ny ×n denote the 2D snapshots of a field dataset (such
as pressure or velocity data from a flow solver), where Si ∈ RNx ×Ny is the field snapshot at time t i and n
represents the number of such snapshots. Nx and Ny are the number of data probes in the respective Carte-
sian axes. For instance, here, the number of fixed (Eulerian) query points introduced in the moving point
cloud. These probes are structured as a spatially uniform space for the CRAN architecture as it relies on
the field uniformity. These probes are optimally selected based on the field convergence and the interface
load recovery using the snapshot-FTLR method as discussed in section 4. The target of the CRAN-based
n+1 n+2
end-to-end learning is to encode-propagate-decode the future values at the field probes: Ŝ , Ŝ , ...
using the training dataset S. The CRAN framework is constructed using the following process:
Step 1: Find the nonlinear encoding feature space for the dataset S
Given the n snapshots of any field data, construct a trainable neural encoder space using layers of linear
convolutional kernels with nonlinear activation (Conv2D) as shown in Fig. 3. The dimensionality of
the field 2D probes is conveniently and gradually down-sampled using the convolutional filters and the
plain feed-forward networks until a finite size of the low-dimensional state is reached, similar to the POD
encoding as discussed earlier. The update of time coefficient can be expressed as
Aic = fc (Si ; θc,enc ), i = 1, 2, ..., n, (16)
where fc is the trainable encoder space that is parametrized by θc,enc . The time coefficients Aic =
ih T 1 2 n
[ai1 i2 h i
c ac ...ac ] ∈ R represent h encoded features at time t . The low-dimensional states Ac = {Ac Ac ... Ac } ∈
R h×n of the dataset S are determined in a self-supervised fashion of the autoencoder, with no energy or
orthogonality constraint unlike the POD. Here, h < n (Nx × Ny ) and h is the unknown hyperparamter
for the optimal feature extraction based on the input dataset S. This projection reduces the order of high-
dimensional field from O(Nx × Ny ) to O(h). After the decomposition, the problem boils down to predict
11
n+1 n+2 n+1 n+2
Âc , Âc , ... using the low-dimensional state Anc and spatially up-sample to Ŝ , Ŝ , ... as generative
outputs.
Step 2: Supervised learning and prediction of the low-dimensional states using LSTM-RNN
Similar to the POD-RNN, the training of the low-dimensional space consists of learning a one-to-one
dynamical operator gc that allows to predict a certain series of the low-dimensional states. Without a loss
of generality, we employ the LSTM-RNN in a closed-loop fashion. The LSTM-RNN is employed as a
state transformation between the low-dimensional states Ai−1 c and Aic . Instantiate the network such that
1 2 n−1
the low-dimensional states {Ac Ac ... Ac } are mapped to the time advanced low-dimensional states
2 3 n
{Âc Âc ... Âc } so that dynamical operator gc is a one-to-one transformation
i
Âc = gc (Ai−1
c ; θc,evolver ), i = 2, ..., n, (17)
where θc,evolver are the weight operators of the LSTM cell. Likewise, once trained, θc,evolver also forms an
offline recurrent parameter space which predicts finite time steps of the self-supervised features obtained
from the neural encoder space.
n+1
Step 3: Generating the high-dimensional time advanced state Ŝ
n+1
Once Âc is generated, there is a need to prolongate the time advanced features at t n+1 using gradual
reverse convolution (DeConv2D) as these features are often hard to interpret. As there is no knowledge of
the mean or the POD basis vectors, we rely on decoding these low-dimensional features gradually using
a decoder space which is a mirror of the encoder space.
n+1 n+1
Ŝ = fc−1 (Âc ; θc,dec ), (18)
where fc−1 is the trainable decoder space that is parametrized by θc,dec . The trained LSTM-RNN is em-
n+1 n+2 n+p
ployed to iteratively generate the desired future states Ŝ , Ŝ ...., Ŝ starting from Sn . This technique
of the CRAN on the flow variables is self-supervised and is capable of predicting the flow variables at the
fixed probes provided that the prediction error is within the acceptable range of accuracy. The complete
predictive process is illustrated in Fig. 3.
Normalization of dataset: We normalise the flow dataset snapshot matrix S = {S1 S2 ... Sn } ∈
RNx × Ny ×n by first finding the fluctuation around the temporal mean using
S0i = Si − S, i = 1, 2, ..., n, (19)
where S = 1n ∑ni=1 Si is the temporal average over the entire dataset and S0 = {S01 S02 ... S0n } ∈ RNx × Ny ×n
are the fluctuations around this mean. Next the scaling of S0 is conducted as
S0i − S0min
S0is = , i = 1, 2, ..., n, (20)
S0max − S0min
where S0max and S0min are the maximum and 01the02 minimum vales in S0 , respectively, and arranged as a
0 0n
matrix. The obtained dataset matrix Ss = Ss Ss . . . Ss is further broken up into a set of Ns finite time
training sequences, where each training sequence consists of Nt snapshots. The training dataset with the
above modifications has the following form:
S = Ss01 Ss02 . . . Ss0Ns ∈ [0, 1]Nx ×Ny ×Nt ×Ns

(21)
h i
where each training sample Ss0 j = S01 02 Nt
s, j Ss, j . . . Ss, j is a matrix consisting of the scaled database.
12
Spatial down-sampling
… Reshape
Conv2D #1 Conv2D #2 Conv2D #𝑛𝑐𝑜𝑛𝑣

Input:
𝐒𝑛 FC # 𝑛𝑓𝑢𝑙𝑙 ෡𝑖+1
LSTM-RNN 𝐀 𝑐
Generative output: Temporal Temporal

𝐀𝑛𝑐 ෡ 𝑐 𝑛+1
𝐀
𝐒෠ 𝑛+1 , … , 𝐒෠ 𝑛+𝑝 encoding decoding
𝐀𝑖𝑐
෡ 𝑐 𝑛+𝑝
𝐀
DeConv2D #𝑛𝑐𝑜𝑛𝑣 DeConv2D #(𝑛𝑐𝑜𝑛𝑣 −1) DeConv2D #1
… Transform FC # 𝑛𝑓𝑢𝑙𝑙
Spatial prolongation
6
Figure 3: Schematic of the CRAN framework for flow field prediction. The encoding is achieved by reducing the input dimension
from Nx ×Ny to Ac via CNNs (Conv2D) and the fully connected networks. The decoding is achieved using the fully connected layers
and the transpose CNNs (DeConv2D). Between the encoding and decoding space, the LSTM-RNN evolves the low-dimensional
state Ac .
Optimization: A loss function is constructed that equally weights the error in the full-state recon-
struction and the evolution of the low-dimensional representations. The target of the training h is to findi
0j
the CRAN parameters θc = {θc,enc , θc,dec , θc,evolver } such that for any sequence Ss = Ss, j Ss, j . . . SNs,tj
01 02
h i
and its corresponding low-dimensional representation A1c, j A2c, j . . . ANc,tj the following error between the
truth and the prediction is minimized:
 2 2
Nt S0is, j − fc−1 ( fc (S0is, j ; θc,enc ); θc,dec ) (1 − β ) Nt Aic, j − gc (Aic, j ; θc,evolver )
β 2 2
θc = argminθc  ∑ 2
+ ∑ 2
Nt Nt − 1 i=2

i=1 S0is, j Aic, j
2 2
(22)
where β = 0.5. The value of β is chosen in order to give equal weightage to the errors in the full state
reconstruction and feature predictions, respectively. This loss function is iteratively minimised for all the
j = 1, 2, . . . , Ns training sequences using the adaptive moment optimization [39].
4. Snapshot-field transfer and load recovery
We introduce a novel procedure for the partitioned mesh-to-mesh field transfer between the full-order
grid and the neural network grid. A convolutional neural network relies on a spatially uniform input data
stream for identifying patterns in datasets. This is accounted for the uniform dimension and operation of
the adaptive CNN kernels. Given that most practical FSI problems are modeled in a highly unstructured
and body conformal mesh, there is a need to develop a flow field data processing step to interpolate the
scattered information as snapshot images before learning them in the CRAN framework. Consequently,
in this section, we introduce a general data processing step that can be utilized for mapping the flow
field variables from an unstructured grid of the full-order model to a uniform grid of the DL-ROM. We
13
(3) Pixelated
ത𝑏 (𝑡)
force: 𝐅
(2) Level-set (4) Damp noises:

