betaFeedback

WorksPhysics and Astronomy

Physics-informed neural networks modeling for systems with moving immersed boundaries: application to an unsteady flow past a plunging foil

Abstract

Recently, physics informed neural networks (PINNs) have been explored extensively for solving various forward and inverse problems and facilitating querying applications in fluid mechanics applications. However, work on PINNs for unsteady flows past moving bodies, such as flapping wings is scarce. Earlier studies mostly relied on transferring to a body attached frame of reference which is restrictive towards handling multiple moving bodies or deforming structures. Hence, in the present work, an immersed boundary aware framework has been explored for developing surrogate models for unsteady flows past moving bodies. Specifically, simultaneous pressure recovery and velocity reconstruction from Immersed boundary method (IBM) simulation data has been investigated. While, efficacy of velocity reconstruction has been tested against the fine resolution IBM data, as a step further, the pressure recovered was compared with that of an arbitrary Lagrange Eulerian (ALE) based solver. Under this framework, two PINN variants, (i) a moving-boundary-enabled standard Navier-Stokes based PINN (MB-PINN), and, (ii) a moving-boundary-enabled IBM based PINN (MB-IBM-PINN) have been formulated. A fluid-solid partitioning of the physics losses in MB-IBM-PINN has been allowed, in order to investigate the effects of solid body points while training. This enables MB-IBM-PINN to match with the performance of MB-PINN under certain loss weighting conditions. MB-PINN is found to be superior to MB-IBM-PINN when a priori knowledge of the solid body position and velocity are available. To improve the data efficiency of MB-PINN, a physics based data sampling technique has also been investigated. It is observed that a suitable combination of physics constraint relaxation and physics based sampling can achieve a model performance comparable to the case of using all the data points, under a fixed training budget.

Introduction

Unsteady flows past flapping wings involving complex flow dynamics [1, 2, 3] are studied extensively to understand the underlying mechanisms of flight in natural flyers, and enable design of effective bio-mimetic propulsive [4] and energy harvesting devices [5]. With the advancements in high performance computing, this has become possible through high fidelity computational fluid dynamics (CFD) simulations. In the context of simulating unsteady flows past moving bodies, broadly, two classes of methods exist to spatially discretise the computational domain, body conformal, and, body non-conformal mesh approaches. Arbitrary Lagrange Eulerian (ALE) framework [6, 7] falls under the first category, where, the mesh deforms in conjunction with the body motion. An example of the latter is Immersed Boundary Method (IBM) [8, 9, 10], in which a fixed Eulerian grid is used to discretise the domain where the solid boundary (described by a set of Lagrangian marker points), is immersed. Overall, IBM enables handling complex geometries and multiple moving bodies more easily by alleviating the need for mesh motion, re-meshing and the associated mesh quality constraints otherwise required in ALE. As a result, IBM incurs comparatively lesser computational and memory costs. Despite that, high fidelity simulations still involve significant computational burdens and are memory intensive for parametric exploration or real time query. These requirements are frequently encountered in solving ill-posed problems where boundary conditions are unknown, or, while solving inverse problems such as parameter estimation, system identification, optimisation [11], as well as for hidden physics recovery [12, 13]. Hence, cost effective alternatives or surrogates to minimise the associated computational and memory costs for solving such class of problems in the domain of unsteady flows past moving bodies are necessary.

Instantaneous pressure field for flow around moving bodies such as natural swimmers or flyers is important due to the complex interactions between pressure associated with the acceleration of the body and the vortices in the flow [14]. In experiments, either flow visualisation data of a passive scalar such as the concentration of contrast agents/smoke/dye is available [1], or, planar velocity data is available through particle image velocimetry (PIV) measurements [15]. These quantities are much easier to measure than the pressure-field. Hence, recovering instantaneous pressure non-invasively from the flow visualisation/PIV velocity data would be extremely useful in understanding the underlying mechanisms and also for estimating the loads on a moving test specimen [16, 15, 17]. In the case of numerical simulations too, near-field pressure recovery from velocity data is desirable. In some variants of IBM [18, 19], a mathematical artifice is used to enable the computation of the aero-/hydro-dynamic loads from the total time derivative of velocity and the momentum forcing terms in the momentum conservation equation. In such cases, the pressure data are not even stored. Therefore, an accurate physics consistent near-field pressure recovery and simultaneous velocity reconstruction from the coarse/sparse velocity data would be highly desirable. This also adds to the effort towards further reducing the high memory costs, often incurred in storing high-fidelity simulation data. Indeed, it would be convenient to query both velocity and pressure once recovered at a specific spatio-temporal coordinate within the region of interest, with minimal efforts.

Nevertheless, surrogate modeling based on simulation data obtained from IBM poses several challenges [20]. Specifically, as a result of employing a fixed Eulerian background grid, the region of the grid bounded by the Lagrangian markers contains fictitious flow-field values which are necessary to the IBM computation. Thus, at any time instant, the Eulerian grid points falling within the range of the solid body motion could either lie in the fictitious flow domain (that shall be called the solid domain from now on) or the fluid domain. Moreover, enforcing the no-slip boundary condition on the solid boundary becomes difficult as the Lagrangian markers are disjoint from the underlying Eulerian grid points. Thus, while handling moving bodies or complex geometries becomes easier with IBM [21], obtaining a surrogate capable of both recovering pressure and reconstructing the velocity accurately becomes challenging.

Recently, machine learning (ML) methods have gained particular interest in the fluid mechanics community towards developing surrogate models as ML is effective in solving complex inverse and ill posed problems while also suitable for real time query applications [11, 22, 23, 24]. There are two paradigms in ML based approaches, (i) data driven [22], and, (ii) physics informed [11, 25, 26, 27]. Data driven approaches are well developed but require large amount of data to capture the essential dynamics of complex physical systems. For some cases, the combination of reduced-order modeling and machine learning allows to distill and exploit the underlying patterns and correlations within large-scale data sets, leading to enhanced predictive capabilities and streamlined simulations, ultimately accelerating the design and optimization processes [28, 29].

But while data driven methods are physics agnostic, other methods embed partial/complete information of the underlying physics (in some manner). The present work focuses on the latter, specifically, the physics informed neural networks (PINNs) [30] which is a class of deep neural networks [31]. Formulated as an optimisation problem, PINNs embed prior knowledge of the governing physics through the objective/loss function [30]. The added physics loss component regularises learning on the available data and also enables solving different forward/inverse problems with minimal change in the architecture [30]. In the context of inverse/ill-posed problems, PINNs are commonly used for, data-driven discovery of PDEs [30], data assimilation [32], and, hidden physics discovery [33]. Recent surveys have presented the diverse set of domains in which PINNs have been applied [26, 27]. Once trained on a given spatio-temporal domain, the model allows for query at any temporal/ spatial location within the given domain. Moreover, PINNs are memory efficient; once trained, the model parameters alone are sufficient to be saved to carry out the querying. Other advantages are, they are automatically differentiable by obtaining a continuous function of input variables, and extremely flexible to implement using the well developed deep learning [31] software platforms capable of automatic differentiation [34] such as Tensorflow [35], PyTorch [36] and Jax [37].

Standard collocation PINNs [30] are sometimes difficult to train for forward problems due to underlying failure modes, broadly classified under: competing optimization objectives, and, propagation failures as discussed in [38, 39]. Various strategies have been proposed lately— such as, loss component balancing [38, 40], modified backbone architectures [38], adaptive sampling [41], domain decomposition [42] and sequential learning [39, 43]— to tackle these problems. A more detailed list of scalable algorithms to train PINNs efficiently can be found in [44, 43]. When no solution data is available, it has been shown that PINNs are relatively easier to train [45] if sparse/coarse simulation results or partially observed experimental data are at one’s disposal. It has also been shown that when simulation data is available, the expressivity of PINNs can be improved by appropriately relaxing the physics loss components, without the need for any architectural modifications [46]. In light of this, with simulation data being available in the present study, it is also of particular interest to investigate strategies that improve the predictive capability of PINNs under a fixed computational budget without the need for architectural changes.

PINNs based approaches have also been recently tested for problems involving moving boundaries/interfaces. A PINN was proposed [47] with a two-network architecture to solve simple benchmark problems on free/moving interfaces (known as Stefan problems), in both forward and inverse settings. One network predicts/learns the previously unknown/known moving interface position, and the other learns the complete solution of the underlying PDEs with appropriate constraints at the interface, enforced through the loss function. A two-phase incompressible flow problem with moving interface was solved in both forward and inverse settings [48], using PINNs with a volume of fluid formulation. Hidden physics recovery was performed using the volume fraction as training variables, where the velocity and pressure fields were recovered as hidden variables. Motivated by the fictitious domain method (FDM), Yang et al. [49] proposed a FDM-PINN approach and demonstrated its capability in solving forward problems with stationary and moving bodies involving linear elliptic/parabolic PDEs. A few studies in the area of vortex induced vibrations of a bluff-body have also been reported [32, 50, 51, 50]. These studies considered a body attached frame of reference to model the flow around the body. Note that transferring CFD data from an Eulerian frame of reference to a body attached frame of reference is not feasible in the case of multiple moving bodies or deforming structures. In such situations, a fixed Eulerian frame of reference (as used in IBM) is beneficial and effective. To combine the advantages of IBM and PINNs, Huang et al. [52] proposed a direct forcing based IB-PINNs approach and applied it on a steady flow problem past a stationary cylinder for Reynolds numbers Re<40Re < 40. The physics losses were minimised in the entire domain considering the solid and fluid regions together. However, it is not clear if in the IB-PINN formulation, including the solid region in the physics loss calculation is necessary when the body position and velocity are known a priori. More recently, in another application on moving body [53], a standard Navier-Stokes (N-S) formulation based PINN considering a non body-attached (inertial) frame of reference was used. PIV data was used and the near-field pressure data around the moving body was recovered effectively in this approach. It would thus be interesting to test the efficacy of IB-PINN [52] in comparison with the standard collocation based PINN in an inertial frame of reference, where the solid region is discarded. Such a comparison has not been undertaken in the literature, to the best of the authors’ knowledge. The role of the solid region grid points in an IBM setting on the overall performance of the PINN is also yet to be understood. These are some of the pertinent issues that need to be addressed in light of the above discussion, in the context of handling flow past moving bodies especially in a non body-attached frame of reference.

The key objectives and contributions of the present study are identified in the following. The application is in the area of unsteady incompressible flows past flapping wings and to the best of the authors’ knowledge, this is the first time PINNs’ performance have been analysed in this context in a body non-conformal setting. Note that PINNs based modeling of the flow in a body attached frame of reference could be restrictive and is not amenable for multiple moving bodies or deforming structures. A body non-conformal surrogate modeling framework using PINNs is developed here for pressure recovery from the velocity data. A fixed Eulerian background computational grid and a Lagrangian description of the solid boundary are considered. Given the generality of the approach, it is capable of handling multiple moving bodies/deforming structures, however, a single moving body has been considered in the present case for demonstrating the framework, for the sake of simplicity. Under this framework, two variants of PINN are investigated, one based on the standard N-S equations considering the fluid region alone, and the other involving the IBM based modified N-S equations, where both fluid and solid regions are considered for physics loss computation. The models are trained on coarsened velocity data and tested on high resolution data generated using an in-house IBM based unsteady flow solver [19]. As discussed earlier in this section, the pressure data are not usually available/stored in the IBM approach when alternative mathematical formulation for load calculation are used [19, 9]. Hence, demonstration of the pressure recovery is an important part of the present investigation. To establish confidence, the pressure recovered using the PINN models are validated against that obtained using an ALE solver. This serves as a rigorous test of the surrogate model’s capability to learn the underlying physics.

To improve the model performance under a fixed computational budget, the efficacy of physics residual loss relaxation is also investigated. A residual loss formulation allowing relative fluid-solid relaxation is explored here, in order to resolve the role of the solid region in the performance of the second variant, when the position and the velocity of solid body are known a priori. To train the PINN models in a data efficient manner while maintaining good accuracy, a physics based training data sampling strategy is explored.

The rest of the paper is organised as follows. In section 2, the immersed boundary aware PINNs based surrogate modeling framework is discussed in the context of pressure recovery and velocity reconstruction. The pre setup considerations for the PINNs including training-testing database generation and hyperparameter tuning are discussed in section 3. Sections 4 and 5 demonstrate the effects of physics constraint relaxation and a physics based vorticity cut off sampling strategy. Finally, the key conclusions and further works are outlined in section 6.

PINNs for unsteady flows past a moving body

Often in flow problems involving a moving body, the solid body is allowed to move either based on a prescribed kinematics or as dictated by the structural dynamical equations coupled with the flow solver. The present study deals with the former scenario. One of the main aims of the present surrogate modelling framework is to recover the near-field pressure as a hidden variable, and reconstruct velocity data at any spatio-temporal coordinate within a domain of interest while enabling real time query. The present study proposes to achieve this by training a PINN in a body non-conformal manner. As in the IBM, where the flow is investigated in an Eulerian grid and the solid is represented in a Lagrangian fashion, a similar strategy is proposed here and hence can also be called an immersed boundary aware (IBA) approach. This prevents the need for mapping the computational domain to a body-attached frame of reference to tackle a moving boundary problem as was done in earlier studies [32, 50]. Under the proposed framework, two formulations, namely, (i) a moving-boundary-enabled standard Navier-Stokes based PINN (MB-PINN), and, (ii) moving-boundary-enabled IBM based PINN (MB-IBM-PINN), are studied.