ζ𝐅ത𝑏 (𝑡)
𝚽
DL grid
𝐒(X, Y, 𝑡)
(1) Field transfer (5) Interface load
recovery: Ψ𝐅 ത𝑏 (𝑡)
CFD grid
𝐬(x 𝑡 , y 𝑡 , 𝑡)
Figure 4: Illustration of an iterative cycle for mesh-to-mesh field transfer and load recovery to determine the CRAN snapshot grid.
See the details of all variables in the main text.
achieve this via interpolation and projection of the field information in an iterative process that allows
a recoverable interface force data loss. Once this loss is observed in the training forces, we correct
them by reconstructing to a higher-order CFD force. We select the interface force as the primary criteria
because the boundary layer forces are highly grid sensitive. This iterative process is cyclically performed
as shown in Fig. 4. We refer to the entire cyclic process as a snapshot-field transfer and load recovery
(snapshot-FTLR) method.
Herein, our intent is to find the best DL-ROM grid that can recover the full-order interface load
correctly and capture the Lagrangian-Eulerian interface. While the loss of data information is observed
in the training data, it is assumed that the loss is unchanged and can be used as correction during the
predictions. The proposed snapshot-FTLR technique can be divided into five key steps, which are as
follows:
1. Field transfer: The first step involves the field transfer from a highly unstructured moving point
cloud (x(t), y(t)) to a reference uniform grid (X, Y). The size of the grid can be chosen Nx × Ny .
We use Scipy’s griddata function [1] to interpolate the scattered CFD data and fit a surface. This
function generates the interpolant using Clough-Tocher scheme [2] by triangulating the scattered
data s(x(t), y(t),t) with Quickhull algorithm and forming a piecewise cubic Bezier interpolating
polynomial on each triangle. The gradients of the interpolant are chosen such that the curvature of
the interpolated surface is approximately minimized [57, 64]. After achieving this minimization,
the field values at the static Eulerian probes Nx × Ny are generated as S(X, Y,t).
2. Level-set: With the aid of learned interface description, a level Φ is assigned for all the cells on the
uniform grid. The exact interface description is known for a stationary solid boundary, which is
then used for synchronous structural motion prediction via the POD-RNN framework. The process
is further demonstrated in section 3.1. This step provides the identification of the solid, fluid and
interface cells on the DL-ROM grid.
14
Fluid cells
Δy (𝑖, 𝑗 + 1)
Interface cells
Δx
1 c 2 Solid cells
(𝑖 − 1, 𝑗) b (𝑖, 𝑗) a (𝑖 + 1, 𝑗)
Face mid-point
4 d 3
𝐒≠0
(𝑖, 𝑗 − 1)
𝐒=0
Figure 5: Identification of interface cells in the DL-ROM grid and finite difference interpolation of field at the cell faces (green
cross). The blue nodes contain the flow field values and the red dots are marked as zero.
3. Pixelated force1 : The next step is the calculation of the forces exerted by the interface cells that
contain the rigid body description. We refer to these forces as pixelated forces. The method of
measuring the pixelated force incorporates the construction of the Cauchy stress tensor in the inter-
n
face pixels. As shown in Fig. 5, for an interface cell k at time instant n, the pixelated force fk for a
Newtonian fluid can be written as
n n n n n
fk = (σa;k − σb;k ).nx ∆y + (σc;k − σd;k ).ny ∆x, (23)
n is the Cauchy stress tensor for any point inside the cell k. As depicted in Fig. 5, we
where σ∗;k
calculate this tensor at the mid-points of cell faces a − d − b − c via finite difference approximation.
nx and ny are normals in the x and y-direction, respectively, whereby ∆x and ∆y denote the cell
n
sizes. For the present case, we only consider the pressure component while calculating fk .
To obtain the total pixelated force signal, these individual pixelated forces are summed over all the
interface cells as
NF
n n
Fb = ∑ fk . (24)
k=1
1 2 n
where NF denotes the number of interface cells. These total pixelated force signals Fb = {Fb Fb ... Fb },
however, can be corrupted with missing or noisy information compared to its higher-order counter-
part. This is accounted due to the loss of sub-grid information that leads to a coarser calculation of
forces on the sharp interface. Thus, these output signals need to be corrected.
4. Damp noises: To correct the pixelated signals, we first apply a moving average Gaussian filter ζ
to smoothen the signals. This is achieved by damping the high-frequency noises in these signals
that are encountered due to a full-order interface movement on a low-resolution uniform grid. This
1 This section has been presented in our previous work from Bukka et al. in [9]. It is outlined here for the sake of completeness.
15
damping smoothens the force propagation for easier reconstruction. To smooth a pixelated force
n
signal Fb at a time step n, we perform the following operations:
(a) collect the time series high-frequency noisy signals Fb of window length 2 f + 1
n− f n−1 n n+1 n+ f
Fb = {Fb ... Fb Fb Fb ... Fb }, (25)
n
(b) select Gaussian-like weights of the specified window length with mean weight wn for Fb
n+ f
w = {wn− f ... wn−1 wn wn+1 ... wn+ f }, with ∑ wi = 1 (26)
i=n− f
n
(c) apply the weighted average filter to damp the noise at Fb
n+ f
n i n
Fb := ∑ wi Fb = ζ Fb . (27)
i=n− f
5. Interface force recovery: The smooth pixelated force propagation ζ Fb can still lack the mean
and derivative effects compared to its full-order counterpart FΓfs . This is where we introduce the
functional mapping Ψ as described in Algorithm 1. If this mapping can recover the bulk forces
correctly, then we can utilize the selected grid. If not, then refine the grid and repeat the above
steps from (1)-(5), until we find the best snapshot grid that recovers these bulk quantities with
acceptable accuracy.
Remark 1. The mesh-to-mesh field transfer and load recovery for the snapshot DL-ROM grid search are
expressed in a similar procedure as Algorithm 2. Once selected, the chosen uniform grid is utilized for
learning the flow variables in the CRAN framework with the POD-RNN operating on the moving interface.
Algorithm 3 essentially employs the aforementioned interface load recovery to extract the higher-order
forces from the synchronized field and interface predictions.
16
Algorithm 1: Functional reconstruction mapping Ψ for a higher-order force recovery on the DL-ROM grid.
Algorithm 1
Input: De-noised pixelated force signals Fb , full-order FSI forces FΓfs

Output: Reconstructed pixelated force signal ΨFb
To reconstruct smooth pixelated force signals Fb :

1. Collect n time steps of denoised pixelated force and full-order FSI forces:
1 2 n−1 n
Fb = {Fb Fb ... Fb Fb };
1 2 n−1 n
FΓfs = {FΓfs FΓfs ... FΓfs FΓfs };
2. Get the mean and fluctuating force components:
0
Fb = Fb − mean(Fb );
0
FΓfs = FΓfs − mean(FΓfs );
3. Define the time-dependent derivative error using E c :
0 0 0 0
E c = (FΓfs − Fb )./(Fb ) with Fb 6= 0;
4. Reconstruct the smooth pixelated force signals with mean and derivative corrections:
0 0
Fb := Fb + mean(FΓfs ) + mean(E c )Fb = ΨFb ;
Algorithm 2: Iterative snapshot DL-ROM grid search via interface force recovery.
Algorithm 2
1 2 n
Input: Full-order field data s = {s1 s2 ... sn } and FSI forces FΓfs = {FΓfs FΓfs ... FΓfs }
Output: Grid selection Nx × Ny
Initialise a uniform grid (Nx × Ny ) of cell size ∆x = ∆y:

while:
1. Project point cloud CFD data s(x(t), y(t)) ∈ Rm×n on snapshot grid S(X, Y):
S ← s s.t. S ∈ RNx ×Ny ×n ;
2. Apply learned level-set function Φ ∈ RNx ×Ny ×n on S(X, Y) element wise:
S := Φ ∗ S;
3. Calculate pixelated force signals Fb from interface cells Γfs in S using Eqs. (23) - (24);
4. Damp the high-frequency noises in pixelated signals Fb via moving average Gaussian filter
ζ:
Fb := ζ Fb using Eqs. (25)-(27);
5. Reconstruct the smooth pixelated force Fb to full-order by functional mapping Ψ:
Fb :≈ ΨFb using Algorithm 1;
6. if ((kFb − FΓfs k/kFΓfs k) ≤ ε ):
break;
else:
Refine the grid Nx := 2Nx , Ny := 2Ny ;
17
Algorithm 3: Extracting FSI forces via predicted fields on the DL-ROM grid and the ALE displacement.
Algorithm 3
n+1 n+2 n+p

Input: Field predictions from CRAN Ŝ = {Ŝ Ŝ ... Ŝ }, displacement predictions from
POD-RNN ŷ = {ŷn+1 ŷn+2 ... ŷn+p }, calculated interface mapping Ψ from training data
n+1 n+2 n+p
Output: Full-order force predictions: Fˆ = {Fˆ fs Fˆ fs ... Fˆ fs }
Γfs Γ Γ Γ
To extract full-order FSI forces:

1. Apply predicted level-set function Φ̂ ∈ RNx ×Ny ×p from POD-RNN on Ŝ ∈ RNx ×Ny ×p element
wise:
Ŝ := Φ ∗ Ŝ;
2. Calculate pixelated force signals Fˆ b from interface cells Γfs in Ŝ using Eqs. (23) - (24);
3. Damp high-frequency noises in extracted pixelated signals Fˆ via moving average gaussian
b
filter ζ :
Fˆ b := ζ Fˆ b using Eqs. (25)-(27);
4. Reconstruct the extracted smooth pixelated force Fˆ b to full-order by known functional map-
ping Ψ:
Fˆ Γfs ≈ ΨFˆ b
5. Application: Vortex-induced vibration (VIV)
This section demonstrates our hybrid partitioned DL-ROM methodology for a benchmark fluid-
structure interaction problem. While the proposed technique can be generalized to any fluid-structure
system involving the interaction dynamics of flexible structures with an unsteady wake-vortex system,
we consider an unsteady elastically-mounted cylinder problem in an external flow. The wake flow behind
the bluff body generates switching vortex loads on the solid, which causes its vibration. This structural
vibration, in turn, affects the near-wake flow dynamics resulting in a two-way synchronized oscillating
system. For this phenomenon of vortex-induced vibration, the flow physics is highly sensitive and directly
related to the interaction dynamics between the downstream vortex patterns and the cylinder’s motion. Of
particular interest is to learn and predict such synchronized wake-body interaction using data-driven com-
puting. We consider the parameter sets corresponding to a limit cycle response as well as non-stationary
vibration response.
The freely vibrating structural system is installed in a 2D computational domain. The center of the
cylinder is at sufficient distances from the boundaries to capture the downstream wake dynamics (see
Fig. 6 (a)). A no-slip and traction continuity condition is enforced on the cylinder surface (Eqs. (4)-
(5)), while a uniform velocity profile U∞ is maintained at the inlet Γin . The traction-free boundary is
implemented on the outlet Γout and slip condition is specified at the top Γtop and bottom Γbottom . We
discretize the fluid domain via unstructured finite element mesh. The final mesh, which is obtained after
following the standard rules of mesh convergence, is presented in Fig. 6 (b). It contains a total of 25, 916
quadrilateral elements with 26, 114 nodes. The full-order simulation is carried out via finite element
incompressible Navier-Stokes solver in the ALE framework. The output details are tabulated in Table 1
for validation. By constructing the Cauchy stress tensor on the fluid-solid surface, the fluid force along
18
Γ𝑡𝑜𝑝
Γ𝑖𝑛 Γ𝑜𝑢𝑡 30𝐷
15𝐷
Γ𝑏𝑜𝑡𝑡𝑜𝑚
4
10𝐷 20𝐷
(a)
(b)
Figure 6: Freely vibrating circular cylinder in a uniform flow: (a) Schematic of an elastically-mounted cylinder undergoing VIV,
(b) representative full-order domain and the near cylinder unstructured mesh.
19
Table 1: An elastically-mounted circular cylinder undergoing vortex-induced vibration: Comparison of the full-order forces and the
displacements in the present study and the benchmark data. CD and CL represent the drag and lift force coefficients, respectively.
Ax and Ay represent the x and y displacements of the cylinder, respectively.
∗ N ∗∗

Mesh Ncyl f luid mean(CD ) (CL )rms (Ax )rms Ay max
Present 124 25,916 2.0292 0.1124 0.0055 0.5452

Ref. [43] 168 25,988 2.0645 0.0901 0.0088 0.5548
∗N
cyl are the number of cylinder surface elements
∗∗ N
f luid are the number of Eulerian fluid elements
the interface is computed. The drag and lift force coefficients are calculated as
1
Z
CD = 1 f 2
(σ f .n).nx dΓ,
2 ρ U ∞ D Γfs
(28)
1
Z
CL = 1 f 2 (σ f .n).ny dΓ,
2 ρ U ∞ D Γfs
where CD and CL denote the drag and lift force coefficients, respectively, and Γfs is the fluid-solid inter-
face. nx and ny are the Cartesian components of the unit normal vector n.
5.1. Limit cycle VIV response

We start by applying our proposed methodology for the VIV at frequency lock-in. A rigid cylinder
is mounted in the transverse and streamwise directions with linear and homogeneous springs and no
damping. This results in alike natural frequencies fnx = fny in both Cartesian directions. Computation is
carried at a fixed cylinder mass ratio m∗ = πρ4m U∞
f D2 = 10, reduced velocity Ur = f n D = 1
U
q∞
k
= 5, and
2π mD
ρ fU∞ D
Reynolds number Re = = 200. m is the solid mass, fn is the natural frequency of the cylinder and
µf
U∞ is the uniform inlet x-velocity. The remaining other constants imply their usual meaning. At these
non-dimensional parameters, a high and sensitive limit cycle oscillations are obtained [43].
The full-order simulation is carried out for tU∞ /D = 250 with a time step of 0.025 tU∞ /D. From
these full-order data, ntr = 5000 snapshots (100-225 tU∞ /D) are used for training and nts = 1000 (225-
250 tU∞ /D) are kept for testing. Following the data generation, we apply our multi-level framework and
the snapshot-FTLR method for predicting the flow fields, the ALE displacements, and the pressure force
coefficients.
5.1.1. Structural displacements via POD-RNN driver

Time instances of the ALE x-displacements or y-displacements Y = y1 y2 . . . yN ∈ Rm×N are de-

0
composed using the POD projection. Here, m = 26, 114 and N = 6000. The fluctuation matrix Y ∈ Rm×N
and the offline database of the mean field ȳ ∈ Rm and the POD modes ν ∈ Rm×N are obtained. Eigenval-
0 0
ues ΛN×N of the covariance matrix Y T Y ∈ RN×N are extracted as their magnitude measures the energy
of the respective POD modes. Cumulative and percentage of the modal energies are plotted in Figs. 7
and 8 with respect to the number of modes. It can be observed that the system energy of > 99.99% is
concentrated in the first 1 − 2 POD modes. This observation implies that the ALE field can potentially be
reduced with just k = 2 POD modes instead of m = 26, 114 fluids nodes. One reason for this observation
is that the majority of the variance about ȳ is concentrated at the moving interface and propagates linearly
to the far-field boundaries. Interestingly, with these two modes, the full-order mesh can be reconstructed
within a maximum absolute error < 1.0 × 10−15 . Fig. 9 shows the dynamical evolution of the dominant
20
(a) (b)
Figure 7: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 5, Re = 200): (a) Cumulative and (b) percentage of
the modal energies for the ALE x-displacements Yx .
(a) (b)
Figure 8: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 5, Re = 200): (a) Cumulative and (b) percentage of
the modal energies for the ALE y-displacements Yy .
21
Table 2: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 5, Re = 200): Network and hyperparameters details of
the closed-loop RNN.
ALE k Cell Nh∗ Optimizer α∗ Epochs Learning Decay step