The details of the problem setup, unsteady flow modeling, the MB-PINN and MB-IBM-PINN architectures with their respective loss formulation and training methodology are presented in the following subsections.

The flapping foil system

Two-dimensional unsteady incompressible flow past a plunging rigid elliptic airfoil at a low Reynolds number is considered. The foil follows a harmonic kinematics and is considered to be of unit chord length (c=1c = 1) with a thickness to chord ratio of 0.125. It is subjected to uniform free-stream; see Figure 1 for a schematic of the problem. The plunging kinematics is given by,

Schematic of the problem setup

Figure 1. Schematics of the problem setup: computational domain Ω\Omega is chosen for the IBM solver, and the truncated domain Ωr\Omega^r is chosen for the surrogate model. Ωf\Omega_f and Ωfr\Omega_f^r are the fluid regions excluding the solid boundary ΓIB\Gamma_{IB} and solid region Ωs\Omega_s at any given time instant. Γinletr\Gamma^r_{\mathrm{inlet}}, Γoutletr\Gamma^r_{\mathrm{outlet}}, Γupperr\Gamma^r_{\mathrm{upper}} and Γlowerr\Gamma^r_{\mathrm{lower}} are the inlet, outlet, upper and lower boundaries of the truncated domain Ωr\Omega_r, respectively.

y(t)=hacos⁡(2πfht),and(1)y(t) = h_a \cos(2\pi f_h t), \qquad\text{and} \tag*{(1)}
y˙(t)=−2πfhhasin⁡(2πfht).(2)\dot{y}(t) = -2\pi f_h h_a \sin(2\pi f_h t). \tag*{(2)}

Here, tt is dimensional time; hah_a and fhf_h are the plunging amplitude and frequency, respectively. Aligning with earlier literature [54, 55], the non-dimensional amplitude h=ha/ch = h_a/c and reduced frequency k=2πfhc/U∞k = 2\pi f_h c/U_\infty are defined, where, U∞U_\infty is the free stream velocity. Further description of the flapping motion will be made in the terms of kk and hh.

Modeling the unsteady flow

The flow around the flapping foil is assumed to be governed by the incompressible Navier-Stokes (N-S) equations given in its non-dimensional form by

∂u∂t+∇⋅(uu)=−∇p+1Re∇2u,(3)\frac{\partial\mathbf{u}}{\partial t}+\nabla\cdot(\mathbf{u}\mathbf{u})=-\nabla p+\frac{1}{Re}\nabla^2\mathbf{u}, \tag*{(3)}
∇⋅u=0.(4)\nabla\cdot\mathbf{u}=0. \tag*{(4)}

Here, u\mathbf{u} represents the non-dimensional velocity vector in the xx-yy space with uu and vv being the xx and yy components of u\mathbf{u}, respectively, and pp is the non-dimensional pressure. Re=U∞cνRe=\frac{U_{\infty}c}{\nu} indicates the Reynolds number with ν\nu being the kinematic viscosity.

A discrete forcing IBM based in-house flow solver [19] has been employed here to solve Eqs. (3) and (4) for generating the training and testing data for the velocity fields. In IBM approach, the immersed solid boundary ΓIB\Gamma_{IB} is represented by a set of Lagrangian markers (figure 2(a)) which do not conform exactly with the Eulerian fluid grid. Importantly, at the solid boundary ΓIB\Gamma_{IB}, a no-slip boundary condition is to be satisfied such that u(x=ΓIB,t)=y˙(t)\mathbf{u}(\mathbf{x}=\Gamma_{IB},t)=\dot{\mathbf{y}}(t). Since this cannot be directly enforced on ΓIB\Gamma_{IB} in the IBM, a momentum forcing term f\mathbf{f} with fxf_x and fyf_y as the respective xx and yy components must be added in Eq. (3). f\mathbf{f} is evaluated such that the no-slip boundary condition is satisfied on ΓIB\Gamma_{IB} using appropriate mapping and interpolation between the Eulerian and Lagrangian points. Additionally, a source/sink term qq is added in Eq. (4) to strictly satisfy the continuity equation. Notably f\mathbf{f} and qq are non-zero only inside Ωs\Omega_s and in the vicinity of ΓIB\Gamma_{IB}, they are zero elsewhere in Ωf\Omega_f as depicted by the filled/unfilled markers and colour coded arrows in figure 2(b). The governing equations in the current IBM approach takes the form as below,

∂u∂t+∇⋅(uu)=−∇p+1Re∇2u+f,(5)\frac{\partial\mathbf{u}}{\partial t}+\nabla\cdot(\mathbf{u}\mathbf{u})=-\nabla p+\frac{1}{Re}\nabla^2\mathbf{u}+\mathbf{f}, \tag*{(5)}
∇⋅u−q=0.(6)\nabla\cdot\mathbf{u}-q=0. \tag*{(6)}

For more details on the IBM solver, please refer to [19, 56].

In the current IBM solver [19], the pressure-field was not required for computing the aerodynamic loads and hence this data was not stored. For a rigorous testing of the pressure recovery capability of the PINN models, an ALE-based flow solver has been used to generate and test the pressure data.

Schematics of the computational domain and mesh grid used in the immersed boundary method

Figure 2. Schematics of (a) a representative computational domain Ω=Ωf∪Ωs\Omega= \Omega_f \cup\Omega_s used in the immersed boundary method. Here Ωs\Omega_s is the shaded area representing the solid region bounded by the solid boundary ΓIB\Gamma_{IB}, and (b) the mesh grid representing the staggered primitive variable arrangement and fluid-solid grid point classifications.

In the ALE technique, the governing equations are solved as it is without the need for any additional forcing or mass source/sink terms in the equations. The ALE based simulations were performed using foam-Extend’s icoDyM-Foam solver [57]. Details of computational domain, boundary conditions and mesh, and a thorough validation of the above IBM and ALE solvers for the low past a flapping foil can be found in the supplementary material.

Network architecture

Among the two formulations of PINNs for moving boundaries being presented here, the MB-PINN is formulated only in the fluid region (discarding the solid region) while training; the MB-IBM-PINN is formulated in both the fluid and the solid body regions. The objective is to fit velocity bulk data obtained from IBM simulations and recover pressure as a hidden variable using both MB-PINN and MB-IBM-PINN while investigating their associated efficacy.

For the standard collocation PINN [30], a deep feed forward Neural network (FNN) is chosen as the backbone. Typically, for flow problems, a FNN is formulated to approximate the flow-field variables as outputs, given a spatio-temporal input [58]. The outputs are expressed as a composition of nonlinear activation functions acting on the spatio-temporal inputs and a network of fully connected hidden layers. Mathematically, this can be expressed as

h0=(x,t)(7)h_0 = (\mathbf{x}, t) \tag*{(7)}
zl=Wlhl−1+bl,(8)z_l = W_l h_{l-1} + b_l, \tag*{(8)}
hl=ϕ(Wlhl−1+bl), with l=1,2,⋯L, and(9)h_l = \phi(W_l h_{l-1} + b_l), \text{ with } l = 1, 2, \cdots L, \text{ and} \tag*{(9)}
o=WL+1hL+bL+1,(10)o = W_{L+1}h_L + b_{L+1}, \tag*{(10)}

where, (x,t)∈Rnx+1(\mathbf{x},t) \in\mathbb{R}^{n_x+1} is the spatio-temporal input with h0∈Rnx+1h_0 \in\mathbb{R}^{n_x+1} being the input layer and nxn_x being the spatial dimension. Here, zl∈Rnlz_l \in\mathbb{R}^{n_l} and hl∈Rnlh_l \in\mathbb{R}^{n_l} are the pre-activation and activation units of the neural network with LL hidden layers and nln_l hidden neurons for 1<l<L1 < l < L, respectively, with ϕ(.)\phi(.) being a non-linear activation function. The quantities Wl∈Rnl−1×nlW_l \in\mathbb{R}^{n_{l-1} \times n_l} and bl∈Rnlb_l \in\mathbb{R}^{n_l} for 1<l<L+11 < l < L + 1, represented by θ\boldsymbol{\theta} as a whole, are the weights and biases of the neural network to be optimised, respectively. Under the IBA framework, for the MB-PINN (see schematic figure Figure 3(a)), the outputs o\boldsymbol{o} are the flow-field variables {u,p}\{\boldsymbol{u}, p\}. Whereas, in MB-IBM-PINN (see schematic figure Figure 3(b)), the outputs o\boldsymbol{o} are the flow-field variables and additionally, the body forcing and mass source/sink terms such that o={u,p,f,q}\boldsymbol{o} = \{\boldsymbol{u}, p, \boldsymbol{f}, q\}. A FNN is often trained by optimizing the network parameters θ\boldsymbol{\theta} with a large but potentially sparse or corrupt set of training data consisting of inputs and their corresponding target labels/outputs represented by {(xi,ti),o^i}i=1Ntrain\{(\boldsymbol{x}^{i}, t^{i}), \hat{\boldsymbol{o}}^{i}\}_{i=1}^{N_{\mathrm{train}}} consisting of NtrainN_{\mathrm{train}} input-output samples with o^\hat{\boldsymbol{o}} representing the true outputs to be fitted by the network. The optimization is carried out by minimising a loss/cost function, L\mathcal{L}, to obtain an optimal set of network parameters, θ\boldsymbol{\theta}, such that

Image not extracted

Figure 3. Schematic of (a) MB-PINN and (b) MB-IBM-PINN network architectures used under the present IBA framework

θoptimal=arg⁡min⁡θ(L).(11)\boldsymbol{\theta}^{\mathrm{optimal}} = \underset{\boldsymbol{\theta}}{\arg\min}(\mathcal{L}). \tag*{(11)}

Numerically, the loss minimisation to obtain the optimal network parameters is often performed by first randomly initialising θ\boldsymbol{\theta} and then iteratively updating it using a variant of a stochastic gradient descent (SGD) based algorithm. Here, the gradients of L\mathcal{L} with respect to θ\boldsymbol{\theta} given by ∇θL\nabla_{\boldsymbol{\theta}}\mathcal{L} are to be calculated. These gradients can be computed using back propagation and automatic differentiation (AD) [34] which is an accurate and efficient way of computing derivatives in a computational graph following the chain rule of differentiation. AD is implemented in most modern deep learning packages, such as Tensorflow [35] which is used in the present study. Generally, given KMBK_{MB} mini-batches (sub sets) of the training data containing NMBN_{MB} samples to evaluate the loss and its gradients with respect to the parameters, the algorithm for the parameter updates based on SGD is given by

θi+1←θi−ηi1NMB∑i′=iNMB+1(i+1)NMB∇Lθ(θ,(xi′,ti′)),i=1,2,⋯KMB(12)\boldsymbol{\theta}^{i+1} \leftarrow\boldsymbol{\theta}^{i} - \eta_i \frac{1}{N_{MB}} \sum_{i'=iN_{MB}+1}^{(i+1)N_{MB}} \nabla\mathcal{L}_{\theta}\left(\boldsymbol{\theta},(x^{i'},t^{i'})\right), \quad i=1,2,\cdots K_{MB} \tag*{(12)}

where ηi\eta_i is the learning rate corresponding to ithi^{\mathrm{th}} epoch where each epoch corresponds to KMB=Ntrain/NMBK_{MB}=N_{train}/N_{MB} SGD iterations. Specifically, ADAM[59], a variant of SGD has been used in the literature [30] which will be followed in the present work as well. The detailed loss formulations of MB-PINN and MB-IBM-PINN are discussed in the following section.

Loss formulation

The loss function L\mathcal{L} of a PINN relying on both data and physical knowledge is represented as

L:=LData+LPhy,(13)\mathcal{L} := \mathcal{L}_{Data} + \mathcal{L}_{Phy}, \tag*{(13)}

where LData\mathcal{L}_{Data} and LPhy\mathcal{L}_{Phy} correspond to the loss contributions from the data and physics components, respectively. Here, Ldata\mathcal{L}_{data} is an indicator of the misfit between the true outputs o^train\hat{\mathbf{o}}^{train} and predicted network outputs o\mathbf{o}, respectively; LPhy\mathcal{L}_{Phy} uses AD to compute the gradients of o\mathbf{o} with respect to the input variables (x,t)(x,t) and compose them together to compute the governing equation residuals. This way, a PINN embeds the prior knowledge of the underlying physics in the training algorithm.

The data and physics loss components can be further split into respective weighted contributions [60, 38, 61] from initial and boundary conditions, interior bulk data and governing equation residuals. The loss formulation in terms of the various components is as follows

LData:=αLBulk+β(LBC+LIC)+γLIB(14)\mathcal{L}_{Data} := \alpha\mathcal{L}_{Bulk} + \beta(\mathcal{L}_{BC} + \mathcal{L}_{IC}) + \gamma\mathcal{L}_{IB} \tag*{(14)}
LPhy:=λ(LPDEmx+LPDEmy+LPDEc).(15)\mathcal{L}_{Phy} := \lambda(\mathcal{L}_{PDE_{mx}} + \mathcal{L}_{PDE_{my}} + \mathcal{L}_{PDE_c}). \tag*{(15)}

Here, LBulk\mathcal{L}_{Bulk}, LBC\mathcal{L}_{BC}, LIC\mathcal{L}_{IC} and LIB\mathcal{L}_{IB} are the loss components corresponding to the predicted interior bulk velocity data, velocity boundary conditions and initial condition on the Eulerian grid. The parameters α,β\alpha,\beta are the weighting coefficients of the bulk data loss, initial and boundary condition loss components, respectively. Here, γ\gamma is the generic weighting coefficient of the additional no-slip velocity boundary condition loss LIB\mathcal{L}_{IB}. It is characteristic to the IBA framework which allows enforcing the no-slip boundary conditions directly on the Lagrangian points and similar to [52] instead of following a set of sequential steps and interpolation as in IBM [19]. This is possible owing to the underlying flexibility of PINNs in choosing collocation points anywhere in the domain to enforce appropriate constraints on them. Thus, both MB-PINN and MB-IBM-PINN allow flexibility in providing both the Eulerian grid points describing the fluid and Lagrangian points describing the solid boundary, respectively. λ\lambda is the weighting coefficient of components of the physics informed loss LPhy\mathcal{L}_{Phy}, which is discussed later in this section, after presenting the data driven loss components.