rate decay
x or y 2 LSTM 256 Adam∗ 0.001 1600 0.2 125 epochs
Nh : LSTM hidden cell dimension
α : Initial learning rate
Adam∗ : Adaptive moment optimization [39]
0
POD modes ν ∈ Rm×k using Aν = ν T Y ∈ Rk×N . The time coefficients are nearly periodic with a limit
cycle oscillation as inherent in the high-dimensional information.
The N temporal coefficients are divided into the training and the testing part. We train a closed-
loop LSTM in a time-delayed fashion for learning these time coefficients. This training implies that the
predicted output is used as an input for the next time step. The input to the closed-loop recurrent neural
network is a vector of dimension k so that all k modes are learned and predicted in one time instant.
The nearly periodic trend in the modes allows an easier hyperparameter construction of the LSTM-RNN,
albeit different phases and amplitudes of the modes exist. Further details about the network can be found
in Table 2. The training took around 5 minutes on a single graphics processing unit, and the prediction
of the time coefficients Âν is depicted in Fig. 10 for testing. We keep the multi-step prediction cycle to
p = 100 while testing the output. This implies that one test time step is used to predict the next 100 time
coefficients until a ground input is fed. It is worth mentioning that an encoder-decoder LSTM architecture
would have been preferred to extract the temporal coefficients in the case of chaotic dynamics. In the
present case of periodic oscillations, a simple closed-loop LSTM-RNN is sufficient for the point cloud
tracking.
i
The predicted modal coefficients at any time step i, Âν ∈ Rk , can simply be reconstructed back to
the point cloud ŷi ∈ Rm using the mean field ȳ ∈ Rm and the k spatial POD modes ν ∈ Rm×k . Fig. 11
compares the predicted and true values of the VIV motion in the x and y-directions of the cylinder. The
results indicate that the POD-RNN can accurately predict the position of the moving FSI interface for a
limit cycle oscillation. These results are inferred in a closed-loop fashion by predicting p = 100 steps
from one demonstrator at a time. The incorporation of the ground data demonstrators in the predictions
can circumvent the compounding effect of errors and thereby boost the long term predictive abilities of
the neural networks [73]. This description completes the moving point cloud prediction for the chosen
parameter set.
5.1.2. Snapshot-FTLR
As mentioned earlier, the POD-RNN driver tracks the moving interface and extracts the Φ level-set
feature. However, the best snapshot DL-ROM grid (Nx × Ny ) must be selected for the field predictions
in the CRAN framework. The snapshot-FTLR estimates the point cloud field data on a uniform grid
where a higher-order load recovery is possible at the practical level. We project the point cloud flow
field data, for instance the pressure fields, s = {s1 s2 ... sN } ∈ Rm×N as spatially uniform snapshots S =
{S1 S2 ... SN } ∈ RNx ×Ny ×N . This uniformity is achieved via the SciPy’s griddata function [1] by mapping
the m dimensional unstructured data on a 2D uniform grid. Here, Nx = Ny = 32, 64, 128, 256, 512, 1024.
Note that the DL-ROM grid for this case is of the size 8D × 8D with ≈ 5D length for the downstream
cylinder wake. To assess the sensitivity of grid distribution, we compare the accuracy of surface fit
provided by the grid data using nearest, linear and cubic interpolations. This is achieved by sampling the
field’s maximum and the minimum values with pixel numbers on the fitted surface process. As seen from
Fig. 12, for any pressure time step tU∞ /D = 125, the nearest levels-off to the true pmax and pmin during
22
(a)
(b)
Figure 9: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 5, Re = 200): Time history of the k = 2 modal
coefficients shown from 100 till 150 tU∞ /D for the (a) ALE x-displacements Yx and the (b) ALE y-displacements Yy .
23
(a)
(b)
Figure 10: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 5, Re = 200): Prediction of the modal coefficients on
the test data for the (a) ALE x-displacements Yx and the (b) ALE y-displacements Yy .
24
(a) (b)
Figure 11: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 5, Re = 200): Predicted and actual (a) x-position
and (b) y-position of the interface normalised by the diameter of cylinder D. The root mean squared error between the true and
predicted is 2.17 × 10−4 and 1.08 × 10−2 for Ax /D and Ay /D, respectively.
grid refinement. This reason is that this method assigns the value of the nearest neighbor in the distributed
triangulation-based information.
Remark 2. The linear and cubic interpolation techniques involve triangulations with C0 and C1 conti-
nuity. The linear method linearly converges to the true values on grid refinement. The cubic interpolation
approach, however, converges well on a coarser grid Nx = Ny = 64, 128, but deviates on subsequent grid
refinement Nx = Ny = 256, 512, 1024 with small relative errors.
Fig. 13 compares the interpolation methods for the pressure field with respect to the full-order for
tU∞ /D = 125 on the 512 × 512 DL pixelated grid. It can be observed that the nearest contains oscillations
compared to the full-order description due to a discontinuous assignment of the field at the selected
probes. A closer look in Fig. 13 (d) depicts some local fitting deviations compared to the full-order for
the cubic due to a higher-order interpolation. Conversely, the linear and the nearest look much closer
to the full-order results. Hence, we rely on the linear technique for a coarse-grain field transfer in the
remainder of our analysis. Fig. 14 demonstrates the grid dependence of the normalized pressure pixelated
forces Fb /0.5ρ f U∞2 D vs the full-order forces FΓfs /0.5ρ f U∞2 D. We observe that the interface coarsening
leads to the pixelated force propagation to contain some noises, especially in the direction where the solid
displacements are large. These noisy signals become dominant on reducing the time step and increasing
the cell size. The reasons are attributed to the loss of the interface information due to lower fidelity and a
linear way of the pixelated force calculation. Interestingly, these noises are somewhat invariant to a sharp
fluid-solid interface propagation on a uniform grid. These force signals are devoid of any high-frequency
noises for the static boundaries as discussed in our previous work [9].
To correct these forces, while still maintaining a lower interface grid resolution, we first apply the
Gaussian filtering ζ on the pixelated force propagation to damp the high-frequency noises. Fig. 15 (a)-(b)
depict the smooth trend ζ Fb in the normalized forces on the various DL-ROM grids for the drag and lift
coefficients via a Gaussian filter length of 20. It is worth noting that the pixelation process results in the
25
(a) (b)
(c) (d)
Figure 12: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 5, Re = 200): Convergence of the pressure field with
the number of pixels for different interpolation schemes at the time step tU∞ /D = 125. (a)-(b) Maximum pressure values and the
respective errors, (c)-(d) minimum pressure values and the respective errors.
26
Full-order grid
(a) (b)
(c) (d)
Figure 13: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 5, Re = 200): Qualitative behavior of the different
interpolation methods for the NS pressure field in the ALE reference at the time step tU∞ /D = 125: (b) Nearest neighbor, (c) linear
and (d) piece-wise cubic type values on the snapshot 512 × 512 uniform grid with respect to (a) full-order CFD grid.
27
(a) (b)
Figure 14: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 5, Re = 200): Total pixelated force propagation (from
100-125 tU∞ /D) on the snapshot DL-ROM grids vs the full-order for (a) the pressure drag coefficient CD,p , and (b) the pressure lift
coefficient CL,p on the training field.
mean and derivative variation in the bulk quantities from the full-order counterpart. These errors reduce
with super-resolution. However, we are interested in refining the uniform grid to an extent where the bulk
quantities can be linearly reconstructed to the full-order within the mean and derivative correction using
Algorithm 1.
The pixelated force and the reconstruction process Ψ on the different resolutions of the DL-ROM
grid is summarized in Table 3. Columns 3 and 4 tabulate the mean and the root mean square of the total
pixelated drag and lift coefficients, respectively, for various grid sizes. Column 5 in Table 3 similarly
denote the mean of time-dependent derivative correction E c observed from the Algorithm 1. Columns 6
and 7 depict the reconstruction error ε in the corrected forces ΨFb and the full-order FΓfs as calculated
from step 6 of Algorithm 2. The reconstruction accuracy is nearly 99.8% and 96.5% for the drag and
the lift, respectively, onward grid 512 × 512. This process facilitates an optimal uniform grid to carry the
neural prediction on while still recovering the bulk quantities within reasonable accuracy.
Remark 3. It is worth mentioning that the propagation of the pixelated force depends on the problem
at hand; however, the present reconstruction offers a general data recovery methodology. Fig. 15 (c)-
(d) depict the force correction by observing the Ψ mapping and correcting on the training forces. We
select the 512 × 512 as the snapshot DL-ROM grid for the flow field predictions. It accounts for a
reasonable force recovery while avoiding the necessity of super-resolution at the interface and bypassing
the unstructured mesh complexity of the moving point cloud.
5.1.3. Near-wake dynamics using the CRAN driver

With the chosen grid Nx = Ny = 512, the end-to-end nonlinear learning based on the CRAN is used
for the flow field predictions. Algorithm 3 can be employed to recover
the bulk quantities from the field
data. As described in section 5.1.2, a point cloud flow dataset s = s1 s2 . . . sN ∈ Rm×N from the full-
order solver is transferred to the uniform grid using the snapshot-FTLR method. The generated field
information S = S1 S2 . . . SN ∈ RNx ×Ny ×N are decomposed into ntr = 5000 training data (100-225

28
(a) (b)
(c) (d)
Figure 15: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 5, Re = 200): Interface load behavior vs time (from
100-125 tU∞ /D) on the snapshot DL-ROM grids. (a)-(b) denote the smoothen force propagation and (c)-(d) depicts recovered
interface force information for drag and lift coefficients.
29
Table 3: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 5, Re = 200): Pixelated forces and the reconstruction
accuracy on the various snapshot DL-ROM grids with respect to the full-order grid.
Grid Cell-size mean(ζ CD,p ) (ζ CL,p )rms mean(E c ) ε(CD,p ) ε(CL,p )
128 0.0625 1.8295 0.1491 0.8534 7.6 ×10−3 3.2 ×10−1

256 0.0313 1.7470 0.1212 0.9124 3.3 ×10−3 1.5 ×10−1
512 0.0157 1.7130 0.1180 0.9307 2.9 ×10−3 4.7 ×10−2
1024 0.0078 1.6979 0.1168 0.9545 2.8 ×10−3 2.8 ×10−2
FOM 2.2 ×10−5 1.6834 0.1174 – – –
Table 4: Freely vibrating circular cylinder in a uniform flowr (m∗ = 10,Ur = 5, Re = 200): Deep CRAN layer details: Convolutional
kernel sizes, numbers and stride. Layers 6,7,8 represent the fully connected feed-forward encoding, while 9,10,11 depict the similar
decoding. The low-dimensional state Ac is evolved between the layers 8 and 9 with the LSTM-RNN.
Conv2D encoder DeConv2D decoder

Layer kernel-size kernels stride Layer kernel-size kernels stride
1 10 × 10 2 4 12 5×5 16 2
2 10 × 10 4 4 13 5×5 8 2
3 5×5 8 2 14 5×5 4 2
4 5×5 16 2 15 10 × 10 2 4
5 5×5 32 2 16 10 × 10 1 4
tU∞ /D) and nts = 1000 testing data (225-250 tU∞ /D). The training time steps ntr are further sparsed
for every 0.05 tU∞ /D, thereby reducing the total number of trainable steps to ntr /2. This is essentially
carried out to speed-up the CRAN I/O data processing, while still maintaining the same time horizon of
training. As detailed in section 3.2 from Eqs. (19)-(21), the standard principles ofdata normalization and
batch-wise arrangement is adopted to generate the scaled featured input S = Ss01 Ss02 . . . Ss0Ns ∈
[0, 1]Nx ×Ny ×Nt ×Ns , where Ns = 100 and Nt = 25. While training,
h a mini-batch i of size ns = 1 is maintained.
For every training iteration, a training sample Ss0 j = S01 S 02 . . . SNt is randomly shuffled from the
s, j s, j s, j
scaled featured input to update and refresh the CRAN neural parameters.
The encoding and decoding space of the CRAN is summarized in Table 4 in a way that the spatial
downsampling and prolongation are achieved gradually. We keep the low-dimensional encoder state
Ac and the LSTM-RNN evolver cells of the sizes: h = Nh = 32, 64, 128 (see Fig. 3). These modes
are obtained from a nonlinear space of the autoencoder network and are found to be the most sensitive
hyperparameter in this architecture. Since these projections are self-supervised, we experiment with the
different sizes of the modes based on the convergence during training. We first instantiate the training of
all the three CRAN models for Ntrain = 500, 000 iterations starting from the pressure fields. We select the
Nh = 128 based on a faster decay in the overall loss (Eq. (22)) compared with the Nh = 32, 64 evolver
states. Nh = 128 pressure model is further optimized until the total iterations reach to Ntrain = 106 with
the overall loss ≈ 4.93 × 10−5 . Once the pressure training is completed, we transfer the learning to the
x-velocity field and optimize the CRAN network subsequently for Ntrain = 750, 000 iterations.
Remark 4. Each CRAN model is trained on a single graphics processing unit Quadro GP100/PCIe/SSE2
with Intel Xeon(R) Gold 6136 central processing unit @ 3.00GHz × 24 processors for nearly 2.7 days.
A long training time is considered so that the CRAN-based framework self-supervises with the maximum
optimization. The training process is stopped when the loss reaches ≈ 1.0 × 10−5 . The test time steps
depict the predictive performance of this network in terms of accuracy check and remarkable speed-ups
30
compared with the FOM counterpart.
Herein, we are interested to demonstrate Nh = 128 trained CRAN model to infer the flow field with
a multi-step predictive cycle of p = Nt = 25. Figs. 17 and 18 show the comparison of the predicted and
the true values of the pressure and x-velocity fields, respectively, at time steps 9300 (232.5 tU∞ /D), 9600
(240 tU∞ /D) and 9960 (249 tU∞ /D) on the test data. The normalized reconstruction error E i , for any time
step say i, is constructed by taking the absolute value of the difference between the true Si and predicted
field Ŝi and normalizing the difference with the L2 norm of the truth using the equation
|Si − Ŝi |
Ei = . (29)
kSi k2
It is evident from these predictions that significant portions of the reconstruction errors are concentrated
in the nonlinear flow separation region of the vibrating cylinder. These errors are in the order of 10−3 for
the pressure and 10−4 for the x-velocity. The field predictions are accurately predicted in the expected
cylinder motion, albeit with a minor deviation in the reconstruction. We observe that for the CRAN
framework, with one input step, the coupled FSI prediction is accurate for the next 25 time steps in the
future, after which the predictions can diverge. Thus, the feedback demonstrators combat the compound-
ing errors and enforce the CRAN predictive trajectory devoid of divergence over the test data.
The synchronous point cloud structural and coarse-grain flow field predictions are employed to calcu-
late the integrated pressure loads CD,p and CL,p via Algorithm 3. The results are shown in Fig. 16 (in the
red line). The interface load algorithm extracts the full-order pressure loads with consistent phases from
the predicted fields on the uniform grid, largely accounting for an accurate flow separation and the near-
wake predictions. Interestingly, we observe some variations in the amplitude of the force signals from a
direct coarse-grain integration over the moving interface even though other quantities are predicted well.
The reason for this behavior is the existence of high-frequency noises in the pixelated forces from the
uniform grid, even in the training data. Since the internal training process of the CRAN does not damp
these force noises directly, the predicted fields can also infer these noises in time. We note that these
spurious noises are the largest when the interface displacement is high and as a result, large residuals
are seen in the force plots. The development of these residuals in the CNN-based predictions can also
be responsible for a finite amount of predictive ability of the CRAN architecture and the requirement of
ground data to combat this problem.
Remark 5. One potential way to correct these residuals is to dampen the noises using a denoising
recurrent neural network. These networks have been shown to perform well to denoise signals in a
variety of applications such as electrocardiographic signal (ECG) [4] and automatic speech recognition
(ASR) [46]. For demonstration, we damp the noises in the coarse-grain force signals using a denoising
LSTM-RNN. We achieve this by learning the mapping of the coarse-grained force signals to the full-order
force using a standard LSTM type recurrent network on finite predicted signals. The filtered signals
in Fig. 16 (shown in the blue line) demonstrate good precision for the drag and reasonable in the lift
compared to the full-order. Since denoising RNNs are beyond the scope of this manuscript, we refer the
reader to the works in [29] and [55] for more details.
5.2. Non-stationary VIV response