The data driven loss components mentioned above are common to both MB-PINN and MB-IBM-PINN, which can be mathematically expressed as

LBulk:=1NBulk∑i=1NBulk∥u(xBulki,ti)−u^(xBulki,ti)∥L22,(16)\mathcal{L}_{Bulk} := \frac{1}{N_{Bulk}} \sum_{i=1}^{N_{Bulk}} \left\lVert\mathbf{u}(x^{i}_{Bulk},t^{i})-\hat{\mathbf{u}}(x^{i}_{Bulk},t^{i}) \right\rVert_{L_{2}}^{2}, \tag*{(16)}
LBC:=1Nk∑k=1Nk(1NBCk∑i=1NBCk∥u(xBCki,ti)−u^(xBCki,ti)∥L22),(17)\mathcal{L}_{BC} := \frac{1}{N_{k}} \sum_{k=1}^{N_{k}} \left( \frac{1}{N_{BC_{k}}} \sum_{i=1}^{N_{BC_{k}}} \left\lVert\mathbf{u}(x^{i}_{BC_{k}},t^{i})-\hat{\mathbf{u}}(x^{i}_{BC_{k}},t^{i}) \right\rVert_{L_{2}}^{2} \right), \tag*{(17)}
LIC:=1NIC∑i=1NIC∥u(xICi,t0)−u^(xICi,t0)∥L22,(18)\mathcal{L}_{IC} := \frac{1}{N_{IC}} \sum_{i=1}^{N_{IC}} \left\lVert\mathbf{u}(x^{i}_{IC},t_{0})-\hat{\mathbf{u}}(x^{i}_{IC},t_{0}) \right\rVert_{L_{2}}^{2}, \tag*{(18)}

and

LIB:=1NIB∑i=1NIB∥u(xIBi,t0)−u^(xIBi,t0)∥L22,(19)\mathcal{L}_{IB} := \frac{1}{N_{IB}} \sum_{i=1}^{N_{IB}} \left\lVert\mathbf{u}(x^{i}_{IB},t_{0})-\hat{\mathbf{u}}(x^{i}_{IB},t_{0}) \right\rVert_{L_{2}}^{2}, \tag*{(19)}

where, u^\hat{\mathbf{u}} is the true velocity data generated for training by the CFD simulation, whereas, u\mathbf{u} is the network predicted velocity flow-field data. The spatial and temporal points are represented by (x,t)(\mathbf{x},t) with x∈Ωr\mathbf{x} \in\Omega^{r} and t∈[0,T]t \in[0,T], respectively. The subscripts shown in the loss expressions for the spatio-temporal points are to indicate independence in sampling the select points of the particular loss component (as depicted by different markers in the schematic Figure 4). Here, (xBulki,ti)(\mathbf{x}^{i}_{Bulk},t^{i}) for i={1,⋯ ,NBulk}i = \{1,\cdots,N_{Bulk}\} corresponds to the set of interior velocity data points excluding the solid region. The set of points (xBCki,ti)(\mathbf{x}^{i}_{BC_{k}},t^{i}) for i={1⋯NBCk}i = \{1\cdots N_{BC_{k}}\} with k={1,2,3}k = \{1,2,3\} correspond to the Dirichlet boundary conditions collected at the inlet, Γinletr\Gamma^{r}_{inlet}, upper and lower boundaries, Γupperr\Gamma^{r}_{upper} and Γlowerr\Gamma^{r}_{lower}, respectively (see Figure 1), of the truncated domain Ωr\Omega^{r} considered. For simplicity, the inlet and the upper-lower boundary condition losses shall be referred to as Linlet\mathcal{L}_{inlet} and Lwalls\mathcal{L}_{walls}, respectively. The set of points (xICi,t0)(\mathbf{x}^{i}_{IC},t_{0}) for i={1,⋯ ,NIC}i = \{1,\cdots,N_{IC}\} are collected at the initial time stamp. Here, (xIBi,ti)(\mathbf{x}^{i}_{IB},t^{i}) for i={1,⋯ ,NIB}i = \{1,\cdots,N_{IB}\} correspond to the set of Lagrangian markers lying on the immersed solid boundary as represented in Figure 4.

A schematic representing sampling of residual, bulk and boundary data points at two different time instants

Figure 4. A schematic representing sampling of residual, bulk and boundary data points at two different time instants in (a) and (b). Although in the present work, bulk data points and fluid region residual points are sampled from same Eulerian grid, the bulk data points and fluid residual points are shown to be disjoint considering a more general case.

In the case of MB-IBM-PINN, additional constraints are required for the momentum forcing term, f\mathbf{f}, and the mass source/sink term, qq, which are also additional outputs in the network (see Figure 3b). It is known that f\mathbf{f} and qq have support only in the solid region bounded by ΓIB\Gamma_{IB}, and are zero in the fluid region at any time instant. However, both f\mathbf{f} and qq are not explicitly known inside the solid region as only velocity u\mathbf{u} data is stored. In [52], in a similar IBM based PINN formulation, velocity penalties were enforced on the Lagrangian markers to satisfy the no-slip boundary condition on the immersed boundary. It was discussed that such an enforcement compensates for the unknown ff and qq terms in the solid region. Hence, inside the solid region, ff and qq are recovered as hidden variables given the velocity penalties are enforced on the immersed solid boundary. This is in contrast with the discrete forcing IBM formulation [19], where, ff and qq are computed in a set of sequential steps to satisfy the no-slip boundary condition. However, since ff and qq are known to be zero in the fluid domain, an additional data driven constraint LIBvar\mathcal{L}_{IBvar} is formulated based on [52], such that

LIBvar:=1Nf∑i=1NIBvar(∥fx(xIBvari,ti)∥L22+∥fy(xIBvari,ti)∥L22+∥q(xIBvari,ti)∥L22),(20)\mathcal{L}_{IBvar}:=\frac{1}{N_f}\sum_{i=1}^{N_{IBvar}}\left(\left\|f_x(\mathbf{x}_{IBvar}^{i},t^i)\right\|_{L_2}^2+\left\|f_y(\mathbf{x}_{IBvar}^{i},t^i)\right\|_{L_2}^2+\left\|q(\mathbf{x}_{IBvar}^{i},t^i)\right\|_{L_2}^2\right), \tag*{(20)}

where, (xIBvari,ti)(\mathbf{x}_{IBvar}^{i},t^i) for i={1,⋯NIBvar}i=\{1,\cdots N_{IBvar}\} is the set of points from the interior of the fluid domain Ωfr\Omega_f^r where the forcing and source/sink terms loss is evaluated. Thus, the data-driven loss formulation of MB-IBM-PINN is as follows

LData:=αLBulk+β(LBC+LIC)+γLIB+ξLIBvar,(21)\mathcal{L}_{Data}:=\alpha\mathcal{L}_{Bulk}+\beta(\mathcal{L}_{BC}+\mathcal{L}_{IC})+\gamma\mathcal{L}_{IB}+\xi\mathcal{L}_{IBvar}, \tag*{(21)}

with ξ\xi being the weighting coefficient of LIBvar\mathcal{L}_{IBvar}.

The physics informed loss components LPDEmx\mathcal{L}_{PDE_{mx}}, and LPDEmy\mathcal{L}_{PDE_{my}} correspond to the xx and yy momentum residuals and LPDEc\mathcal{L}_{PDE_c} corresponds to the continuity residual, respectively, which are expressed as follows

LPDEmx:=1NRes∑i=1NRes∥rmx(xResi,ti)∥L22,(22)\mathcal{L}_{PDE_{mx}}:=\frac{1}{N_{Res}}\sum_{i=1}^{N_{Res}}\left\|r_m^x(\mathbf{x}_{Res}^{i},t^i)\right\|_{L_2}^2, \tag*{(22)}
LPDEmy:=1NRes∑i=1NRes∥rmy(xResi,ti)∥L22,(23)\mathcal{L}_{PDE_{my}}:=\frac{1}{N_{Res}}\sum_{i=1}^{N_{Res}}\left\|r_m^y(\mathbf{x}_{Res}^{i},t^i)\right\|_{L_2}^2, \tag*{(23)}

and

LPDEc:=1NRes∑i=1NRes∥rcy(xResi,ti)∥L22.(24)\mathcal{L}_{PDE_c}:=\frac{1}{N_{Res}}\sum_{i=1}^{N_{Res}}\left\|r_c^y(\mathbf{x}_{Res}^{i},t^i)\right\|_{L_2}^2. \tag*{(24)}

For MB-PINN, the following forms of the residuals are used

rmx(xResi,ti)=ut+uux+vuy−px−1/Re⁡(uxx+uyy),(25)r_m^x(\mathbf{x}_{Res}^i,t^i) = u_t + uu_x + vu_y - p_x - 1/\operatorname{Re}(u_{xx} + u_{yy}), \tag*{(25)}
rmy(xResi,ti)=vt+uvx+vvy−py−1/Re⁡(vxx+vyy),(26)r_m^y(\mathbf{x}_{Res}^i,t^i) = v_t + uv_x + vv_y - p_y - 1/\operatorname{Re}(v_{xx} + v_{yy}), \tag*{(26)}
rc(xResi,ti)=ux+vy.(27)r_c(\mathbf{x}_{Res}^i,t^i) = u_x + v_y. \tag*{(27)}

In MB-IBM-PINN, the momentum and continuity residuals are expressed as

rmx(xResi,ti)=ut+uux+vuy−px−1/Re⁡(uxx+uyy)−fx,(28)r_m^x(\mathbf{x}_{Res}^i,t^i) = u_t + uu_x + vu_y - p_x - 1/\operatorname{Re}(u_{xx} + u_{yy}) - f_x, \tag*{(28)}
rmy(xResi,ti)=vt+uvx+vvy−py−1/Re⁡(vxx+vyy)−fy,(29)r_m^y(\mathbf{x}_{Res}^i,t^i) = v_t + uv_x + vv_y - p_y - 1/\operatorname{Re}(v_{xx} + v_{yy}) - f_y, \tag*{(29)}
rc(xResi,ti)=ux+vy−q,(30)r_c(\mathbf{x}_{Res}^i,t^i) = u_x + v_y - q, \tag*{(30)}

where, (xResi,ti)(\mathbf{x}_{Res}^i,t^i) for i={1,⋯NRes}i = \{1,\cdots N_{Res}\} is the set of points from the interior of the domain Ωr\Omega^r where the governing equation residuals are evaluated to compute the physics loss LPhy\mathcal{L}_{Phy}.

In MB-PINN, solid points from Ωs\Omega_s at any given temporal snapshot are discarded while computing the bulk data and physics losses. This is achieved by classifying the fluid-solid points based on the solid boundary location a priori as a data preprocessing step. Thus, the set of bulk data points (xBulki,ti)i=1NBulk(\mathbf{x}_{Bulk}^i,t^i)_{i=1}^{N_{Bulk}} and residual collocation points (xResi,ti)i=1NRes(\mathbf{x}_{Res}^i,t^i)_{i=1}^{N_{Res}}, only from Ωrffluid\Omega_{rf}^{fluid} are fed into the network while training for MB-PINN. It is to be noted that unless the solid boundary points are known a priori, it is not possible to classify the domain as solid or fluid. MB-IBM-PINN involves a more general formulation based on the governing equations (equations (5) and (6)), where the whole snapshot domain for the residual loss, including Ωs\Omega_s, is considered. For the pressure recovery problem, the boundary position and the velocity is known a priori. In such a scenario, it is of interest to understand the role of the solid region in the performance of MB-IBM-PINN and compare with MB-PINN.

The bulk data and residual collocation points in the IBA framework come from an underlying Eulerian grid. Here, due to solid body motion, the solid region Ωs\Omega_s and fluid region Ωfr\Omega_f^r are expected to change in time such that Ωr=Ωfr∪Ωs\Omega^r = \Omega_f^r \cup\Omega_s. Thus, the underlying Eulerian grid points in the region covered by the entire range of solid body motion, can either be found in Ωs\Omega_s or in Ωr\Omega^r at any time instant as shown in 4(a) and 4(b). Earlier works, including [52, 53], do not discuss the role of solid region in the model performance. Thus, in order to determine the exact contribution from each region on the quality of predictions, a fluid-solid residual weighting is proposed in the present study for the MB-IBM-PINN formulation. Here, the residual losses are individually evaluated on the fluid and the solid regions at any time instant respectively with the corresponding loss weights.