In section 5.1, we discussed the application of our hybrid partitioned DL methodology for a VIV
response at a mass ratio of m∗ = 10, reduced velocity of Ur = 5 and Reynolds number of Re = 200.
Although a high and sensitive VIV response is obtained at these parameters, the full-order VIV motion
is a simple single frequency limit cycle oscillation. Hence, to further strengthen the application of our
methodology, we now apply the partitioned DL-ROM framework for a Ur = 7 while keeping the same
31
(a) (b)
Figure 16: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 5, Re = 200): Interface force prediction on the test
time steps from the predictive framework (from 225-250 tU∞ : (a) Pressure drag coefficient CD,p , and (b) pressure lift coefficient
CL,p . The red lines depict the coarse-grain forces obtained from Algorithm 3 and blue lines depict the corrected forces using a
denoising LSTM-RNN.
Reynolds number and the mass ratio. For the considered training and testing dataset, the full-order VIV
motion consists of small magnitudes of non-stationary amplitude responses with no limit cycle oscillation
at these new parameters.
The full-order time series dataset is similarly generated until tU∞ /D = 250. A total of 2000 snapshots
are generated at every 0.125 tU∞ /D for the pressure field and the ALE displacements. We keep the same
time duration of the training and testing range but with a reduced number of time steps: ntr = 1000
snapshots (100-225 tU∞ /D) for training and nts = 200 (225-250 tU∞ /D) for testing. At these set of
training and testing range, we note that there is no saturation amplitude in the dataset, which makes it
a good test case to replicate the complexities involved in the VIV motion that are not necessarily single
frequency limit cycle response.
1 As2 discussed in section 5.1.1, time instances of the ALE x-displacements or y-displacements Y =
y y . . . yN ∈ Rm×N for this full-order dataset are decomposed using the POD projection in order to
reduce the point cloud information. Here, m = 26, 114 and N = 1200. Cumulative modal energies for the
ALE x-displacements and y-displacements are plotted in Fig. 19 (a) and (c), respectively, with respect to
the number of modes. For the chosen parameters, it is interesting to observe that the POD spectrum for
the ALE displacements are still fast decaying (similar to the limit cycle VIV oscillation in Figs. 7 and 8)
and the majority of the system energy of > 99.9% can be described by the first 2 POD modes. This helps
in reducing the ALE field with k = 2 POD modes instead of m = 26, 114 fluid nodes for this test case.
Notably, with these two modes, the full-order mesh displacement can be reconstructed within a maximum
absolute error < 1.0 × 10−14 .
Fig. 19 (b) and (d) depict the dynamical evolution of the 2 POD modes for the ALE x-displacements
and y-displacements, respectively, obtained from the POD analysis. As inherent in the high-dimensional
VIV motion, we note that these time coefficients propagate as non-stationary variable amplitude responses
with small frequency differences in every oscillation. The point cloud tracking, hence, reduces to effective
learning and prediction of these dynamical coefficients accurately in time.
32
(a)
(b)
(c)
Figure 17: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 5, Re = 200): Comparison of the predicted and the
true fields along with the normalized reconstruction error E i at tU∞ /D = (a) 232.5, (b) 240 (c) 249 for the pressure field (P).
33
(a)
(b)
(c)
Figure 18: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 5, Re = 200): Comparison of the predicted and the
true fields along with the normalized reconstruction error E i at tU∞ /D = (a) 232.5, (b) 240 (c) 249 for the x-velocity field (U).
34
(a) (b)
(c) (d)
Figure 19: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 7, Re = 200): Cumulative modal energies and the time history of k = 2
modal coefficients shown from 100 till 150 tU∞ /D. (a)-(b) for the ALE x-displacements Yx , and (c)-(d) for the ALE y-displacements Yy .
35
(a) (b)
(c) (d)
Figure 20: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 7, Re = 200): Prediction of the modal coefficients and the interface
motion on the test data. (a) and (c) denote the predicted and actual time coefficients of the ALE x-displacements Yx and ALE y-displacements Yy ,
respectively. (b) and (d) denote the corresponding x and y position of the interface, normalised by diameter of the cylinder D. With p = 25, the root
mean squared error (RMSE) between the true and predicted is 5.58 × 10−4 and 2.22 × 10−3 for Ax /D and Ay /D, respectively. With p = 200, the
RMSE between the true and predicted is 9.11 × 10−4 and 4.44 × 10−3 for Ax /D and Ay /D, respectively.
36
(a) (b)
Figure 21: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 7, Re = 200): Recovered interface load behavior vs time shown from
100-125 tU∞ /D using the snapshot-FTLR for (a) the drag, and (b) the lift coefficients on the 512 × 512 snapshot DL-ROM grid with respect to
full-order.
Using a closed-loop LSTM-RNN with the same hyperparameters as in Table 2, we learn the POD time
coefficients over the 1000 training time steps by optimising the evolver loss (Eq. 15) as detailed in section
3.1. With the training of 5 minutes on a graphics processing unit, the prediction of the modal coefficients
is depicted in Fig. 20 (a) and (c) for the ALE x-displacements and y-displacements, respectively, over the
test data. We keep the multi-step prediction cycle of length p = 25 (one input infers 25 test steps) and
p = 200 (one input infers all 200 test steps). We note that the predictions of these non-stationary time
coefficients are in good agreement with the ground truth with slight improvements in p = 25 predictions
as compared to p = 200. The predicted modal coefficients are reconstructed back to the point cloud using
the offline database of the mean field and the POD modes to track the interface. Fig. 20 (b) and (d) depict
the comparison of the predicted and true values of the VIV motion in the x and y directions of the cylinder,
respectively. The results indicate that the POD-RNN on the ALE motion can predict the position of the
moving FSI interface with excellent accuracy for this parameter set.
The application of the snapshot-FTLR as a load recovery for this parameter set is showcased in Fig. 21
over the training forces. For the sake of uniformity, we have selected the same resolution of the snapshot
DL-ROM grid 512 × 512 obtained via linear interpolation. Via Gaussian filtering of length 5 and the
functional reconstruction mapping Ψ (using Algorithm 1), the reconstructed force signals obtained show
a good agreement for the drag and lift coefficients with the full-order values. As a result, we transfer
the field point cloud to uniform and structured DL snapshots for the CRAN driver. For demonstration
purposes, we only consider the pressure fields for testing the predictive abilities of the CRAN driver.
The generated pressure field information S = S1 S2 . . . SN ∈ RNx ×Ny ×N using the snapshot-FTLR
are decomposed into ntr = 1000 training data (100-225 tU∞ /D) and nts = 200 testing data (225-250
tU∞ /D). As detailed in section 3.2, the training flow dataset is normalized and batch-wise arranged to
generate the scaled featured input with number of batches Ns = 40 and the RNN time sequence Nt = 25.
To train the CRAN architecture on this flow dataset, we load the optimized Nh = 128 CRAN parameters
for the pressure model for the limit cycle case in section 5.1 and optimize further. This transfer of learning
helps in optimizing the CRAN framework for the new parameter set in Ntrain = 50, 000 iterations, which
took a training time of nearly 7 hours on single graphics processing unit with the final loss of ≈ 1.0×10−5 .
The test time steps depict the predictive performance of this network in terms of accuracy check.
37
(a)
(b)
Figure 22: Freely vibrating circular cylinder in a uniform flow (m∗ = 10,Ur = 7, Re = 200): (a) Comparison of the predicted and the
true fields along with the normalized reconstruction error E i at tU∞ /D = 250 for the pressure field (P), and (b) interface drag and
lift prediction on the test time steps from the predictive framework with multi-step predictive cycle p = 25. The CRAN predictions
are accurate for the first 25-30 time steps in the future, after which the predictions can diverge.
38
Table 5: Freely vibrating circular cylinder in a uniform flow: Summary of the training and testing times of the DL-ROM compo-
nents along with force computation as compared to the full-order computational time. The testing of the POD-RNN, CRAN and
Algorithm 3 denote the run time needed to predict p = 25 instances of displacements, fields and forces, respectively. The combined
predictions denote the total run time for all these predictions.
Parameter Training∗ Training∗ Testing∗ Testing∗ Algo. 3∗∗ Combined FOM∗∗ Speed-up
(Re = 200, (POD- (CRAN) (POD- (CRAN) (Forces) predictions predictions factor
m∗ = 10) RNN) RNN)
(hours) (hours) (seconds) (seconds) (seconds) (seconds) (seconds)
Ur = 5 1.5×10−1 6.0 × 101 2.8×10−1 6.4×10−1 8.9×10−1 1.8 4.3 × 102 2.4 × 102
Ur = 7 1.5×10−1 7.5 3.1×10−1 6.5×10−1 6.9×10−1 1.7 4.3 × 102 2.5 × 102
∗ Single GPU processor
∗∗ Single CPU processor
Fig. 22 (a) depict the comparison of the predicted and the true values of the pressure field at the
last test time step 250 tU∞ /D obtained with a multi-step predictive cycle of p = Nt = 25. From these
predictions, we note that that the field predictions are in excellent agreement with the true field, with
a reconstruction error E i in the order of 10−3 . Similar to the limit cycle VIV oscillation, we note the
coupled FSI prediction from the CRAN is accurate for the first 25-30 time steps in the future, after which
the predictions can diverge. The feedback demonstrators combat the compounding errors and enforce
the CRAN predictive trajectory devoid of divergence over the test data. Figs. 22 (b) depict the evolution
of the integrated pressure forces CD,p and CL,p on the interface over the test data using Algorithm 3.
From a direct integration of the predicted fields, we obtain a reasonable drag and lift prediction. The
variation in the predicted amplitude from the force signals are minor for this case due to a smaller interface
displacement in this coupled dataset. As a result, the noises in the forces are of a smaller magnitude.
Finally, we provide an estimate of computational cost for the hybrid partitioned DL-ROM together
with full-order simulations. In Table 5, we report the training times, the testing times and the total time
needed to predict the solid displacements, the flow predictions and the forces (using Algorithm 3). We
recall that the multi-level DL-ROM solution can offer remarkable speed-ups in computing as compared
with the FOM, by nearly 250 times. While a canonical laminar flow of vibrating cylinder is considered,
the proposed hybrid DL-ROM does not make any assumptions with regard to geometry and boundary
conditions.
6. Conclusions
We have presented a hybrid partitioned deep learning framework for the reduced-order modeling of
moving interfaces and fluid-structure interaction. The proposed novel DL-ROM relies on the proper
orthogonal decomposition combined with recurrent neural networks and convolutional recurrent autoen-
coder network. We decouple structural and fluid flow representations for easier and modular learning of
two physical fields independently. While POD-RNN provides an accurate extraction of the fluid-structure
interface, the CRAN enables extraction of the fluid flow fields. We have successfully demonstrated the
inference capability of the proposed DL-ROM framework by predicting the time series of the unsteady
pressure field of an FSI set-up of an elastically-mounted circular cylinder. Using coarse to fine-grained
learning of fluid-structure interaction, a low-dimensional inference of the flow fields and the full-order
mesh description with load recovery has been discussed.
By analyzing a prototype FSI model for limit cycle and non-stationary structural motions, we have
first shown that the POD-RNN component infers the point cloud dynamics by projecting the ALE dis-
placements on the POD modes. A closed-loop LSTM-RNN effectively propagates the evolution of the
POD time coefficients and tracks the full-order grid motion accurately. We then analyze an iterative low
interface resolution DL-ROM grid search for the CNNs that preserves the full-order pressure stresses on
39
the moving fluid-structure interface via snapshot-FTLR. We have shown that this snapshot-FTLR method
selects the grid for the CRAN architecture but can exhibit spurious oscillations and data loss in the loads
on a moving FSI system. These errors are shown to depend on the uniform grid resolution, and a detailed
study has been performed to recover and filter the losses. A 16-layered trained CRAN network is shown
to predict the flow dynamics involving a coupled FSI with accurate estimates of flow prediction. The
CRAN network extrapolates the field accurately for 25-30 time steps from one input data, after which
the predictions diverge owing to some noises in the CRAN reconstruction. This leads to instability in
the force calculation on the moving interface. Once trained, the proposed DL-ROM technique can be
implemented using only a fraction of computational resources to that of a full-order model for online
predictions.
Since the proposed hybrid DL-ROM framework is our first attempt for a moving fluid-structure in-
terface, there are several possible extensions of the proposed approach. For example, the POD-RNN has
been applied for the rigid body equations involving a simplified VIV motion. One could extend our frame-
work to flexible structures with geometric and/or material nonlinearity via the CRAN driver. Moreover,
other neural architectures such as graph neural networks [60] and transformers [71] can be considered
within the proposed partitioned DL-ROM framework. Instead of POD, hyperreduction techniques such
as energy-conserving sampling and weighting [3] can be used for evolving nonlinear structural dynam-
ics. In our current work, the DL-ROM methodology models VIV in a 2D non-body conformal uniform
mesh. Although the snapshot-FTLR method allows a structured grid search for CNNs and preserves
interface description, it can create a larger number of unknowns in the DL space if refined elsewhere.
In future work, it is worth considering a 3D extension of the DL-ROM methodology and an adaptive
refinement/coarsening of the non-interface cells in the DL-ROM grid to reduce the number of unknowns.
Acknowledgement
The authors would like to acknowledge the Natural Sciences and Engineering Research Council of
Canada (NSERC) for the funding. This research was supported in part through computational resources
and services provided by Advanced Research Computing at the University of British Columbia. We
would also like to thank the reviewers for their constructive comments that has helped in improving the
quality of the manuscript substantially.
References
[1] Scipy reference guide, release 1.4.1. https: // docs. scipy. org/ doc/ scipy-1. 4. 1/
scipy-ref-1. 4. 1. pdf , 2019.
[2] P. Alfeld. A trivariate clough—tocher scheme for tetrahedral data. Computer Aided Geometric
Design, 1(2):169–181, 1984.
[3] S. S. An, T. Kim, and D. L. James. Optimizing cubature for efficient integration of subspace defor-
mations. ACM Transactions on Graphics, 165, 2008.
[4] K. Antczak. Deep recurrent neural networks for ecg signal denoising. arXiv preprint
arXiv:1807.11551, 2018.
[5] G. Berkooz, P. Holmes, and J. L. Lumley. The proper orthogonal decomposition in the analysis of
turbulent flows. Annual review of fluid mechanics, 25(1):539–575, 1993.
[6] M. M. Bronstein, J. Bruna, T. Cohen, and P. Veličković. Geometric deep learning: Grids, groups,
graphs, geodesics, and gauges. arXiv preprint arXiv:2104.13478, 2021.
40
[7] M. E. Brown, D. J. Lary, A. Vrieling, D. Stathakis, and H. Mussa. Neural networks as a tool for
constructing continuous ndvi time series from avhrr and modis. International Journal of Remote
Sensing, 29(24):7141–7158, 2008.
[8] S. R. Bukka. Data-driven computing for the stability analysis and prediction of fluid-structure
interaction. PhD thesis, 2019.
[9] S. R. Bukka, R. Gupta, A. R. Magee, and R. K. Jaiman. Assessment of unsteady flow predictions
using hybrid deep learning based reduced-order models. Physics of Fluids, 33(1):013601, 2021.
[10] S. R. Bukka, A. R. Magee, and R. K. Jaiman. Reduced order model for unsteady fluid flows via
recurrent neural networks. In ASME 2019 38th International Conference on Ocean, Offshore and
Arctic Engineering. American Society of Mechanical Engineers, 2019.
[11] J. Burkardt, M. Gunzburger, and H. Lee. Pod and cvt-based reduced-order modeling of navier–
stokes flows. Computer Methods in Applied Mechanics and Engineering, 196(1-3):337–355, 2006.
[12] B. Chang, M. Chen, E. Haber, and E. H. Chi. Antisymmetricrnn: A dynamical system view on
recurrent neural networks. arXiv preprint arXiv:1902.09689, 2019.
[13] S. Chaturantabut and D. C. Sorensen. Discrete empirical interpolation for nonlinear model reduc-
tion. SIAM Journal on Scientific Computing, 32:2737–2764, 2010.
[14] H. Chen, Y. Zhang, M. K. Kalra, F. Lin, Y. Chen, P. Liao, J. Zhou, and G. Wang. Low-dose ct with
a residual encoder-decoder convolutional neural network. IEEE transactions on medical imaging,
36(12):2524–2535, 2017.
[15] T. Chen and H. Chen. Universal approximation to nonlinear operators by neural networks with
arbitrary activation functions and its application to dynamical systems. IEEE Transactions on Neural
Networks, 6(4):911–917, 1995.
[16] W. Chen, Q. Wang, J. S. Hesthaven, and C. Zhang. Physics-informed machine learning for reduced-
order modeling of nonlinear problems. Journal of Computational Physics, page 110666, 2021.
[17] G. Cybenko. Approximation by superpositions of a sigmoidal function. Mathematics of Control,
Signals, and Systems, 2(4):303–314, 1989.
[18] A. Daw, R. Q. Thomas, C. C. Carey, J. S. Read, A. P. Appling, and A. Karpatne. Physics-guided
architecture (pga) of neural networks for quantifying uncertainty in lake temperature modeling. In
Proceedings of the 2020 SIAM International Conference on Data Mining, pages 532–540. SIAM,
2020.
[19] J. Donea. Arbitrary Lagrangian–Eulerian finite element methods. Computational Methods for Tran-
sient Analysis, pages 474–516, 1983.
[20] N. B. Erichson, M. Muehlebach, and M. W. Mahoney. Physics-informed autoencoders for lyapunov-