Thus, the LPhy\mathcal{L}_{Phy} for MB-IBM-PINN is decomposed as,

LPhy=λfluid(LPDEmxfluid+LPDEmyfluid+LPDEcfluid)+λsolid(LPDEmxsolid+LPDEmysolid+LPDEcsolid),(31)\mathcal{L}_{Phy} = \lambda_{fluid}\left(\mathcal{L}^{fluid}_{PDE_{m_x}} + \mathcal{L}^{fluid}_{PDE_{m_y}} + \mathcal{L}^{fluid}_{PDE_c}\right) + \lambda_{solid}\left(\mathcal{L}^{solid}_{PDE_{m_x}} + \mathcal{L}^{solid}_{PDE_{m_y}} + \mathcal{L}^{solid}_{PDE_c}\right), \tag*{(31)}

where, λ#\lambda_{\#}, LPDEmx#\mathcal{L}^{\#}_{PDE_{m_x}}, LPDEmy#\mathcal{L}^{\#}_{PDE_{m_y}}, and LPDEc#\mathcal{L}^{\#}_{PDE_c} correspond to the residual loss weighting coefficient, xx and yy momentum residuals and continuity residual evaluated in the respective #={solid,fluid}\# = \{solid, fluid\} regions. Hence, in this case, two sets of residual collocation points, one from Ωfr\Omega_f^r given by (xResi,ti)i=1NResfluid(\mathbf{x}_{Res}^{i},t^{i})_{i=1}^{N_{Res}^{fluid}}, and one from region Ωs\Omega_s given by (xResj,tj)j=1NRessolid(\mathbf{x}_{Res}^{j},t^{j})_{j=1}^{N_{Res}^{solid}}, are needed for the evaluation of the physics losses. This allows one to sample different proportions of collocation points from each region. Additionally, the relative weighting allows balancing of the relative contributions from the fluid and the solid regions. This is keeping in mind that density of the points in a particular region is equivalent to weighting the loss locally in that region [41]. To make consistent comparison with the MB-IBM-PINN model and given that MB-PINN residuals are evaluated only in the fluid region, its LPhy\mathcal{L}_{Phy} can be expressed in terms of #=fluid region\# = \text{fluid region} such that

LPhy:=λfluid(LPDEmxfluid+LPDEmyfluid+LPDEcfluid).(32)\mathcal{L}_{Phy} := \lambda_{fluid}\left(\mathcal{L}_{PDE_{m_x}}^{fluid} + \mathcal{L}_{PDE_{m_y}}^{fluid} + \mathcal{L}_{PDE_c}^{fluid}\right). \tag*{(32)}

In every mini batch of collocation points corresponding to LBulk\mathcal{L}_{Bulk} and LPhyfluid\mathcal{L}_{Phy}^{fluid} from the same fluid region Ωfr\Omega_f^r are mostly disjoint from each other due to random sampling. However, there still exists some overlap since the underlying grid from which these points are sampled, is kept the same in this study. Notably, given that PINNs are essentially meshless and there is freedom to select points from anywhere in the domain for the physics losses, one could also sample from a different grid than the Eulerian grid in order to keep the residual and the bulk data points truly disjoint [41]. This allows finer control over the sampling methodology and sampling levels for each loss component individually. The proportion and refinement levels for each loss component can also be varied for each individual component thereby implicitly weighting the loss components [41]. Moreover, each individual component can be tracked and their weights can be updated adaptively, as reported in [38, 40, 41]. Adaptive methods have not been investigated in the present study to focus on understanding the effect of systematic variations of loss weights and data sampling. Each loss component was monitored independently to understand how every component evolves with iterations.

The details of database generation for training and testing the models and hyperparameter tuning are discussed in § 3. The numerical settings considered for MB-PINN and MB-IBM-PINN along with the baseline results, and computational strategies considered to further improve the baseline performance are discussed in § 4 and § 5.

Pre-setup considerations for PINN

Training and testing database generation

Training and testing data for the velocity-field have been obtained from the IBM solver described in section 2.2. In this regard, the flow around the plunging airfoil was simulated at Re=500Re = 500, k=2πk = 2\pi and h=0.16h = 0.16 corresponding to kh=1.0kh = 1.0. This resulted in a periodic flow-filed with a reverse Kármán vortex street in the wake of the flapping foil. An ALE based OpenFOAM solver was employed to generate the test data set for pressure, as was mentioned in section 2.2. It is known that PINNs do suffer from the issue of propagation failures [42, 39], leading to training difficulties on large spatial/temporal domains. Moreover, earlier works [1, 62, 63, 64] have shown that the key flow features that largely dictate the dynamics and the aerodynamic loads in flapping problems are strongest in the vicinity of the body. For the considered flapping kinematic parameters, the resulting flow is periodic [55], making the vortex structures and their interactions repeat in time. Keeping these in mind, to setup the training and testing databases, a truncated spatial domain (Figure 1) is chosen such that (x,y)∈[−1c,3.5c]×[−1c,1c](x,y) \in[-1c, 3.5c] \times[-1c, 1c] ensuring that at least two trailing-wake couples are contained in the truncated domain along with the flapping foil. The time domain is restricted to two plunging cycles for t/T>25.0t/T > 25.0 and normalised such that t/T∈[0,2]t/T \in[0, 2] with a snapshot sampling interval of Δt/T=0.025\Delta t/T = 0.025. Here, the snapshot interval is chosen such that the formations of the leading-edge vortex (LEV) and the trailing-edge vortex (TEV) are captured smoothly within one time period of vortex shedding. Note that the transient effects in the flow-field are over by t/T>25.0t/T > 25.0 and the vortex shedding pattern is stabilised. The high resolution IBM velocity data from a non-uniform grid is interpolated onto a coarse uniform grid to bring down the quality of the training data. This serves as a rigorous test of the IBA surrogate since the coarse grid is neither embedded in the IBM nor the ALE grid.

Thus, three benchmark data sets are generated from the CFD simulations, (i) reference IBM velocity data (Ref-IBM), (ii) reference ALE pressure data (Ref-ALE) and (iii) coarsened and interpolated (CI) velocity data. In the case of CI data set, the flow-field data from the IBM solution are interpolated onto a coarse grid containing 270×120270 \times120 spatial grid points with a uniform spatial resolution (Figure 5(a)) approximately 5 times lower near the solid boundary than the original IBM grid. However, the resolution of CI grid is sufficient such that the key vortex structures are not lost. Secondly for training, the time interval Δt/T=0.05\Delta t/T = 0.05 is chosen such that temporal resolution is two times coarser as compared to Ref-IBM. Different spatial resolutions of CI, Ref-IBM, and Ref-ALE are shown in Figure 5(b)-(Figure 5d), respectively. The details of the spatio-temporal resolution of these data sets are summarized in Table 1. These data sets correspond to the truncated spatio-temporal domain and not full entire domain considered for CFD simulations.

Spatial grid of CI data set with zoomed views of CI, Ref-IBM, and Ref-ALE

Figure 5. (a) Spatial grid of CI data set with the Lagrangian markers in blue, inlet and wall boundary points in magenta, and the solid region in red colours, respectively. Zoomed view of the spatial grid for (b) CI, (c) Ref-IBM and (d) Ref-ALE presents the comparative spatial resolution near the solid boundary.

Data setsNxN_xNyN_yNtN_tΔt/T\Delta t/T(Nx×Ny×Nt)(N_x \times N_y \times N_t)
Ref-IBM651500810.0252.6365e07
Ref-ALE--810.0254.05e06
CI270120410.051.3284e06

Table 1. Training and testing data set resolution within the truncated domain considered in this study.

The training of the models is performed solely based on the velocity data of CI. The velocity-field reconstructions are compared with Ref-IBM, whereas the pressure recovered by the PINN models is compared with Ref-ALE.

Among the total Nx×Ny×Nt=1.3284e06N_x \times N_y \times N_t = 1.3284\mathrm{e}06 data points in CI, NBulk=1.2915e06N_{Bulk} = 1.2915\mathrm{e}06 number of points were used to calculate LBulk\mathcal{L}_{Bulk} for training after discarding the solid region. This NBulkN_{Bulk} is just SData=NBulkNRef×100=4.89%S_{Data} = \frac{N_{Bulk}}{N_{Ref}} \times100 = 4.89\% of the NRef=Nx×Ny×NtN_{Ref} = N_x \times N_y \times N_t flow-field data points available in the Ref-IBM data set. Here, in CI data set, the boundary points resolution is such

that NIB=1000N_{IB}=1000, Nwall=540N_{wall}=540, and Ninlet=120N_{inlet}=120. At the initial time stamp t/T=0t/T=0, NIC=31500N_{IC}=31500 flow field data points are sampled discarding the solid region. Here, the residual collocation points are sampled from the CI grid such that there are Nresfluid=1.2915e06N_{res}^{fluid}=1.2915e06 and Nressolid=1.476e04N_{res}^{solid}=1.476e04 spatio-temporal points in the fluid and solid region respectively. Throughout the study, the number of boundary points and the number of residual collocation points are not varied.

By choosing the grid as in CI, there is a significant reduction in the number of training data points that encompass the flow-field information. This is in fact similar to a super-resolution setup with reconstruction of the flow-field from a coarse data [65]. It is noted that data coarseness is often more pronounced in the super-resolution studies. In the current study, we attempt to train the proposed PINN models only based on velocity data from the coarse CI data set to predict the velocity and pressure to be validated against Ref-IBM and Ref-ALE data sets, respectively. In other words, a data reconstruction problem coupled with hidden variable recovery is at hand. It is worth noting that the pressure data is not used while training throughout the study and is only recovered as a hidden variable.

To evaluate the model performance, different error metrics such as root mean squared error (aRMSE), mean absolute error (aMAE), coefficient of determination (aR2aR^2) and relative root mean squared error (arRMSE also called overall relative L2 error) are considered. These are computed over velocity reconstruction and pressure recovery against the reference IBM and ALE data sets (Ref-IBM and Ref-ALE), respectively. Here, the prefix “a” for every error metric discussed here denotes averaging of the error metrics over uu, vv evaluated against Ref-IBM data and pp evaluated against Ref-ALE data. In addition, snapshot wise relative root mean squared errors (rRMSE), and the contours of normalised point-wise normalised absolute errors are presented to determine the temporal and spatial error behaviour, respectively. Here, the point-wise absolute errors are normalised with respect to the maximum absolute value of the corresponding flow-field component. Mathematical details of the above mentioned error metrics can be found in the supplementary material. Moreover, the capability of the proposed surrogate models in accurately resolving the vorticity field is investigated as well. This is crucial in the context of flow past flapping wings where the vortex structures play an important role in dictating the dynamics.

Hyper parameter tuning

To determine an appropriate number of hidden layers and hidden neurons, the underlying FNN backbone (now referred as MB-FNN) was considered. Here, MB-FNN was trained on the CI data set to reconstruct velocity fields alone in a purely data driven manner without considering the physics losses.

The number of hidden layers and hidden neurons were chosen after a grid search on combinations of L=[6,8,10]L = [6, 8, 10] layers and nl=[60,80,100]n_l = [60, 80, 100] hidden neurons respectively. The resultant models were evaluated on the basis of velocity reconstruction alone on the Ref-IBM data set (accuracy details presented in Table 2). It was observed that with the converged L=10L = 10 hidden layers and nl=100n_l = 100 hidden neurons, remarkable relative RMSE of 0.51% in xx-component of velocity uu and 1.31% in yy-component of velocity vv were obtained (see Table 3). This shows that MB-FNN is quite expressive. Throughout the study, these hyper-parameters are kept the same for both MB-PINN and MB-IBM-PINN models as well.

Table 2: Accuracy details for the MB-FNN models (without any physical constraints) for different number of hidden layers LL and hidden neurons per layer nln_l. The model is tested on the Ref-IBM data set.

LLnln_laaRMSEaaMAEaR2aR^2arRMSE
Accuracy
6601.74e-027.9e-039.9929e-012.47
6801.455e-026.3e-039.9947e-011.26
61001.34e-025.55e-039.9956e-011.16
8601.17e-025.15e-039.9967e-011.67
8807.9e-033.25e-039.9984e-011.13
81006.75e-032.55e-039.9987e-010.98
10609.75e-054.44e-039.9976e-011.4
10807.01e-032.8e-039.9987e-011.01
101006.15e-032.2e-039.9989e-010.91

Table 2. Accuracy details for the MB-FNN models (without any physical constraints) for different number of hidden layers LL and hidden neurons per layer nln_l. The model is tested on the Ref-IBM data set.

Table 3: Velocity component wise accuracy of the best MB-FNN model. The model is tested on the Ref-IBM data sets.

ModelAccuracy
MB-FNNRMSEMAER2R^2rRMSE
uu5.8e-032.1e-039.9997e-010.51
vv6.5e-032.3e-039.9981e-011.31

Table 3. Velocity component wise accuracy of the best MB-FNN model. The model is tested on the Ref-IBM data sets.

For the MB-PINN model with L=10L = 10 layers and nl=100n_l = 100 hidden neurons, θ=9.1603e04\theta= 9.1603e04 parameters need to be calibrated while training. Whereas, due to three additional outputs in MB-IBM-PINN, θ=9.1906e04\theta= 9.1906e04 parameters need to be calibrated while training. Unless specified, a maximum computational budget of 5e055e05 iterations per training cycle in learning rate steps of [1e−03,5e−04,1e−04][1e - 03, 5e - 04, 1e - 04] has been considered with a mini batch size of 1.5e031.5e03, throughout the study. Training of all the models were carried out on a single NVIDIA Tesla A100 GPU card with 40GB memory and 6912 CUDA cores. Whereas, testing was carried out on a desktop grade Quadro Pascal P2200 card with 5GB memory and 1280 CUDA cores. This is to investigate whether a comparable performance from MB-PINN and MB-IBM-PINN can be obtained with the same architecture and computational budget, for not only velocity reconstruction but also pressure recovery as a hidden variable. The multi-part physics loss weighting studies have been carried out to answer this question.