stable fluid flow prediction. arXiv preprint arXiv:1905.10866, 2019.
[21] N. Geneva and N. Zabaras. Modeling the dynamics of pde systems with physics-constrained deep
auto-regressive networks. Journal of Computational Physics, 403:109056, 2020.
[22] F. J. Gonzalez and M. Balajewicz. Deep convolutional recurrent autoencoders for learning low-
dimensional feature dynamics of fluid systems. arXiv preprint arXiv:1808.01346, 2018.
41
[23] Ian Goodfellow, Yoshua Bengio, Aaron Courville, and Yoshua Bengio. Deep learning, volume 1.
MIT press Cambridge, 2016.
[24] A. Goyal and Y. Bengio. Inductive biases for deep learning of higher-level cognition.
arxiv.org/pdf/2011.15091.pdf, 2021.
[25] R. Gupta, S. R. Bukka, and R. K. Jaiman. Assessment of hybrid data-driven models to predict
unsteady flows. In APS Division of Fluid Dynamics Meeting Abstracts, pages K09–017, 2020.
[26] P. S. Gurugubelli and R. K. Jaiman. Self-induced flapping dynamics of a flexible inverted foil in a
uniform flow. Journal of Fluid Mechanics, 781:657–694, 2015.
[27] Y. Ham, J. Kim, and J. Luo. Deep learning for multi-year enso forecasts. Nature, 573(7775):568–
572, 2019.
[28] J. Han, A. Jentzen, and E. Weinan. Solving high-dimensional partial differential equations using
deep learning. Proceedings of the National Academy of Sciences, 115(34):8505–8510, 2018.
[29] S. Han, Z Meng, X. Zhang, and Y. Yan. Hybrid deep recurrent neural networks for noise reduction
of mems-imu with static and dynamic conditions. Micromachines, 12(2):214, 2021.
[30] J. S. Hesthaven and S. Ubbiali. Non-intrusive reduced order modeling of nonlinear problems using
neural networks. Journal of Computational Physics, 363:55–78, 2018.
[31] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780,
1997.
[32] K. Hornik, M. Stinchcombe, and H. White. Universal approximation of an unknown mapping and
its derivatives using multilayer feedforward networks. Neural networks, 3(5):551–560, 1990.
[33] R. K. Jaiman, M. Z. Guan, and T. P. Miyanawala. Partitioned iterative and dynamic subgrid-scale
methods for freely vibrating square-section structures at subcritical reynolds number. Computers &
Fluids, 133:68–89, 2016.
[34] V. Joshi, P. S. Gurugubelli, Y. Z. Law, R. K. Jaiman, and P. F. B. Adaikalaraj. A 3d coupled
fluid-flexible multibody solver for offshore vessel-riser system. In International Conference on
Offshore Mechanics and Arctic Engineering, volume 51210, page V002T08A009. American Society
of Mechanical Engineers, 2018.
[35] V. Joshi and R. K. Jaiman. A hybrid variational Allen-Cahn/ALE scheme for the coupled analysis of
two-phase fluid-structure interaction. International Journal for Numerical Methods in Engineering,
117(4):405–429, 2019.
[36] V. Joshi, R. K. Jaiman, and C. Ollivier-Gooch. A variational flexible multibody formulation for par-
titioned fluid–structure interaction: Application to bat-inspired drones and unmanned air-vehicles.
Computers & Mathematics with Applications, 80(12):2707–2737, 2020.
[37] A. Karpatne, W. Watkins, J. Read, and V. Kumar. Physics-guided neural networks (pgnn): An
application in lake temperature modeling. arXiv preprint arXiv:1710.11431, 2, 2017.
[38] S. R. Kashyap and R. K. Jaiman. A robust and accurate finite element framework for cavitating
flows with fluid-structure interaction. 103:19–39, 2021.
[39] D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
42
[40] I. E. Lagaris, A. Likas, and D. I. Fotiadis. Artificial neural networks for solving ordinary and partial
differential equations. IEEE transactions on neural networks, 9(5):987–1000, 1998.
[41] Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. nature, 521(7553):436, 2015.
[42] Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anand-
kumar. Fourier neural operator for parametric partial differential equations. arXiv preprint
arXiv:2010.08895, 2020.
[43] Z. Li, W. Yao, K. Yang, R. K. Jaiman, and B. C. Khoo. On the vortex-induced oscillations of a
freely vibrating cylinder in the vicinity of a stationary plane wall. Journal of Fluids and Structures,
65:495–526, 2016.
[44] L. Lu, P. Jin, and G. E. Karniadakis. Deeponet: Learning nonlinear operators for identifying
differential equations based on the universal approximation theorem of operators. arXiv preprint
arXiv:1910.03193, 2019.
[45] S. Lunz, O. Öktem, and C. Schönlieb. Adversarial regularizers in inverse problems. In Advances in
Neural Information Processing Systems, pages 8507–8516, 2018.
[46] A. Maas, Q. V. Le, T. M. O’neil, O. Vinyals, P. Nguyen, and A. Y. Ng. Recurrent neural networks
for noise reduction in robust asr. 2012.
[47] W. Mallik, R. K. Jaiman, and J. Jelovica. Kinematically consistent recurrent neural networks for
learning inverse problems in wave propagation. arxiv.org/pdf/2110.03903.pdf, 2021.
[48] T. P. Miyanawala and R. K. Jaiman. An efficient deep learning technique for the navier-stokes
equations: Application to unsteady wake flow dynamics. arXiv preprint arXiv:1710.09099, 2017.
[49] T. P. Miyanawala and R. K. Jaiman. A low-dimensional learning model via convolutional neural
networks for unsteady wake-body interaction. arXiv preprint arXiv:1807.09591, 2018.
[50] T. P. Miyanawala and R. K. Jaiman. A novel deep learning method for the predictions of current
forces on bluff bodies. In International Conference on Offshore Mechanics and Arctic Engineering,
volume 51210, page V002T08A003. American Society of Mechanical Engineers, 2018.
[51] T. P. Miyanawala and R. K. Jaiman. Decomposition of wake dynamics in fluid–structure interaction
via low-dimensional models. Journal of Fluid Mechanics, 867:723–764, 2019.
[52] T. P. Miyanawala and R. K. Jaiman. A hybrid data-driven deep learning technique for fluid-structure
interaction. In International Conference on Offshore Mechanics and Arctic Engineering, volume
58776, page V002T08A004. American Society of Mechanical Engineers, 2019.
[53] A. T. Mohan and D. V. Gaitonde. A deep learning based approach to reduced order modeling for
turbulent flow control using lstm neural networks. arXiv preprint arXiv:1804.09269, 2018.
[54] D. Mokbel, H. Abels, and S. Alanda. A phase-field model for fluid–structure interaction. Journal
of computational physics, 372:823–840, 2018.
[55] M. C. Mozer, D. Kazakov, and R. V. Lindsey. State-denoised recurrent neural networks. arXiv
preprint arXiv:1805.08394, 2018.
[56] N. Muralidhar, M. R. Islam, M. Marwah, A. Karpatne, and N. Ramakrishnan. Incorporating prior
domain knowledge into deep neural networks. In 2018 IEEE International Conference on Big Data
(Big Data), pages 36–45. IEEE, 2018.
43
[57] G. M. Nielson. A method for interpolating scattered data based upon a minimum norm network.
Mathematics of Computation, 40(161):253–271, 1983.
[58] E. J. Parish and K. Duraisamy. A paradigm for data-driven predictive modeling using field inversion
and machine learning. Journal of Computational Physics, 305:758–774, 2016.
[59] C. S. Peskin. The immersed boundary method. Acta numerica, 11:479–517, 2002.
[60] T. Pfaff, M. Fortunato, A. Sanchez-Gonzalez, and P. W. Battaglia. Learning mesh-based simulation
with graph networks. arXiv preprint arXiv:2010.03409, 2020.
[61] F. Pichi, F. Ballarin, G. Rozza, and J. S. Hesthaven. An artificial neural network approach to bifur-
cating phenomena in computational fluid dynamics. arXiv preprint arXiv:2109.10765, 2021.
[62] M. Raissi, P. Perdikaris, and G. E. Karniadakis. Physics-informed neural networks: A deep learning
framework for solving forward and inverse problems involving nonlinear partial differential equa-
tions. Journal of Computational Physics, 378:686–707, 2019.
[63] S. R. Reddy, A. R. Magee, and R. K. Jaiman. Deep convolutional recurrent autoencoders for flow
field prediction. arXiv, pages arXiv–2003, 2020.
[64] R. J. Renka, R. L. Renka, and A. K. Cline. A triangle-based c1 interpolation method. The Rocky
Mountain journal of mathematics, pages 223–237, 1984.
[65] S. H. Rudy, S. L. Brunton, J. L. Proctor, and J. N. Kutz. Data-driven discovery of partial differential
equations. Science Advances, 3(4):e1602614, 2017.
[66] L. Ruthotto and E. Haber. Deep neural networks motivated by partial differential equations. Journal
of Mathematical Imaging and Vision, pages 1–13, 2019.
[67] O. San and R. Maulik. Machine learning closures for model order reduction of thermal fluids.
Applied Mathematical Modelling, 60:681–710, 2018.
[68] J. Schmidhuber. Deep learning in neural networks: An overview. https://arxiv.org/abs/1404.7828,