Role of multi-part physics loss weighting

To obtain a good model performance under a fixed computational budget, an efficient strategy could be relaxing the physics loss components [48, 46], or inversely, weighting the data driven loss components higher than the physics counterpart [38, 66, 40]. The former approach has been considered in the present study, where the weights of all the data driven components are taken to be unity and only the physics loss components are relaxed by lowering their appropriate weighting coefficients. Dynamic weighting strategies have been proposed in some of the recent studies [38, 40, 44, 60, 61]. In the present study, a manual tuning strategy of the residual loss weights is adopted to understand their contributions towards pressure recovery in a systematic manner.

To recall from § 2, data sets at the solid region are discarded for computing both LBulk\mathcal{L}_{Bulk} and LPhy\mathcal{L}_{Phy} in MB-PINN, thus only λfluid\lambda_{fluid} is in consideration. In MB-IBM-PINN, LBulk\mathcal{L}_{Bulk} computation is the same as in MB-PINN, i.e., only the fluid region data points are considered. The solid region is taken into account in LPhy\mathcal{L}_{Phy} computation along with the fluid region, with λsolid\lambda_{solid} and λfluid\lambda_{fluid} as respective weights (see Eq. (31)). In the present study, a multi-part physics loss weighting strategy thus refers to the fluid-solid partitioning of the physics loss and relative weighting between them as discussed in § 2. This choice of a fluid-solid partitioning is done to better quantify the impact of the solid region residual losses. As a result, there is an extra hyper parameter, λsolid\lambda_{solid}, to vary in MB-IBM-PINN.

In the beginning, without any loss balancing (i.e. λfluid=λsolid=1.0\lambda_{fluid} = \lambda_{solid} = 1.0) it was observed that the predictions from both MB-PINN and MB-IBM-PINN were significantly worse, with the snapshot wise relative errors being more than 20% for velocity reconstruction (Figure 6(a)-(Figure 6b)), and over 30% for pressure recovery (Figure 6(c)). Thus, reasonable baseline models for MB-PINN and MB-IBM-PINN were first obtained with an overall mean relative error, given by rRMSE, below 10% for velocity reconstruction, and below 20% for pressure recovery. This was achieved in the case of λfluid=0.1\lambda_{fluid} = 0.1 for MB-PINN. To match the MB-PINN’s performance, for MB-IBM-PINN with the same λfluid=0.1\lambda_{fluid} = 0.1, a relatively lower λsolid=0.001\lambda_{solid} = 0.001 was required, which suggests the need of undermining the solid region loss. In the further discussions, the MB-PINN and MB-IBM-PINN models with λfluid=0.1\lambda_{fluid} = 0.1, and λsolid=0.001\lambda_{solid} = 0.001, shall be referred to as the baseline cases. In the relative error plots (Figure 6), the MB-PINN and MB-IBM-PINN curves almost overlap each other except at a few later time instances, where MB-IBM-PINN does slightly worse for the selected baseline λfluid\lambda_{fluid} and λsolid\lambda_{solid} values. The equivalence of MB-PINN and MB-IBM-PINN is also confirmed by the similar order of the convergence of the overall training and individual loss components LBulk\mathcal{L}_{Bulk}, scaled LPhy\mathcal{L}_{Phy}, LIC\mathcal{L}_{IC}, Linlet\mathcal{L}_{inlet}, and Lwall\mathcal{L}_{wall} in Figure 7(a) and Figure 7(b), respectively.

Relative errors in time for velocity components and recovered pressure

Figure 6. Relative errors (rRMSE) in time for (a, b) xx and yy velocity components with respect to Ref-IBM data set, and (c) for recovered pressure with respect to Ref-ALE data set. Here, MB-PINN (represented as MB) and MB-IBM-PINN (represented as MB-IBM) models are compared, considering λfluid=1\lambda_{fluid}=1 and λfluid=0.1\lambda_{fluid}=0.1.

Smoothened individual loss component convergence

Figure 7. Smoothened individual loss component convergence shown for the baseline (a) MB-PINN with λfluid=0.1\lambda_{fluid}=0.1 and (b) MB-IBM-PINN with λfluid=0.1\lambda_{fluid}=0.1, λsolid=0.001\lambda_{solid}=0.001.

The true and predicted velocity and pressure contours are compared in Figure 8(a) and Figure 8(b). The corresponding maximum value normalised point-wise absolute error contours in the near-field region are shown in Figure 8(c) for a test time stamp t/T=0.375t/T = 0.375, which is previously unseen during the training. At this time stamp, the velocity and pressure fields are queried on high resolution Ref-IBM and Ref-ALE data sets test grids, respectively. It can be seen in Figure 8(c) that the errors are highly pronounced in the region where there is an LEV. This is expected as strong gradients of velocity and pressure exist in such regions of strong vorticity. The issue of difficulty of PINN in capturing strong/sharp gradient has also been reported in the earlier literature [38]. For the test time t/T=0.375t/T = 0.375, a characteristic strong LEV on the upper surface of the foil followed by a secondary vortex structure in the mid-region of the upper surface is observed (see Figure 9(a)). The baseline models are able to resolve the primary vortex structures very well, while the secondary structure is still blurred in the predictions as seen in Figure 9(b) and Figure 9(c). With regard to the pressure recovery, although the baseline results seem promising, the errors are still high. In order to see if larger training iterations (NiterN_{iter}) per cycle could improve the results, in another set of experiments with λfluid=0.1\lambda_{fluid} = 0.1 and λsolid=0.001\lambda_{solid} = 0.001 (see Table 4 for accuracy details), it is seen that the improvements are not too significant considering the increased computational budget. It remains to be seen if further relaxation of the residual losses can improve the pressure recovery and the overall model performance.

ModelNiterN_{iter}Accuracy
MB-PINN (λfluid=0.1\lambda_{fluid} = 0.1)7.5e05aRMSEaMAEaR2aR^2arRMSE
MB-PINN (λfluid=0.1\lambda_{fluid} = 0.1)7.5e055.64e-023.02e-029.95e-015.39
MB-PINN (λfluid=0.1\lambda_{fluid} = 0.1)1e065.24e-022.79e-029.96e-015.01
MB-IBM-PINN (λfluid=0.1\lambda_{fluid} = 0.1, λsolid=0.001\lambda_{solid} = 0.001)7.5e058.79e-024.87e-029.87e-018.32
MB-IBM-PINN (λfluid=0.1\lambda_{fluid} = 0.1, λsolid=0.001\lambda_{solid} = 0.001)1e068.19e-024.54e-029.89e-017.74

Table 4. Accuracy details of the MB-PINN and MB-IBM-PINN models under the effect of larger NiterN_{iter}.

Image not extracted

Figure 8. Comparison of true and predicted (a, b) velocity and (c) pressure snapshots queried on the Ref-IBM and Ref-ALE grids, respectively, for a test time stamp t/T=0.375t/\mathrm{T} = 0.375. The inshot boxes A# for # = 1, 2, ⋯\cdots 6 correspond to the near-field region, where, the normalised point-wise absolute errors in the flow-field are prominent.

Image not extracted

Figure 9. Comparison of (a) IBM obtained vorticity snapshot with the predictions of (b) baseline MB-PINN (λfluid=0.1\lambda_{fluid} = 0.1) and (c) MB-IBM-PINN (λfluid=0.1,λsolid=0.001\lambda_{fluid} = 0.1, \lambda_{solid} = 0.001) for a test time stamp t/T=0.375t/\mathrm{T} = 0.375. The inshot boxes Z correspond to the near-field region surrounding the LEV and the secondary vorticity structure.

The detailed error measures for the relaxation studies are presented in Table 5. In the case of MB-PINN, there is a significant decrease in the overall error measures as λfluid\lambda_{fluid} is decreased from 1 to 0.1. However, beyond λfluid=0.001\lambda_{fluid} = 0.001, the pressure recovery breaks down as seen in Table 5. It is observed that at λfluid=0.0001\lambda_{fluid} = 0.0001 the snapshot-wise relative errors (see Figure 10) are either as good as the baseline case or even worse in certain situations. Thus the most optimal results have been obtained for λfluid=0.001\lambda_{fluid} = 0.001 in the case of MB-PINN, with arRMSE = 3.22% which can be considered as the best performing MB-PINN model within the present set of experiments.

Relative errors in time for velocity components and pressure

Figure 10. Relative errors (rRMSE) in time for (a,b) xx and yy velocity components, and (c) pressure pp predicted by MB-PINN (represented as just MB) and MB-IBM-PINN (represented as MB-IBM) models with different λfluid\lambda_{fluid} and λsolid\lambda_{solid} combinations depicting the effect of further residual relaxation.

Modelλfluid\lambda_{fluid}λsolid\lambda_{solid}aRMSEaMAEaR2aR^2arRMSE (in %)
MB-PINN1-2.38e-011.37e-019.23e-0122.96
MB-PINN0.1-7.64e-024.09e-029.91e-017.27
MB-PINN0.01-4.09e-022.28e-029.97e-013.78
MB-PINN0.001-3.58e-022.07e-029.98e-013.22
MB-PINN0.0001-7.59e-024.87e-029.91e-016.56
MB-IBM-PINN110.361e-012.22e-018.24e-0134.11
MB-IBM-PINN10.12.60e-011.52e-019.04e-0124.81
MB-IBM-PINN10.012.25e-011.32e-019.28e-0121.56
MB-IBM-PINN10.0012.08e-011.19e-019.41e-0119.99
MB-IBM-PINN0.112.83e-011.71e-018.85e-0125.76
MB-IBM-PINN0.10.11.76e-019.94e-029.54e-0116.36
MB-IBM-PINN0.10.011.35e-017.60e-029.73e-0112.66
MB-IBM-PINN0.10.0019.14e-025.07e-029.87e-018.63
MB-IBM-PINN0.0010.0014.15e-012.69e-017.27e-0132.41
MB-IBM-PINN0.0010.00013.27e-012.078e-018.22e-0125.91
MB-IBM-PINN0.0010.000012.71e-011.719e-018.76e-0121.02
MB-IBM-PINN0.00103.59e-022.03e-029.98e-013.21

Table 5. Accuracy of MB-PINN and MB-IBM-PINN for different relaxation coefficients. Best performing models are highlighted in bold-faced fonts.

For MB-IBM-PINN, even for a fixed λfluid\lambda_{fluid}, the prediction accuracy improves overall when the λsolid\lambda_{solid} is lowered. This is observed for different λfluid\lambda_{fluid} values as shown in Table 5. To obtain a reasonable accuracy in velocity reconstruction and pressure recovery, indicated by arRMSE, λsolid\lambda_{solid} is expected to be at least one order lower than that of λfluid\lambda_{fluid}. This is true for the baseline λfluid=0.1\lambda_{fluid} = 0.1, with λsolid=0.001\lambda_{solid} = 0.001 that is two orders lower, where the accuracy

of the model almost matches MB-PINN model for same λfluid\lambda_{fluid} as shown in the relative error plots in Figure 6(a)-(Figure 6c). This requirement could be because relaxation is also implicitly related to choosing a relatively lower proportion of collocation points in the region [41]. This is not applicable in the case of MB-PINN since the solid region is not included in the formulation. By fixing λfluid=0.001\lambda_{fluid}=0.001 as in the optimal MB-PINN model, the percentage arRMSE blows up even at λsolid=0.0001\lambda_{solid}=0.0001 and 0.000010.00001. Given that a significantly lower λsolid\lambda_{solid} is indeed required to improve the accuracy further, the need for solid region losses in MB-IBM-PINN is hence investigated by choosing a limiting λsolid=0\lambda_{solid}=0 keeping λfluid=0.001\lambda_{fluid}=0.001. In this case, the MB-IBM-PINN model gives the best performance, matching that of MB-PINN. The arRMSE drastically drops as seen in Table 5.

It can also be seen that for the optimal MB-PINN and MB-IBM-PINN models, LData\mathcal{L}_{Data} (Figure 11(a)) comes closer to that of purely data driven MB-FNN model. Whereas, the scaled LPhy\mathcal{L}_{Phy} are the lowest for both variants of PINN at the optimal values of relaxation (see Figure 11(b)). The relative error

Overall effect of relaxation on loss convergence

Figure 11. Overall effect of relaxation on loss convergence studied for MB-PINN and MB-IBM-PINN models, considering different λfluid\lambda_{fluid} and λsolid\lambda_{solid} values. The combined data driven losses LData\mathcal{L}_{Data} are presented in comparison with the MB-FNN model in (a), and the scaled physics-informed losses LPhy\mathcal{L}_{Phy} in (b), respectively.

in time for velocity reconstruction and pressure recovery are plotted across different selected levels of λfluid\lambda_{fluid} and λsolid\lambda_{solid} for both PINN models in Figure 10(a)-(Figure 10c). It is seen that for MB-IBM-PINN, the relative errors in velocity reconstruction and pressure recovery closely follow the optimal MB-PINN model at all time instants for λfluid=0.001\lambda_{fluid}=0.001 and λsolid=0\lambda_{solid}=0.

It is thus seen that MB-IBM-PINN can match the performance of MB-PINN under certain combinations of λfluid\lambda_{fluid} and λsolid\lambda_{solid}, under the same training budget and network architecture. At optimal λfluid\lambda_{fluid} and for λsolid=0\lambda_{solid}=0, the solid region points can be effectively discarded in MB-IBM-PINN, for an equivalent performance to MB-PINN. This is because the no-slip boundary condition in both the PINN variants is directly enforced through LIB\mathcal{L}_{IB}, rather than satisfied indirectly as in IBM (see § 2.2), where the solid region points are required. This also suggests that MB-IBM-PINN might not be necessary when one is already aware of the solid body’s position and velocity a priori. Moving forward, MB-PINN can be used efficiently for pressure recovery from velocity data even if obtained from an underlying IBM solver whenever the solid body position and velocity is known a priori. MB-PINN model is allowed here because the forcing terms and source/sink terms are zero in the fluid region and non-zero in the solid region as enforced in LIBvar\mathcal{L}_{IBvar} for MB-IBM-PINN. However, there might be scenarios where the solid body/moving boundary position and velocity data might not be available a priori. This is true for applications where inferring the flow-field and the body position might be needed only from the wake data or for the prediction of an evolving moving boundary inside a closed duct/channel as in biomedical applications [67]. In such situations, one might be able to use MB-IBM-PINN without a fluid-solid splitting of residual losses, however, an overall physics loss relaxation would still need to be invoked.

Note that the surrogate model in the present study is able to handle the immersed moving boundary using the Lagrangian markers and the fluid using an Eulerian grid, respectively. The surrogate model with residual relaxation performs very well for queries on the high resolution grids and additionally recover pressure as a hidden variable. However, the number of data points used are still large even with the coarse CI data sets with about 1.328e06 data points in total. Therefore, given the localised features of the flow-field, the possibility of a physics based sampling of the data points is explored in the next section.

Improving data efficiency through physics-based data sampling

As mentioned earlier, capturing strong/sharp gradients may become challenging in PINNs [38], and hence, physics constraint relaxation was explored in § 4. In the baseline MB-PINN model, the secondary vortex structure was not resolved accurately while also smudging the LEV as shown in Figure 9. This was also reflected in the locally high point-wise errors in Figure 8 (marked in rectangular boxes). These errors were localised around the regions of strong LEVs and TEVs which were representative of strong velocity gradients in the flow-field. Many authors proposed adaptive sampling strategies to mitigate this [41, 66, 68, 69, 70]. It was noted that the sampling of spatial points based on a prior distribution is equivalent to point-wise weighting; [41] dealt with PDEs in three dimensions, while [70] proposed a deep adaptive sampling based PINNs (DAS-PINNs) for higher dimensional PDEs.

However, the above approaches entail iterative sampling of newer residual collocation points (and not data points) to mitigate the training difficulty. Recently, it was shown in [45] how the loss landscape of PINN becomes smoother in the presence of very sparse/coarse simulation/experimental data in the bulk, as compared to no data, thereby making PINNs easier to train. In the present study, a data-driven approach was thus adopted. In fact, high number of data points were also used in the optimal MB-PINN model identified in the previous section. To improve data-efficiency and reduce associated data storage requirement, the possibility of using the underlying features of the flow-field data to train the network for MB-PINN is explored in this section.

Vorticity magnitude levels are good indicators of strong velocity gradients and the current study proposes a data efficient vorticity cutoff based undersampling approach. Here a vorticity cutoff ωz∗\omega_z^* is chosen such that the key features can be retained locally while the regions of low vorticity are undersampled. The vorticity fields are computed from the velocity data using a second-order central-difference scheme. Once the vorticity fields are obtained, a reasonable absolute vorticity cutoff value of ∣ωz∗∣=1.0|\omega_z^*| = 1.0 is chosen. For each vorticity snapshot, only those N∣ωz∣≥∣ωz∗∣N_{|\omega_z|\geq|\omega_z^*|} fluid points corresponding to ∣ωz∣≥∣ωz∗∣|\omega_z| \geq|\omega_z^*| are first retained, which correspond to the regions of key primary and secondary vortex structures (see Figure 12(a)). Out of the remaining N∣ωz∣<∣ωz∗∣N_{|\omega_z|<|\omega_z^*|} fluid points corresponding to ∣ωz∣<∣ωz∗∣|\omega_z|<|\omega_z^*|, a randomly undersampled set of N∣ωz∣<∣ωz∗∣sampleN_{|\omega_z|<|\omega_z^*|}^{sample} points based on a percentage sampling ratio, Sωz=N∣ωz∣<∣ωz∗∣sampleN∣ωz∣<∣ωz∗∣×100S_{\omega_z}=\frac{N_{|\omega_z|<|\omega_z^*|}^{sample}}{N_{|\omega_z|<|\omega_z^*|}}\times100 is constructed and combined with the earlier retained data points. Note that ∣ωz∗∣|\omega_z^*| does not always need to be 1.0, this may take a different value for a different flow problem and should be selected based on the flow-field vorticity levels of the specific problem of interest. In the present study, Sωz=100%,50%,10%,5%,1%S_{\omega_z}=100\%, 50\%, 10\%, 5\%, 1\% and 0%0\% are considered. For the extreme case Sωz=0%S_{\omega_z}=0\%, strong vorticity points in the near-field (NF) are further undersampled with a sampling ratio defined by, SωzNF=N∣ωz∣>∣ωz∗∣sampleN∣ωz∣>∣ωz∗∣×100S_{\omega_z}^{NF}=\frac{N_{|\omega_z|>|\omega_z^*|}^{sample}}{N_{|\omega_z|>|\omega_z^*|}}\times100, such that SωzNF=10%S_{\omega_z}^{NF}=10\%. The spatial resolutions of the undersampled data sets are presented at a representative time stamp in Figure 12(b)-(Figure 12h). For each SωzS_{\omega_z}, the corresponding data sets are generated (see Table 6) with the same number of temporal snapshots as in the CI data set. The resultant models are tested on Ref-ALE for pressure recovery and Ref-IBM for velocity reconstruction. For all the cases considered here, velocity data on the boundaries, Γinletr\Gamma_{inlet}^{r}, Γupperr\Gamma_{upper}^{r} and Γlowerr\Gamma_{lower}^{r} (depicted by black marker points in Figure 12(b)-(Figure 12h)), are also used during the training step, along with the N∣ωz∣≥∣ωz∗∣N_{|\omega_z|\geq|\omega_z^*|} and N∣ωz∣<∣ωz∗∣sampleN_{|\omega_z|<|\omega_z^*|}^{sample} points.

Data setSωzS_{\omega_z}SωzNFS_{\omega_z}^{NF}NBulkN_{Bulk}NresN_{res}SDataS_{Data}
CI1001.2915e064.89
CI-S50507.5112e052.85
CI-S10103.1885e051.21
CI-S551002.6481e051.3062e061.00
CI-S112.2158e050.84
CI-S002.1078e050.79
CI-S0-US100102.1074e040.08

Table 6. Data sets generated through vorticity cutoff based sampling.

Image not extracted

Figure 12. Vorticity cutoff based undersampling strategy: (a) the selection of strong and weak vorticity regions is shown for a representative vorticity snapshot at t/T=0.0t/T = 0.0, which is characterised by a LEV at the bottom surface, (b)-(h) vorticity cutoff based undersampled data locations corresponding to the data sets in Table 6, (i, j) relative errors in velocity reconstruction, and (k) relative error in recovered pressure.

To understand the effectiveness of the vorticity based undersampling, the MB-PINN models are first trained with the different data sets mentioned in Table 6 for λfluid=0.1\lambda_{fluid}=0.1. Here, Sωz=100%S_{\omega_z}=100\% and λfluid=0.1\lambda_{fluid}=0.1 corresponds to the baseline MB-PINN model discussed in § 4. The detailed error measures are presented in Table 7. It is observed that, all the undersampled cases except the extreme case of CI-S0-US10, perform better in terms of the arRMSE compared to the baseline MB-PINN. With λfluid=0.1\lambda_{fluid}=0.1, the best results are obtained for Sωz=5%S_{\omega z}=5\%, with an overall arRMSE of 4.82% which is slightly higher than the earlier reported best case of λfluid=0.001\lambda_{fluid}=0.001 (without vorticity cutoff based undersampling) in Table 5 with an arRMSE of 3.217%. Next, keeping Sωz=5%S_{\omega z}=5\% fixed, as λfluid\lambda_{fluid} is reduced to 0.01, arRMSE again becomes comparable to that of λfluid=0.001\lambda_{fluid}=0.001 in Table 5, that too with only SData=1.00%S_{Data}=1.00\% for the CI-S5 data set as opposed to 4.89% in the case of the CI data set. The results for λfluid=0.001\lambda_{fluid}=0.001 are also reasonably good at Sωz=5%S_{\omega z}=5\%. However, the MB-PINN model demonstrates the best performance for λfluid=0.01\lambda_{fluid}=0.01 and Sωz=5%S_{\omega z}=5\%. Even in the snapshot-wise rRMSE plots for velocity reconstruction (fig-

Data baseSωzS_{\omega_z} (in %)aRMSEaMAEaR2aR^2arRMSE (in %)
CI1007.64e-024.09e-029.91e-017.27
CI-S50507.02e-023.93e-029.92e-016.55
CI-S10106.79e-023.99e-029.92e-016.29
CI-S555.01e-022.99e-029.96e-014.82
CI-S115.41e-023.47e-029.95e-015.01
CI-S005.11e-023.30e-029.96e-014.83
CI-S0-US10 (SωzNF=10%S_{\omega_z}^{NF}=10\%)09.6e-024.98e-029.87e-019.90
CI-S5 (λfluid=0.01\lambda_{fluid}=0.01)53.59e-022.25e-029.97e-013.22
CI-S5 (λfluid=0.001\lambda_{fluid}=0.001)54.78e-022.67e-029.96e-014.77

Table 7. Accuracy details of MB-PINN models for different levels of vorticity cutoff based undersampling. If not mentioned specifically, λfluid\lambda_{fluid} is taken as λfluid=0.1\lambda_{fluid}=0.1.

ures 12(i) and 12(j)}) and pressure recovery (figure 12(k)), it is seen that $\lambda_{fluid}=0.01$ and $S_{\omega_z}=5\%$ case performs the best overall. Interestingly, for the extreme case of $S_{\omega_z}=0\%$ and for $S_{\omega_z}^{NF}=10\%$, although the relative velocity reconstruction errors worsen ({xref}figures 12(i) and {xref}12(j)), the relative pressure errors have not as much and are contained below 15% ({xref}figure 12(k)`). This indicates that even with SData=0.08%S_{Data}=0.08\% with respect to Ref-IBM, MB-PINN is still able to recover pressure with reasonable accuracy. Note that both physics loss weighting and vorticity cutoff based sampling contribute towards relaxation of the physics constraints. When both are employed together, a relatively higher value of λfluid=0.01\lambda_{fluid}=0.01 is found to be sufficient to match the best model performance.

Hence, it is possible to leverage the underlying flow features to improve data efficiency, while maintaining the expressivity. This is also confirmed by the loss convergence plots in Figure 13. It can be clearly seen how the combined data-driven losses LData\mathcal{L}_{Data} for Sωz=5%S_{\omega_z}=5\% with λfluid=0.01,0.001\lambda_{fluid}=0.01,0.001 are closer to that of MB-FNN model in Figure 13(a) and correspondingly the scaled LPhy\mathcal{L}_{Phy} are lowest in Figure 13(b). The loss convergence of individual loss components for the best MB-PINN models in the relaxation and undersampling experiments are presented in Figure 13(c) and 13(d). Although the losses are slightly higher for the undersampled case, the order of convergence is still similar for both the optimal models considered.

Loss convergence plots for select MB-PINN models

Figure 13. Loss convergence plots for select MB-PINN models considering different levels of vorticity undersampling combined with residual relaxation: (a) Ldata\mathcal{L}_{data} with that of the best MB-FNN model, and (b) scaled Lphy\mathcal{L}_{phy}. Unless otherwise specified, λfluid=0.1\lambda_{fluid}=0.1 by default. Convergence of individual loss components for the optimal MB-PINN models with (c) λfluid=0.001\lambda_{fluid}=0.001, Sωz=100%S_{\omega z}=100\%, and (d) λfluid=0.01\lambda_{fluid}=0.01, Sωz=5%S_{\omega z}=5\%.