2014.
[69] J. Schmidt-Hieber. The kolmogorov–arnold representation theorem revisited. Neural Networks,
137:119–126, 2021.
[70] J. A. Sethian. Level set methods and fast marching methods: evolving interfaces in computational
geometry, fluid mechanics, computer vision, and materials science, volume 3. Cambridge university
press, 1999.
[71] Y. Tay, M. Dehghani, D. Bahri, and D. Metzler. Efficient transformers: A survey. arXiv preprint
arXiv:2009.06732, 2020.
[72] E. J. Tuegel, A. R. Ingraffea, T. G. Eason, and S. Spottswood. Testing the manifold hypothesis. Int.
J of Aerospace Eng, 2011:1687–5966, 2011.
[73] A. Venkatraman, M. Hebert, and J. A. Bagnell. Improving multi-step prediction of learned time
series models. In Twenty-Ninth AAAI Conference on Artificial Intelligence, 2015.
[74] Q. Wang, J. S. Hesthaven, and D. Ray. Non-intrusive reduced order modeling of unsteady flows
using artificial neural networks with application to a combustion problem. Journal of computational
physics, 384:289–307, 2019.
44
[75] R. Wang. Physics-guided deep learning for dynamical systems: A survey. arXiv preprint
arXiv:2107.01272, 2021.
[76] R. Wang, K. Kashinath, M. Mustafa, A. Albert, and R. Yu. Towards physics-informed deep learning
for turbulent flow prediction. In Proceedings of the 26th ACM SIGKDD International Conference
on Knowledge Discovery & Data Mining, pages 1457–1466, 2020.
[77] J. Willard, X. Jia, S. Xu, M. Steinbach, and V. Kumar. Integrating physics-based modeling with
machine learning: A survey. arXiv preprint arXiv:2003.04919, 2020.
[78] Z. Yu. A dlm/fd method for fluid/flexible-body interactions. Journal of computational physics,
207(1):1–27, 2005.
[79] Y. Zhu, N. Zabaras, P. Koutsourelakis, and P. Perdikaris. Physics-constrained deep learning for
high-dimensional surrogate modeling and uncertainty quantification without labeled data. Journal
of Computational Physics, 394:56–81, 2019.
45

A Hybrid Partitioned Deep Learning Methodology For Moving Interface and Fluid-Structure Interaction

Uploaded by

Copyright:

Available Formats

A Hybrid Partitioned Deep Learning Methodology For Moving Interface and Fluid-Structure Interaction

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Hybrid Partitioned Deep Learning Methodology For Moving Interface and Fluid-Structure Interaction

Uploaded by

Copyright:

Available Formats

A hybrid partitioned deep learning methodology for moving interface

and fluid-structure interaction

Rachit Guptaa , Rajeev Jaimana,∗

Preprint submitted to Journal of Computers & Fluids November 2, 2021

2. Full-order vs reduced-order modeling

uf (t) = us (t) , (4)

2.2. Reduced-order modeling

F(z) = C + Bz + F0 (z). (7)

θ AE = argminθ AE [E(Z, F (Z; θ AE ))]. (9)

3. Hybrid partitioned DL-ROM framework

𝐒 𝑛−1 (𝐲ො 𝑛+𝑝−1 , 𝐒෠ 𝑛+𝑝−1 )

Flow field, rigid body motion & force

3.1. The POD-RNN framework: structural motion

Step 1: Construct the POD modes ν for the dataset Y

yi ≈ y + νAiν , i = 1, 2, ..., n, (10)

where Aiν = [ai1 i2 ik T k

Step 3: Reconstruction to the point cloud ŷn+1

is straightforward to reconstruct the full-order ALE displacement at t n+1 as follows:

consists of the normalised time coefficients.

Aic = fc (Si ; θc,enc ), i = 1, 2, ..., n, (16)

S0i = Si − S, i = 1, 2, ..., n, (19)

S = Ss01 Ss02 . . . Ss0Ns ∈ [0, 1]Nx ×Ny ×Nt ×Ns

Conv2D #1 Conv2D #2 Conv2D #𝑛𝑐𝑜𝑛𝑣

Generative output: Temporal Temporal

4. Snapshot-field transfer and load recovery

(2) Level-set (4) Damp noises:

Input: De-noised pixelated force signals Fb , full-order FSI forces FΓfs

To reconstruct smooth pixelated force signals Fb :

Initialise a uniform grid (Nx × Ny ) of cell size ∆x = ∆y:

n+1 n+2 n+p

To extract full-order FSI forces:

5. Application: Vortex-induced vibration (VIV)

Γ𝑖𝑛 Γ𝑜𝑢𝑡 30𝐷

Present 124 25,916 2.0292 0.1124 0.0055 0.5452

5.1. Limit cycle VIV response

5.1.1. Structural displacements via POD-RNN driver

ALE k Cell Nh∗ Optimizer α∗ Epochs Learning Decay step

5.1.3. Near-wake dynamics using the CRAN driver

Grid Cell-size mean(ζ CD,p ) (ζ CL,p )rms mean(E c ) ε(CD,p ) ε(CL,p )

128 0.0625 1.8295 0.1491 0.8534 7.6 ×10−3 3.2 ×10−1

Conv2D encoder DeConv2D decoder

5.2. Non-stationary VIV response

[20] N. B. Erichson, M. Muehlebach, and M. W. Mahoney. Physics-informed autoencoders for lyapunov-

[68] J. Schmidhuber. Deep learning in neural networks: An overview. https://arxiv.org/abs/1404.7828,

You might also like