From the point of view of a fluid dynamicist, the accuracy of the prediction of the key vortex structures are of interest. To this end, flow-field measures, such as the circulation of the LEV (ΓLEV\Gamma_{LEV}), self-induced velocity (UdipoleU_{dipole}) of the first and the second dipoles in the wake are computed here. Errors in the predicted and derived quantities, ΓLEV\Gamma_{LEV} and UdipoleU_{dipole} are good indicators of how well the primary vortex structures have been resolved by the surrogate models. Given a uniform grid of cell size dx=dydx=dy, and the vorticity ωz\omega_z at any cell center, the calculation of LEV circulation ΓLEV\Gamma_{LEV} within the bounding box surrounding the LEV is given by

ΓLEV=∑i=1NLEVωzidxdy,(33)\Gamma_{LEV}=\sum_{i=1}^{N_{LEV}}\omega_z^i dxdy, \tag*{(33)}

where, NLEVN_{LEV} corresponds to the number of grid points in a bounding box containing the LEV. Also, the self induced velocity of a dipole A-B can be calculated as

Udipole=Γavg2πξAB,where,(34)U_{\mathrm{dipole}} = \frac{\Gamma_{\mathrm{avg}}}{2\pi\xi_{AB}}, \qquad\text{where,} \tag*{(34)}
Γavg=12(∣ΓA∣+∣ΓB∣),and,(35)\Gamma_{\mathrm{avg}} = \frac{1}{2}\left(|\Gamma_A|+|\Gamma_B|\right), \qquad\text{and,} \tag*{(35)}
ξAB=(xB−xA)2+(yB−yA)2.(36)\xi_{AB} = \sqrt{(x_B-x_A)^2+(y_B-y_A)^2}. \tag*{(36)}

Here, A and B indicate successive vortex cores that together form a dipole in the trailing-wake, with corresponding circulations ΓA\Gamma_A and ΓB\Gamma_B, respectively. The dipoles are marked by the two rectangular boxes in the wake; see Figure 14(a). ξAB\xi_{AB} is the distance between the centers of vortices A and B with

Comparison of vorticity contours from IBM data and MB-PINN model predictions

Figure 14. Comparison of vorticity contours (a) obtained from IBM data with the predictions of MB-PINN models in (b)-(l) for different λfluid\lambda_{fluid} and SωzS_{\omega_z} at a typical test time stamp t/T=0.375t/T = 0.375. The dotted boxes indicate the regions of strong LEV and vortex dipoles. The base line MB-PINN model prediction in (c) is highlighted by a blue bounding box, whereas, the best MB-PINN model predictions are highlighted by a red bounding box in (f), and (l), respectively.

ModelAccuracy
MB-FNNRMSEMAER2R^2rRMSE
uu5.8e-032.1e-039.999e-010.51
vv6.5e-032.3e-039.998e-011.31
λfluid=0.1\lambda_{fluid}=0.1 (Baseline)
MB-PINNRMSEMAER2R^2rRMSE
uu2.7e-021.1e-029.994e-012.41
vv3.0e-021.2e-029.961e-016.01
pp1.7e-011.0e-019.781e-0113.41
λfluid=0.001\lambda_{fluid}=0.001
MB-PINNRMSEMAER2R^2rRMSE
uu9.6e-033.8e-039.999e-010.84
vv8.3e-033.3e-039.997e-011.67
pp8.9e-025.5e-019.943e-017.14
λfluid=0.01\lambda_{fluid}=0.01, Sωz=5%S_{\omega_z}=5\%
MB-PINNRMSEMAER2R^2rRMSE
uu8.6e-034.9e-039.999e-010.75
vv8.2e-034.7e-039.997e-011.64
pp9.1e-025.8e-029.938e-017.28

Table 8. Accuracy of the best MB-PINN models in comparison with purely data driven MB-FNN model and the baseline MB-PINN model.

(xA,yA)(x_A, y_A) and (xB,yB)(x_B, y_B) being the respective center coordinates. For more details on the calculation of the above measures, one can refer to our earlier study [64]. Here, the above flow-field measures are calculated at a typical test time instant, t/T=0.375t/T = 0.375, which was unseen during the training. The results are presented qualitatively in Figure 14 and quantitatively in Table 9.

ModelϵΓLEV\epsilon_{\Gamma_{LEV}}ϵU1\epsilon_{U^1}ϵU2\epsilon_{U^2}
Relaxation study (different λfluid\lambda_{fluid})
λfluid=1\lambda_{fluid}=141.0060.7610.264
0.17.9940.3090.248
0.011.0330.1290.051
0.0010.2580.1750.047
0.00010.7580.1320.043
Undersampling study (different SωzS_{\omega_z} in %)
Sωz=50S_{\omega_z}=5010.4250.3970.244
108.1390.5850.188
51.3810.0130.018
11.9680.1050.003
01.0930.0160.149
0, SωzNF=10S_{\omega_z}^{NF}=103.0010.3990.031
5, λfluid=0.01\lambda_{fluid}=0.010.0730.0090.051
5, 0.0010.1490.0790.024

Table 9. Percentage errors ϵ#\epsilon_{\#} in predicting LEV circulation ΓLEV\Gamma_{LEV}, self induced dipole velocity UdipoleU_{dipole} for the first two dipoles in the wake at a typical test time stamp t/T=0.375t/T=0.375, for MB-PINN models. Here, #=ΓLEV,U1,U2\#=\Gamma_{LEV}, U^1, U^2. The true values of the derived quantities are ΓLEV=1.450\Gamma_{LEV}=1.450, Udipole1=0.394U^1_{dipole}=0.394, and Udipole2=0.379U^2_{dipole}=0.379.

It is observed that undersampling combined with relaxation (Sωz=5%S_{\omega_z} = 5\%, λfluid=0.01\lambda_{fluid} = 0.01) improves the LEV resolution and captures the other vortex structures remarkably well. While this is achieved, the data efficiency as compared to the optimal MB-PINN model obtained in the relaxation study, is also maintained. Notably, all the models predicted the vortex centers exactly at the locations of the respective true data. The primary and secondary vortex structures and the shear layers are resolved better compared to the baseline model in Figure 14(c), either when the residuals are just further re-

laxed (Figure 14(f)) or when relaxed and combined with the selective vorticity based sampling strategy (Figure 14(l)).

On comparing the test predictions obtained from the optimal MB-PINN models (Figure 15(a)-(Figure 15c)), one can see that even the secondary structure has been captured accurately at time stamp previously unseen by the network during training. The point-wise normalised errors in pressure in the vicinity of the LEV (as seen in the boxes B1 in Figure 15(b)-(Figure 15c)) are slightly worse than that of λfluid=0.001\lambda_{fluid}=0.001 and Sωz=100%S_{\omega_z}=100\% (as seen in the boxes B2 in Figure 15(b)-(Figure 15c)). In the velocity and pressure slices (see Figure 16(a)-(Figure 16c)) taken at x/c=0.0cx/c=0.0c, for the test time stamp t/T=0.375t/T=0.375, queried on the CI grid it is seen that the optimal models obtained in the relaxation and undersampling experiments are equivalent and closely match the true data. Moreover it can be seen in the insets near the boundary D1 in Figure 16(a), and D4 in Figure 16(b) that the no-slip velocity boundary condition is closely satisfied for the optimal MB-PINN models. As seen in Figure 16(c), the optimal models are very close to that of the true data in pressure recovery while for λfluid=0.0001\lambda_{\mathrm{fluid}} = 0.0001, it breaks down away from the solid boundary. In the slices of velocity (inshots D11-D14 of Figure 16(d) and Figure 16(e)) taken at a far field location x/c=1.5cx/c = 1.5c from the foil center, all the models are more or less similar in velocity reconstruction. Whereas, slight differences are seen in pressure recovery (as seen in insets D13 and D14 of Figure 16(f)) across optimal and sub optimal cases though it still follows the overall trend of the true data. Stark differences are observed only in the near-field region across parameters. However, it is still remarkable considering that the model was trained using an undersampling approach with about 20% of data points

Image not extracted

Figure 15. Comparison of (a) true IBM velocity and ALE pressure with the (b) predictions obtained from best MB-PINN models for a testing time stamp t/T=0.375t/T = 0.375, and (c) corresponding maximum value normalised point wise absolute error contours. Here, the rectangular boxes in (b) and (c) representing the near-field region in (a) marked by B1 correspond to λfluid\lambda_{\mathrm{fluid}} with Sωz=100%S_{\omega_z} = 100\% while that of B2 correspond to λfluid=0.01\lambda_{\mathrm{fluid}} = 0.01 with Sωz=5%S_{\omega_z} = 5\%.

Velocity and pressure slices for a test time stamp

Figure 16. Velocity and pressure slices for a test time stamp t/T=0.375t/T = 0.375 queried on the CI grid for different values of relaxation and vorticity undersampling. The obtained slices are compared with interpolated IBM velocity and ALE pressure data for a slice passing through, (a)-(c) the centre of the foil at x/c=0.0cx/c = 0.0c where the rectangular region marked by a gray patch represents the solid body and (d)-(f) with slices at a far field location x/c=1.5cx/c = 1.5c. The zoomed inshots are presented using the correspondingly labeled close up shots D1-D14

compared to the CI data set and about 1% to that of Ref-IBM, while still maintaining an equivalent level of performance.

Owing to the localised nature of the strong velocity gradients in the flow-field, the vorticity cutoff based data sampling leads to reasonable data efficiency (requires a significantly less number of data points for training), while maintaining the expressivity of the PINN model. There is almost a 5 times decrease in the data requirement for the optimal combination of λfluid=0.01\lambda_{\mathrm{fluid}} = 0.01 and Sωz=5%S_{\omega_z} = 5\%. Even at Sωz<5%S_{\omega_z} < 5\%, while reasonable accuracy level can still be obtained. It is also envisioned that, by training efficient PINN models on reduced snapshot domains of interest which have data points stored in an undersampled manner when the solver writes the data, one could potentially obtain huge savings on the memory as well. For a typical example calculation for the present problem, we observed the memory requirement per snapshot to go down by two orders of magnitude.

Summary and conclusions

An immersed boundary aware (body non conformal/ hybrid Lagrange-Eulerian) surrogate modeling framework has been proposed here for pressure recovery and velocity reconstruction in the unsteady flow-field past a flapping body at low Reynolds number. Physics informed neural networks are chosen as the prospective candidates. Taking inspiration from the immersed boundary method, the present framework considers a background Eulerian grid with the moving body immersed in it. This framework removes the limitations of case specific computational domain transformations to body attached frame of reference. Under the IBA framework, two formulations of PINNs, MB-PINN and MB-IBM-PINN have been presented based on the standard Navier-Stokes equations and IBM based modified formulation of the Navier-Stokes equations, respectively. The present problem encompasses key characteristics of the flows past moving bodies: strong vortices with localised regions of strong gradients and time varying flow behaviour. The above mentioned characteristics pose challenges while training for PINNs in general.

In the present work, the MB-PINN and MB-IBM-PINN models are specifically evaluated for velocity reconstruction and pressure field recovery as a hidden variable. To improve the model predictions under a fixed budget, a multi-part physics loss weighting strategy has been adopted. It was discussed how, in the case where the solid body position and velocity is known a priori, It has been observed that the MB-PINN approach is quite reliable where the solid body data points are discarded. However, MB-IBM-PINN with a suitable relative weighting of the solid region residuals, can also perform equally well in terms of the data-driven loss components. MB-PINN out-performs MB-IBM-PINN in terms of computational time per iteration due to a relatively larger computational graph in the case of the latter. Given a fixed computational training budget, it has also been explored if the localised errors corresponding to the strong flow-field gradients in the vicinity of the solid body can be minimised and if the error minimisation is possible with fewer data points. Towards this, a physics-based vorticity cut-off sampling strategy is adopted since the locations with strong vorticity levels are indicators of strong gradients. Despite the test data set being generated from a completely different CFD solver, the best proposed MB-PINN and MB-IBMPINN models exhibit remarkable performance in pressure recovery. It is also observed that an accurate pressure recovery is obtained irrespective of the decrease in the overall bulk data points in the training data set. Hence, it is envisioned that when bulk data is available, a combination of loss components weighting with a physics based vorticity cut-off undersampling should be employed.

Although the strategies presented here require hand tuning, recent advancements in the PINNs such as adaptive training, residual collocation sampling and loss balancing methodologies could be employed in conjunction. This would limit human intervention during training while also reducing the number of runs, making the surrogate framework scalable. It would also be of interest to consider flow or kinematic parameters as additional network inputs and test the generalisability of the model on some unseen parametric instances enabling parametric differentiable surrogates.

References

  1. [1]J. Lai, M. Platzer, Jet characteristics of a plunging airfoil, AIAA journal 37 (12) (1999) 1529–1537.DOI
  2. [2]D. D. Chin, D. Lentink, Flapping wing aerodynamics: from insects to vertebrates, Journal of Experimental Biology 219 (7) (2016) 920–932.DOI
  3. [3]C. Bose, S. Gupta, S. Sarkar, Transition to chaos in the flow-induced vibration of a pitching–plunging airfoil at low reynolds numbers: Ruelle–Takens–Newhouse scenario, International Journal of Non-Linear Mechanics 109 (2019) 189–203.DOI
  4. [4]C. Xie, N. Gao, Y. Meng, Y. Wu, C. Yang, A review of bird-like flapping wing with high aspect ratio, Chinese Journal of Aeronautics (2022).DOI
  5. [5]Q. Xiao, Q. Zhu, A review on flow energy harvesters based on flapping foils, Journal of fluids and structures 46 (2014) 174–191.DOI
  6. [6]J. Sarrate, A. Huerta, J. Donea, Arbitrary Lagrangian–Eulerian formulation for fluid–rigid body interaction, Computer Methods in Applied Mechanics and Engineering 190 (24-25) (2001) 3171–3188.DOI
  7. [7]C. Farhat, P. Geuzaine, C. Grandmont, The discrete geometric conservation law and the nonlinear stability of ALE schemes for the solution of flow problems on moving grids, Journal of Computational Physics 174 (2) (2001) 669–694.DOI
  8. [8]C. S. Peskin, The immersed boundary method, Acta numerica 11 (2002) 479–517.
  9. [9]J. Kim, D. Kim, H. Choi, An immersed-boundary finite-volume method for simulations of flow in complex geometries, Journal of computational physics 171 (1) (2001) 132–150.DOI
  10. [10]R. Mittal, G. Iaccarino, Immersed boundary methods, Annu. Rev. Fluid Mech. 37 (2005) 239–261.
  11. [11]S. L. Brunton, B. R. Noack, P. Koumout sakos, Machine learning for fluid mechanics, Annual Review of Fluid Mechanics 52 (2020) 477–508.DOI
  12. [12]M. Raissi, A. Yazdani, G. E. Karniadakis, Hidden fluid mechanics: A Navier-Stokes informed deep learning framework for assimilating flow visualization data, arXiv preprint arXiv:1808.04327 (2018).arxiv.org/abs/1808.04327
  13. [13]S. Pawar, S. E. Ahmed, O. San, A. Rasheed, Data-driven recovery of hidden physics in reduced order modeling of fluid flows, Physics of Fluids 32 (3) (2020) 036602.
  14. [14]T. L. Daniel, Unsteady aspects of aquatic locomotion, American Zoologist 24 (1) (1984) 121–134.DOI
  15. [15]J. O. Dabiri, S. Bose, B. J. Gemmell, S. P. Colin, J. H. Costello, An algorithm to estimate unsteady and quasi-steady pressure fields from velocity field measurements, Journal of Experimental Biology 217 (3) (2014) 331–336.DOI
  16. [16]X. Liu, J. Katz, Vortex-corner interactions in a cavity shear layer elucidated by time-resolved measurements of the pressure field, Journal of Fluid Mechanics 728 (2013) 417–457.DOI
  17. [17]A. Pirnia, J. McClure, S. Peterson, B. Helenbrook, B. Erath, Estimating pressure fields from planar velocity data around immersed bodies; a finite element approach, Experiments in Fluids 61 (2) (2020) 1–16.DOI
  18. [18]J. Lee, J. Kim, H. Choi, K.-S. Yang, Sources of spurious force oscillations from an immersed boundary method for moving-body problems, Journal of computational physics 230 (7) (2011) 2677–2695.
  19. [19]D. Majumdar, C. Bose, S. Sarkar, Capturing the dynamical transitions in the flow-field of a flapping foil using immersed boundary method, Journal of Fluids and Structures 95 (2020) 102999.DOI
  20. [20]M. Balajewicz, C. Farhat, Reduction of nonlinear embedded boundary models for problems with evolving interfaces, Journal of Computational Physics 274 (2014) 489–504.DOI
  21. [21]W. Kim, H. Choi, Immersed boundary methods for fluid-structure interaction: A review, International Journal of Heat and Fluid Flow 75 (2019) 301–309.DOI
  22. [22]R. Vinuesa, S. L. Brunton, Enhancing computational fluid dynamics with machine learning, Nature Computational Science 2 (6) (2022) 358–366.DOI
  23. [23]R. Vinuesa, S. Brunton, Emerging trends in machine learning for computational fluid dynamics, arXiv preprint arXiv:2211.15145 (2022).arxiv.org/abs/2211.15145
  24. [24]S. Cheng, C. Quilodran-Casas, S. Ouala, A. Farchi, C. Liu, P. Tandeo, R. Fablet, D. Lucor, B. Iooss, J. Brajard, D. Xiao, T. Janjic, W. Ding, Y. Guo, A. Carrassi, M. Bocquet, R. Arcucci, Machine learning with data assimilation and uncertainty quantification for dynamical systems: a review, IEEE/CAA J. Autom. Sinica 10 (6) (2023).arxiv.org/abs/2303.10462
  25. [25]J. Willard, X. Jia, S. Xu, M. Steinbach, V. Kumar, Integrating scientific knowledge with machine learning for engineering and environmental systems, ACM Computing Surveys 55 (4) (2022) 1–37.arxiv.org/abs/2003.04919
  26. [26]G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, L. Yang, Physics-informed machine learning, Nature Reviews Physics 3 (6) (2021) 422–440.
  27. [27]S. Cuomo, V. S. Di Cola, F. Giampaolo, G. Rozza, M. Raissi, F. Piccialli, Scientific machine learning through physics-informed neural networks: where we are and what’s next, Journal of Scientific Computing 92 (3) (2022) 88.arxiv.org/abs/2201.05624
  28. [28]S. El Garroussi, S. Ricci, M. De Lozzo, N. Goutal, D. Lucor, Tackling random fields non-linearities with unsupervised clustering of polynomial chaos expansion in latent space: application to global sensitivity analysis of river flooding, Stochastic Environmental Research and Risk Assessment 36 (3) (2022) 693–718.DOI
  29. [29]B. X. Nony, M. C. Rochoux, T. Jaravel, D. Lucor, Reduced-order modeling for parameterized large-eddy simulations of atmospheric pollutant dispersion, Stochastic Environmental Research and Risk Assessment 37 (6) (2023) 2117–2144.arxiv.org/abs/2208.01518
  30. [30]M. Raissi, P. Perdikaris, G. E. Karniadakis, Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, Journal of Computational Physics 378 (2019) 686–707.DOI
  31. [31]I. Goodfellow, Y. Bengio, A. Courville, Deep learning, MIT press, 2016.
  32. [32]M. Raissi, Z. Wang, M. S. Triantafyllou, G. E. Karniadakis, Deep learning of vortex-induced vibrations, Journal of Fluid Mechanics 861 (2019) 119–137.DOI
  33. [33]M. Raissi, A. Yazdani, G. E. Karniadakis, Hidden fluid mechanics: Learning velocity and pressure fields from flow visualizations, Science 367 (6481) (2020) 1026–1030.DOI
  34. [34]A. G. Baydin, B. A. Pearlmutter, A. A. Radul, J. M. Siskind, Automatic differentiation in machine learning: a survey, The Journal of Machine Learning Research 18 (1) (2017) 5595–5637.
  35. [35]M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin, S. Ghemawat, G. Irving, M. Isard, et al., Tensorflow: A system for large-scale machine learning, 2016, pp. 265–283.arxiv.org/abs/1605.08695
  36. [36]A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, et al., Pytorch: An imperative style, high-performance deep learning library, Advances in neural information processing systems 32 (2019).arxiv.org/abs/1912.01703
  37. [37]J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. VanderPlas, S. Wanderman-Milne, Q. Zhang, JAX: composable transformations of Python+NumPy programs (2018). URL http://github.com/google/jax
  38. [38]S. Wang, Y. Teng, P. Perdikaris, Understanding and mitigating gradient flow pathologies in physics-informed neural networks, SIAM Journal on Scientific Computing 43 (5) (2021) A3055–A3081.arxiv.org/abs/2001.04536
  39. [39]A. Krishnapriyan, A. Gholami, S. Zhe, R. Kirby, M. W. Mahoney, Characterizing possible failure modes in physics-informed neural networks, Advances in Neural Information Processing Systems 34 (2021) 26548–26560.arxiv.org/abs/2109.01050
  40. [40]X. Jin, S. Cai, H. Li, G. E. Karniadakis, NSFnets (Navier-Stokes flow nets): Physics-informed neural networks for the incompressible Navier-Stokes equations, Journal of Computational Physics 426 (2021) 109951.DOI
  41. [41]C. Wu, M. Zhu, Q. Tan, Y. Kartha, L. Lu, A comprehensive study of non-adaptive and residual-based adaptive sampling for physics-informed neural networks, arXiv preprint arXiv:2207.10289 (2022).DOI
  42. [42]A. D. Jagtap, E. Kharazmi, G. E. Karniadakis, Conservative physics-informed neural networks on discrete domains for conservation laws: Applications to forward and inverse problems, Computer Methods in Applied Mechanics and Engineering 365 (2020) 113028.DOI
  43. [43]M. Penwarden, A. D. Jagtap, S. Zhe, G. E. Karniadakis, R. M. Kirby, A unified scalable framework for causal sweeping strategies for physics-informed neural networks (pinns) and their temporal decompositions, arXiv preprint arXiv:2302.14227 (2023).DOI
  44. [44]K. Shukla, M. Xu, N. Trask, G. E. Karniadakis, Scalable algorithms for physics-informed neural and graph networks, Data-Centric Engineering 3 (2022) e24.arxiv.org/abs/2205.08332
  45. [45]V. Gopakumar, S. Pamela, D. Samaddar, Loss landscape engineering via data regulation on pinns, arXiv preprint arXiv:2205.07843 (2022).arxiv.org/abs/2205.07843
  46. [46]D. Lucor, A. Agrawal, A. Sergent, Simple computational strategies for more effective physics-informed neural networks modeling of turbulent natural convection, Journal of Computational Physics 456 (2022) 111022.DOI
  47. [47]S. Wang, P. Perdikaris, Deep learning of free boundary and stefan problems, Journal of Computational Physics 428 (2021) 109914.DOI
  48. [48]A. B. Buhendwa, S. Adami, N. A. Adams, Inferring incompressible two-phase flow fields from the interface motion using physics-informed neural networks, Machine Learning with Applications 4 (2021) 100029.arxiv.org/abs/2101.09833
  49. [49]Q. Yang, Y. Yang, T. Cui, Q. He, FDM-PINN: Physics-informed neural network based on fictitious domain method, International Journal of Computer Mathematics (just-accepted) (2021) 1.DOI
  50. [50]X.-D. Bai, W. Zhang, Machine learning for vortex induced vibration in turbulent flow, Computers & Fluids 235 (2022) 105266.arxiv.org/abs/2103.05818
  51. [51]H. Tang, Y. Liao, H. Yang, L. Xie, A transfer learning-physics-informed neural network (TL-PINN) for vortex-induced vibration, Ocean Engineering 266 (2022) 113101.DOI
  52. [52]Y. Huang, Z. Zhang, X. Zhang, A direct-forcing immersed boundary method for incompressible flows based on physics-informed neural network, Fluids 7 (2) (2022) 56.DOI
  53. [53]M. A. Calicchia, R. Mittal, J.-H. Seo, R. Ni, Reconstructing the pressure field around swimming fish using a physics-informed neural network, Journal of Experimental Biology 226 (8) (2023) jeb244983.
  54. [54]G. C. Lewin, H. Haj-Hariri, Modelling thrust generation of a two-dimensional heaving airfoil in a viscous flow, Journal of Fluid Mechanics 492 (2003) 339.DOI
  55. [55]M. S. U. Khalid, I. Akhtar, H. Dong, N. Ahsan, X. Jiang, B. Wu, Bifurcations and route to chaos for flow over an oscillating airfoil, Journal of Fluids and Structures 80 (2018) 262–274.DOI
  56. [56]C. L. Shah, D. Majumdar, S. Sarkar, Performance enhancement of an immersed boundary method based FSI solver using OpenMP, in: Annual CFD Symposium, NAL, Bangalore, 2019.
  57. [57]H. Jasak, A. Jemcov, Z. Tukovic, et al., OpenFOAM: A c++ library for complex physics simulations, in: International workshop on coupled methods in numerical dynamics, Vol. 1000, IUC Dubrovnik Croatia, 2007, pp. 1–20.
  58. [58]S. Cai, Z. Mao, Z. Wang, M. Yin, G. E. Karniadakis, Physics-informed neural networks (PINNs) for fluid mechanics: A review, Acta Mechanica Sinica (2022) 1–12.arxiv.org/abs/2105.09506
  59. [59]D. P. Kingma, J. Ba, ADAM: A method for stochastic optimization (2014). arXiv:1412.6980.arxiv.org/abs/1412.6980
  60. [60]A. A. Heydari, C. A. Thompson, A. Mehmood, SoftAdapt: Techniques for adaptive loss weighting of neural networks with multi-part loss functions, arXiv preprint arXiv:1912.12355 (2019).arxiv.org/abs/1912.12355
  61. [61]R. Bischof, M. Kraus, Multi-objective loss balancing for physics-informed deep learning, arXiv preprint arXiv:2110.09813 (2021).DOI
  62. [62]M. F. Platzer, K. D. Jones, J. Young, J. C. Lai, Flapping wing aerodynamics: progress and challenges, AIAA journal 46 (9) (2008) 2136–2149.
  63. [63]C. Bose, S. Sarkar, Investigating chaotic wake dynamics past a flapping airfoil and the role of vortex interactions behind the chaotic transition, Physics of fluids 30 (4) (2018) 047101.DOI
  64. [64]D. Majumdar, C. Bose, S. Sarkar, Transition boundaries and an order-to-chaos map for the flow field past a flapping foil, Journal of Fluid Mechanics 942 (2022) A40.DOI
  65. [65]K. Fukami, K. Fukagata, K. Taira, Super-resolution reconstruction of turbulent flows with machine learning, Journal of Fluid Mechanics 870 (2019) 106–120.DOI
  66. [66]T. N. K. Nguyen, T. Dairay, R. Meunier, C. Millet, M. Mougeot, Fixed-budget online adaptive mesh learning for physics-informed neural networks. towards parameterized problem inference, arXiv preprint arXiv:2212.11776 (2022).DOI
  67. [67]F. F. Weller, A free boundary problem modeling thrombus growth: model development and numerical simulation using the level set method, Journal of mathematical biology 61 (2010) 805–818.DOI
  68. [68]Z. Gao, L. Yan, T. Zhou, Failure-informed adaptive sampling for pinns, arXiv preprint arXiv:2210.00279 (2022).arxiv.org/abs/2210.00279
  69. [69]W. Peng, W. Zhou, X. Zhang, W. Yao, Z. Liu, RANG: A residual-based adaptive node generation method for physics-informed neural networks, arXiv preprint arXiv:2205.01051 (2022).arxiv.org/abs/2205.01051
  70. [70]K. Tang, X. Wan, C. Yang, DAS-PINNs: A deep adaptive sampling method for solving high-dimensional partial differential equations, Journal of Computational Physics 476 (2023) 111868.arxiv.org/abs/2112.14038

Paper details

vlnc:W_mvgxh584
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents