betaFeedback

WorksPhysics and Astronomy

Determination of Navier's slip parameter using data assimilation

Abstract

One of the crucial aspects of patient-specific blood flow simulations is to specify material parameters and boundary conditions. The choice of boundary conditions can have a substantial impact on the character of the flow. While no-slip is the most popular wall boundary condition, some amount of slip, which determines how much fluid is allowed to flow along the wall, might be beneficial for better agreement with flow patterns in medical images. However, even if one assumes the simple Navier's boundary conditions on the wall, in which the relationship between tangential components of the normal traction and the velocity is linear, the determination of the specific value of the slip parameter is often difficult. In this work, we present and test an optimal control method to estimate Navier's slip parameter on the wall and the velocity profile at the inlet using artificially generated flow domain and flow data. The results show that it is possible to recover the flow patterns and Navier's slip parameter by using sufficiently accurate discretization even from data containing a substantial amount of noise.

Acknowledgements The work was supported by the Czech Science Foundation, project 23-05207S. In addition, Alena Jarolímová was supported by the Grant Agency of Charles University, grant nr. 282222 and project SVV-2023-260711. We thank Josef Málek and Patrick Farrell for their help and insightful discussions.

Introduction

Performing patient-specific simulations can be a helpful tool for diagnostic purposes in clinical practice. However, the choice of an appropriate model can have crucial consequences on the results of the simulation. More specifically, it is necessary to choose all the parameters in the model precisely enough to reflect the complex biological processes of a particular patient. Since some of the model parameters are patient-specific and hard or even impossible to measure directly, one can use data from medical imaging to compensate for the lack of such information.

The phase-contrast magnetic resonance imaging (PC MRI) technique can provide 4D data of the velocity field (3D space + time). However, its spatial and temporal resolution is limited compared to the resolution of computational fluid simulations, and it also contains a substantial amount of noise. Therefore, using such data to determine the unknown quantities directly may be unreliable. Instead, it is possible to use the data to recover the model parameters by solving an inverse problem.

Several data assimilation methods have been used to solve inverse problems in cardiovascular settings, some of which are discussed, for example, in [6]. Kalman filter based approach, which is usually used to estimate lumped parameters of the model as done in [29, 31], represents one of the popular approaches. Another option, which we follow in this work, is to reformulate the problem in an optimal control setting, minimizing an error functional constrained to the initial- and boundary-value problem governed by a system of partial differential equations (PDE), describing steady or unsteady fluid flows. For example, various options for solving optimal control problems were discussed in [19] and [21]. A PDE-constrained optimization can be reformulated as one large system expressing the Karush–Kuhn–Tucker optimality conditions (KKT). The complete KKT system based on a steady-state flow problem is solved using Picard’s and Newton’s method in [13]. In recent years, sequential quadratic programming algorithms gained popularity for solving the KKT systems in cardiovascular settings [18, 17] [33]. The KKT system for the steady state can also be solved iteratively as done in [23]. For time-dependent problems, the KKT system is usually transformed into an iterative process, see for example [15] or [24].

A crucial part of any data assimilation method is the choice of a fluid flow model. While the choice of the Navier–Stokes fluid model seems reasonable for the description of flow in aortic vessels (as non-Newtonian variants where the viscosity depends on the shear rate have rather minor effects), it is known, see for example [22, 11], that the choice of boundary conditions can substantially impact the blood flow patterns. Even though the no-slip boundary condition is the most popular wall boundary condition in the cardiovascular modelling community, some partial slip along the wall can be observed in certain experiments [30]. Such choice of boundary condition can also partially compensate the errors caused by the inaccuracy of image segmentation, which has been addressed in [29] using a special choice of boundary condition and in [24] by treating the shape of the geometry as an unknown determined using the optimal control.

We are now in the position to formulate our basic question motivating this study: Assuming that the Navier–Stokes fluid slips on the wall according to Navier’s slip boundary condition, can we determine the slip parameter using 4D MRI data? In addition, can we also use the same set of 4D MRI data to determine the velocity of the inlet? In this context, it is worth mentioning that, in their recent study [25], Málek and Rajagopal found criteria (expressed in terms of macroscopic quantities such as the pressure gradient, the volumetric flow rate, the viscosity, the density and geometric dimensions) stating/determining when the Navier–Stokes fluid in a five simple unidirectional flows exhibits no-slip. Even more, in their following study [26], assuming that the flow rate is bigger than the ”critical” flow rate corresponding to the no-slip boundary condition and assuming Navier’s slip boundary condition, they developed a methodology to determine the value of the slip parameter.

In this work, we test the performance of an adjoint-based variational assimilation method for 4D PC-MRI data. We aim to assess whether Navier’s slip boundary condition is more suitable for modelling blood in large arteries. Moreover, we present numerical experiments to verify the reliability of the approach. In particular, the developed code is able to determine the slip parameter for simple flows in accordance with the methodology developed in [26]. For the sake of simplicity, we restrict ourselves to steady flows. Our main objective is to develop an efficient method supported by its careful testing that could have the potential to be used on actual patient-specific data and geometry.

The structure of this work is the following. In § 2, we present our approach based on a general optimal control methodology as described in [6]. We also formulate the underlying boundary-value problem governed by the Navier-Stokes equations. The setup of the numerical experiments is presented in § 3, and the results are discussed in § 4. § 5 consists of a conclusion and discussion of future work.

Methodology

Fluid flow model

Since we plan to apply the method to blood flow in large arteries, we consider a tubular domain Ω⊂R3\Omega\subset\mathbb{R}^{3} representing an artery containing no bifurcations. The boundary of the domain can be split into three parts: inlet Γin\Gamma_{\mathrm{in}}, outlet Γout\Gamma_{\mathrm{out}} and wall Γwall\Gamma_{\mathrm{wall}}, see Figure 1.

Image not extracted

Figure 1. A scheme of the domain with the boundary types.

We consider the steady-state incompressible Navier–Stokes equations for the velocity v:Ω→R3\mathbf{v} : \Omega\to\mathbb{R}^{3} and the pressure p:Ω→Rp : \Omega\to\mathbb{R}. A Dirichlet boundary condition is prescribed at the inlet Γin\Gamma_{\mathrm{in}} using a given function vin:Γin→R3\mathbf{v}_{\mathrm{in}} : \Gamma_{\mathrm{in}} \to\mathbb{R}^{3} and a directional do-nothing boundary condition is prescribed at the outlet Γout\Gamma_{\mathrm{out}}. At the wall Γwall\Gamma_{\mathrm{wall}}, Navier’s slip boundary condition is used to describe the behaviour of the fluid in the tangential direction along the wall, accompanied by the impermeability condition in the normal direction. Together, the governing equations describing the underlying boundary-value problem read

ρ(∇v)v−div⁡T(v,p)=0in Ω,T(v,p)=−pI+2μD(v)in Ω,div⁡v=0in Ω,v=vinon Γin,(T(v,p)n)⋅n=12ρ(v⋅n)2andvt=0on Γout,θvt+γ∗(1−θ)(T(v,p)n)t=0andv⋅n=0on Γwall.(1)\begin{aligned} \rho(\nabla\mathbf{v})\mathbf{v} - \operatorname{div}\mathbb{T}(\mathbf{v},p) &= 0 &&\text{in }\Omega,\\ \mathbb{T}(\mathbf{v},p) &= -p\mathbf{I} + 2\mu\mathbb{D}(\mathbf{v}) &&\text{in }\Omega,\\ \operatorname{div}\mathbf{v} &= 0 &&\text{in }\Omega,\\ \mathbf{v} &= \mathbf{v}_{\mathrm{in}} &&\text{on }\Gamma_{\mathrm{in}},\\ (\mathbb{T}(\mathbf{v},p)\mathbf{n})\cdot\mathbf{n} &= \frac{1}{2}\rho(\mathbf{v}\cdot\mathbf{n})^{2} &&\text{and}\quad\mathbf{v}_{t}=0\quad\text{on }\Gamma_{\mathrm{out}},\\ \theta\mathbf{v}_{t}+\gamma_{*}(1-\theta)(\mathbb{T}(\mathbf{v},p)\mathbf{n})_{t} &= 0 &&\text{and}\quad\mathbf{v}\cdot\mathbf{n}=0\quad\text{on }\Gamma_{\mathrm{wall}}. \tag*{(1)} \end{aligned}

where T\mathbb{T} is the Cauchy stress tensor, D(v)=12(∇v+(∇v)T)\mathbb{D}(\mathbf{v}) = \frac{1}{2}(\nabla\mathbf{v}+(\nabla\mathbf{v})^{T}) is the symmetric part of the velocity gradient, ρ>0\rho> 0 and μ>0\mu> 0 are constant density and viscosity prescribed to match the properties of human blood and n\mathbf{n} is the vector normal to a given part of the boundary. Moreover, for x∈Rx \in\mathbb{R}, we denote the positive and negative part by x+=max⁡{x,0}x_{+}=\max\{x,0\} and x−=min⁡{x,0}x_{-}=\min\{x,0\} respectively. Similarly, for u∈R3\mathbf{u} \in\mathbb{R}^{3} we denote the normal component un=(u⋅n)n\mathbf{u}_{n}=(\mathbf{u}\cdot\mathbf{n})\mathbf{n} and the tangential component ut=u−un\mathbf{u}_{t}=\mathbf{u}-\mathbf{u}_{n}. The parameter θ∈[0,1]\theta\in[0,1] defines how much slip can be permitted along the wall, where θ=0\theta=0 and θ=1\theta=1 correspond to free slip and no-slip, respectively. The parameter γ∗∈(0,∞)\gamma_{*}\in(0,\infty) is set experimentally, and its optimal choice has been discussed in [10].

The value of θ\theta and the velocity profile vin\mathbf{v}_{\mathrm{in}} defining the flow at the inlet of the computational geometry are unknown patient-specific parameters which we want to estimate. Let us denote them as m=(θ,vin)\mathbf{m}=(\theta,\mathbf{v}_{\mathrm{in}}). In the context of data assimilation, the unknown parameters m\mathbf{m} are often called control variables. Using this simplified notation, we can denote the equations defined in (1) by F(w,m)=0\mathbf{F}(\mathbf{w},\mathbf{m})=0, where w=(v,p)\mathbf{w}=(\mathbf{v},p) is the solution for given control variables m=(θ,vin)\mathbf{m}=(\theta,\mathbf{v}_{\mathrm{in}}).

Variational data assimilation

The model contains several parameters that need to be provided to solve the system (1) for a specific patient. It is possible to use tabulated values for density ρ\rho and viscosity μ\mu of human blood or even perform experiments to obtain more accurate values for the given patient. On the other hand, the inlet velocity vin\mathbf{v}_{\mathrm{in}} and the slip parameter θ\theta have to be determined by solving an inverse problem instead, as illustrated in Figure 2.

Diagram of the difference between forward and inverse problem

Figure 2. A diagram of the difference between forward and inverse problem

The variational assimilation is a technique to determine unknown parameters of a model from physical measurements. For real patient-specific cases, we want to use 4D phase-contrast magnetic resonance images (4D PC-MRI), which contain information about the velocity field in 3D space and time. The data received using this technique have low resolution (around 2 mm) and contain substantial noise. Since we are solving a steady problem, we would use only a single temporal slice of the 4D PC-MRI data. Therefore, we obtain a vector-valued function defined over the whole domain Ω\Omega, which we denote by dMRI∈(L∞(Ω))3\mathbf{d}_{\mathrm{MRI}} \in(L^{\infty}(\Omega))^3. However, in this work, we will use only artificially generated data to be able to compare the results with the ground truth. These artificially generated data are obtained by solving (1) with a priori chosen data θ\theta and vin\mathbf{v}_{\mathrm{in}} and then extrapolated to a coarser mesh and convoluted with random noise.

The governing boundary-value problem F(w,m)=0F(\mathbf{w},\mathbf{m}) = 0, see equation (1), is used to describe the physical process of the flow. Since we want to compare the solution w\mathbf{w} to the measured data dMRI\mathbf{d}_{\mathrm{MRI}}, we define an operator T\mathcal{T}, which describes the measurement process. Then, we can apply the operator to w\mathbf{w} to obtain simulated measurements T(w)\mathcal{T}(\mathbf{w}) as shown in Figure 3.

Image not extracted

Figure 3. A scheme of variational data assimilation

The main idea of variational assimilation is to minimize a certain distance of the simulated measurements T(w)\mathcal{T}(\mathbf{w}) from the actual measurements dMRI\mathbf{d}_{\mathrm{MRI}} with respect to the unknown model parameters m\mathbf{m}. The governing equations defined by F(w,m)=0F(\mathbf{w},\mathbf{m}) = 0 act as a constraint to the optimization establishing the relation between m\mathbf{m} and w\mathbf{w}. Let us denote the error functional as

J(w)=12ℓ3∥T(w)−dMRI∥L2(Ω)2,(2)\mathcal{J}(\mathbf{w}) = \frac{1}{2\ell^3}\left\lVert\mathcal{T}(\mathbf{w})-\mathbf{d}_{\mathrm{MRI}}\right\rVert_{L^2(\Omega)}^2, \tag*{(2)}

where ℓ\ell denotes a characteristic length of the problem geometry.1 The optimization problem then can be stated as

min⁡(w,m)∈AJ(w)where A:={(w,m)∈W×M;F(w,m)=0},(3)\min_{(\mathbf{w},\mathbf{m})\in\mathcal{A}} \mathcal{J}(\mathbf{w}) \quad\text{where } \mathcal{A}:=\{(\mathbf{w},\mathbf{m})\in W\times M;F(\mathbf{w},\mathbf{m})=0\}, \tag*{(3)}

for W=[H1(Ω)]3×L2(Ω)W=[H^{1}(\Omega)]^{3}\times L^{2}(\Omega) and M=[0,1]×H1(Γin)M=[0,1]\times H^{1}(\Gamma_{\mathrm{in}}). In other words, we need to solve a PDE-constrained optimization problem.

Since we are working with a steady case, we chose the measurement operator simply as

T(w)=v.(4)\mathcal{T}(\mathbf{w})=\mathbf{v}. \tag*{(4)}

The error functional J(w)\mathcal{J}(\mathbf{w}) is often combined with a Tikhonov regularization R(m)\mathcal{R}(\mathbf{m}):

JR(w,m)=J(w)+R(m).(5)\mathcal{J}_{R}(\mathbf{w},\mathbf{m})=\mathcal{J}(\mathbf{w})+\mathcal{R}(\mathbf{m}). \tag*{(5)}

The regularization of a given parameter is usually a norm of the parameter or its distance from some prescribed value. A correct choice of regularization can help the parameters to be in the desired function space and make the problem well-posed. It also helps to resolve the situation where multiple solutions would be viable otherwise.

In our approach, the regularization R(m)\mathcal{R}(\mathbf{m}) consists of two terms:

R(m)=Rα(m)+Rβ(m).(6)\mathcal{R}(\mathbf{m})=\mathcal{R}_{\alpha}(\mathbf{m})+\mathcal{R}_{\beta}(\mathbf{m}). \tag*{(6)}

The first term is designed to penalize the gradient of inlet velocity vin\mathbf{v}_{\mathrm{in}} in the tangential direction of the inlet, and its purpose is to encourage the algorithm to select a smooth inlet velocity profile, i.e.,

Rα(m)=α2∥∇vin∥L2(Γin)2.(7)\mathcal{R}_{\alpha}(\mathbf{m})=\frac{\alpha}{2}\lVert\nabla\mathbf{v}_{\mathrm{in}}\rVert_{L^{2}(\Gamma_{\mathrm{in}})}^{2}. \tag*{(7)}

The second term also corresponds to vin\mathbf{v}_{\rm in}, and it is defined as a L2L^2 norm of the distance of vin\mathbf{v}_{\rm in} from a given velocity profile vanalytic\mathbf{v}_{\rm analytic} over the inlet boundary Γin\Gamma_{\rm in}.

Rβ(m)=β2ℓ2∥vin−vanalytic(V,θ)∥L2(Γin)2(8)\mathcal{R}_{\beta}(\mathbf{m})=\frac{\beta}{2\ell^2}\left\|\mathbf{v}_{\rm in}-\mathbf{v}_{\rm analytic}(V,\theta)\right\|_{L^2(\Gamma_{\rm in})}^2 \tag*{(8)}

The velocity profile vanalytic\mathbf{v}_{\rm analytic} is the analytical solution to the steady flow in a straight tube with a given average inflow velocity VV and Navier’s slip on the walls given by parameter θ\theta, see [22], and it represents our estimate of how the inlet velocity profile might look like:

vanalytic(V,θ)=V4μγ∗(1−θ)R+2θ(R2−r2)4μγ∗(1−θ)R+θR2n.(9)\mathbf{v}_{\rm analytic}(V,\theta)=V\frac{4\mu\gamma_*(1-\theta)R+2\theta(R^2-r^2)}{4\mu\gamma_*(1-\theta)R+\theta R^2}\mathbf{n}. \tag*{(9)}

Here, RR is the radius of the inlet, rr is the distance from the centre axis, and VV is computed from dMRI\mathrm{dMRI} as the average inflow through the inlet. The motivation behind Rβ\mathcal{R}_{\beta} is to increase the link between θ\theta and vin\mathbf{v}_{\rm in} in the optimization process. It also acts as a counterweight to Rα\mathcal{R}_{\alpha}, which enforces vin\mathbf{v}_{\rm in} to be a constant function.

Substituting the relation (2) for J\mathcal{J} and (7)-(8) for R\mathcal{R} into the definition of the regularized error functional (5) then gives us

JR((v,p),(θ,vin))=12ℓ3∥v−dMRI∥L2(Ω)2+α2∥∇vin∥L2(Γin)2+β2ℓ2∥vin−vanalytic(V,θ)∥L2(Γin)2(10)\mathcal{J}_{\mathcal{R}}((\mathbf{v},p),(\theta,\mathbf{v}_{\rm in}))=\frac{1}{2\ell^3}\left\|\mathbf{v}-\mathbf{d}_{\rm MRI}\right\|_{L^2(\Omega)}^2+\frac{\alpha}{2}\left\|\nabla\mathbf{v}_{\rm in}\right\|_{L^2(\Gamma_{\rm in})}^2 +\frac{\beta}{2\ell^2}\left\|\mathbf{v}_{\rm in}-\mathbf{v}_{\rm analytic}(V,\theta)\right\|_{L^2(\Gamma_{\rm in})}^2 \tag*{(10)}

The weights α\alpha and β\beta have to be selected carefully so that the regularization terms R(m)\mathcal{R}(\mathbf{m}) in (10) are in good balance with the first term J(w)\mathcal{J}(\mathbf{w}). The choice of the weights α\alpha and β\beta is discussed in more detail in section 4.4.

Numerical solution

In order to solve the PDE-constrained optimization problem

min⁡(w,m)∈AJR(w,m)where A:={(w,m)∈W×M;F(w,m)=0},(11)\min_{(\mathbf{w},\mathbf{m})\in\mathcal{A}}\mathcal{J}_{\mathcal{R}}(\mathbf{w},\mathbf{m}) \quad\text{where } \mathcal{A}:=\{(\mathbf{w},\mathbf{m})\in W\times M; F(\mathbf{w},\mathbf{m})=0\}, \tag*{(11)}

the constraint can be removed by enforcing it implicitly in the functional. Since w\mathbf{w} is the solution of model F(w,m)=0F(\mathbf{w},\mathbf{m})=0, w\mathbf{w} can be considered an implicit function of the control variables as w(m)\mathbf{w}(\mathbf{m}), assuming the implicit function theorem applies. Therefore, we can define a reduced functional as

J^R(m):=JR(w(m),m)=J(w(m))+R(m).(12)\widehat{\mathcal{J}}_{\mathcal{R}}(\mathbf{m}):=\mathcal{J}_{\mathcal{R}}(\mathbf{w}(\mathbf{m}),\mathbf{m})=\mathcal{J}(\mathbf{w}(\mathbf{m}))+\mathcal{R}(\mathbf{m}). \tag*{(12)}

The reduced functional J^R(m)\widehat{\mathcal{J}}_{\mathcal{R}}(\mathbf{m}) already contains the PDE constraint implicitly, which means that the original problem (11) is equivalent to an unconstrained optimization problem

min⁡m∈AJ^R(m).(13)\min_{\mathbf{m}\in\mathcal{A}}\widehat{\mathcal{J}}_{\mathcal{R}}(\mathbf{m}). \tag*{(13)}

There are many iterative algorithms to solve an unconstrained optimization problem. Since the model F(w,m)=0F(\mathbf{w},\mathbf{m})=0 has to be solved in each iteration to evaluate the reduced functional J^R(w(mi),mi)\widehat{J}_{R}(\mathbf{w}(\mathbf{m}^{i}),\mathbf{m}^{i}), it is reasonable to choose an algorithm that can minimize the error functional in an efficient way. Therefore, a gradient-based optimization algorithm is used since it is known to converge in fewer iterations than gradient-free algorithms. The iterative process can then be summarized as follows:

Algorithm 1: Optimization iterative process

Data: m0\mathbf{m}^{0}

for i=0,1,2…i=0,1,2\ldots do

    if stopping criterion is fulfilled then

        stop the optimization

    else

        find w(mi)\mathbf{w}(\mathbf{m}^{i}) by solving the PDE equations with mi\mathbf{m}^{i};

        evaluate J^R(mi)=JR(w(mi),mi)\widehat{J}_{R}(\mathbf{m}^{i})=J_{R}(\mathbf{w}(\mathbf{m}^{i}),\mathbf{m}^{i});

        compute ∂J^R∂m(mi)\frac{\partial\widehat{J}_{R}}{\partial\mathbf{m}}(\mathbf{m}^{i});

        determine mi+1\mathbf{m}^{i+1} using the chosen optimization algorithm;

The optimization algorithm has to be able to fulfil the constraint on the control variable θ∈[0,1]\boldsymbol{\theta}\in[0,1]. Therefore, we chose the L-BFGS-B algorithm [9], which can support box constraints. The goal of the algorithm is to find a minimum of a nonlinear function f(x):Rn→Rf(\mathbf{x}):\mathbb{R}^{n}\to\mathbb{R} subject to l≤x≤u\mathbf{l}\leq\mathbf{x}\leq\mathbf{u}, where the vectors l\mathbf{l} and u\mathbf{u} are the lower and upper bound for the variable x\mathbf{x}, respectively. The algorithm is an extension of a quasi-Newtonian unconstrained optimization algorithm BFGS, an iterative method based on approximation of the Hessian matrix.

Each optimization iteration requires solving not only the PDE represented by F(m,w)=0F(\mathbf{m},\mathbf{w})=0 but also computing the gradient ∂J^R∂m(mi)\frac{\partial\widehat{J}_{R}}{\partial\mathbf{m}}(\mathbf{m}^{i}), which can be done using the adjoint equations. The first step in deriving the adjoint equations is to differentiate the definition of the reduced functional (12) with respect to the control variables m\mathbf{m}.

∂J^R∂m=∂JR∂w(∂w∂m)+∂JR∂m.(14)\frac{\partial\widehat{J}_{R}}{\partial\mathbf{m}}=\frac{\partial J_{R}}{\partial\mathbf{w}}\left(\frac{\partial\mathbf{w}}{\partial\mathbf{m}}\right)+\frac{\partial J_{R}}{\partial\mathbf{m}}. \tag*{(14)}

Since the operator w(m)\mathbf{w}(\mathbf{m}) has been defined implicitly using the governing equations F(m,w)=0F(\mathbf{m},\mathbf{w})=0, we cannot explicitly evaluate ∂w∂m\frac{\partial\mathbf{w}}{\partial\mathbf{m}}. However, we can differentiate the relation F(w(m),m)=0F(\mathbf{w}(\mathbf{m}),\mathbf{m})=0 to obtain a relation for ∂w∂m\frac{\partial\mathbf{w}}{\partial\mathbf{m}}:

∂F∂w(∂w∂m)+∂F∂m=0.(15)\frac{\partial F}{\partial\mathbf{w}}\left(\frac{\partial\mathbf{w}}{\partial\mathbf{m}}\right)+\frac{\partial F}{\partial\mathbf{m}}=0. \tag*{(15)}

Let us define an adjoint operator ∂F∗∂w\frac{\partial F^{*}}{\partial\mathbf{w}} as

⟨(∂F∂w)∗(λ),y⟩=⟨λ,∂F∂w(y)⟩∀λ,y∈Y,(16)\left\langle\left(\frac{\partial F}{\partial\mathbf{w}}\right)^{*}(\lambda),y\right\rangle=\left\langle\lambda,\frac{\partial F}{\partial\mathbf{w}}(y)\right\rangle\quad\forall\lambda,y\in Y, \tag*{(16)}

for the appropriate function space YY. Then, we consider an adjoint problem

⟨(∂F∂w)∗(λ),y⟩=∂JR∂w(y)∀y∈Y.(17)\left\langle\left(\frac{\partial F}{\partial\mathbf{w}}\right)^*(\lambda),\mathbf{y}\right\rangle= \frac{\partial J_R}{\partial\mathbf{w}}(\mathbf{y}) \quad\forall\mathbf{y} \in Y. \tag*{(17)}

for an adjoint variable λ∈Y\lambda\in Y. Combining relations (14), (17), the definition of the adjoint operator (16) and the relation (15), we arrive at

∂J^R∂m=∂JR∂w(∂w∂m)+∂JR∂m=⟨(∂F∂w)∗(λ),∂w∂m⟩+∂JR∂m=⟨λ,∂F∂w(∂w∂m)⟩+∂JR∂m=−⟨λ,∂F∂m⟩+∂JR∂m.(18)\begin{aligned} \frac{\partial\widehat{J}_R}{\partial\mathbf{m}} &= \frac{\partial J_R}{\partial\mathbf{w}}\left(\frac{\partial\mathbf{w}}{\partial\mathbf{m}}\right) + \frac{\partial J_R}{\partial\mathbf{m}} \\ &= \left\langle\left(\frac{\partial F}{\partial\mathbf{w}}\right)^*(\lambda),\frac{\partial\mathbf{w}}{\partial\mathbf{m}}\right\rangle+ \frac{\partial J_R}{\partial\mathbf{m}} \\ &= \left\langle\lambda,\frac{\partial F}{\partial\mathbf{w}}\left(\frac{\partial\mathbf{w}}{\partial\mathbf{m}}\right)\right\rangle+ \frac{\partial J_R}{\partial\mathbf{m}} \\ &= -\left\langle\lambda,\frac{\partial F}{\partial\mathbf{m}}\right\rangle+ \frac{\partial J_R}{\partial\mathbf{m}}. \tag*{(18)} \end{aligned}

To conclude, the computation of the gradient requires one solve of the adjoint problem (17). Then, the adjoint variable λ\lambda can be used to evaluate the gradient using the relation

∂J^R∂m=−⟨λ,∂F∂m⟩+∂JR∂m(19)\frac{\partial\widehat{J}_R}{\partial\mathbf{m}} = -\left\langle\lambda,\frac{\partial F}{\partial\mathbf{m}}\right\rangle+ \frac{\partial J_R}{\partial\mathbf{m}} \tag*{(19)}

derived in (18).

Weak formulation and finite element discretization

The problem (1) is solved using the finite element method. The domain Ω\Omega is approximated by a polyhedral domain Ωh\Omega_h. The impermeability condition on the wall and zero tangential part of the velocity condition at the outflow are implemented using the non-symmetric Nitsche method with penalization [7] [28]. The choice of the particular variant of the Nitsche method was discussed in [10]. Denoting β∗>0\beta_* > 0 as the penalization parameter and h>0h > 0 as the cell diameter of the mesh, the weak formulation reads

Find (v−vin∗,p)∈Vh×Ph such that∫Ωρ(∇v)v⋅ϕ dx+∫ΩT(v,p):∇ϕ dx+∫Ωqdiv⁡v dx+∫Γwallθγ∗(1−θ)vt⋅ϕt ds−∫Γwall(T(v,p)n)n⋅ϕn ds+∫Γwallvn⋅(T(ϕ,q)n)n ds+β∗μh∫Γwallvn⋅ϕn ds−∫Γout12ρ(v⋅n)−2 (ϕ⋅n) ds−∫Γout(T(v,p)n)t⋅ϕt ds+∫Γoutvt⋅(T(ϕ,q)n)t ds+β∗μh∫Γoutvt⋅ϕt ds=0(20)\begin{aligned} &\text{Find }(\mathbf{v}-\mathbf{v}_{\mathrm{in}}^{*},p)\in V_h\times P_h\text{ such that}\\ &\int_{\Omega}\rho(\nabla\mathbf{v})\mathbf{v}\cdot\boldsymbol{\phi}\,\mathrm{d}x+\int_{\Omega}\mathbf{T}(\mathbf{v},p):\nabla\boldsymbol{\phi}\,\mathrm{d}x+\int_{\Omega}q\operatorname{div}\mathbf{v}\,\mathrm{d}x\\ &\quad+\int_{\Gamma_{\mathrm{wall}}}\frac{\theta}{\gamma^{*}(1-\theta)}\mathbf{v}_{t}\cdot\boldsymbol{\phi}_{t}\,\mathrm{d}s-\int_{\Gamma_{\mathrm{wall}}}(\mathbf{T}(\mathbf{v},p)\mathbf{n})_{n}\cdot\phi_{n}\,\mathrm{d}s\\ &\quad+\int_{\Gamma_{\mathrm{wall}}}\mathbf{v}_{n}\cdot(\mathbf{T}(\boldsymbol{\phi},q)\mathbf{n})_{n}\,\mathrm{d}s+\frac{\beta_{*}\mu}{h}\int_{\Gamma_{\mathrm{wall}}}\mathbf{v}_{n}\cdot\phi_{n}\,\mathrm{d}s\\ &\quad-\int_{\Gamma_{\mathrm{out}}}\frac{1}{2}\rho(\mathbf{v}\cdot\mathbf{n})^{2}_{-}\,(\boldsymbol{\phi}\cdot\mathbf{n})\,\mathrm{d}s-\int_{\Gamma_{\mathrm{out}}}(\mathbf{T}(\mathbf{v},p)\mathbf{n})_{t}\cdot\boldsymbol{\phi}_{t}\,\mathrm{d}s\\ &\quad+\int_{\Gamma_{\mathrm{out}}}\mathbf{v}_{t}\cdot(\mathbf{T}(\boldsymbol{\phi},q)\mathbf{n})_{t}\,\mathrm{d}s+\frac{\beta_{*}\mu}{h}\int_{\Gamma_{\mathrm{out}}}\mathbf{v}_{t}\cdot\boldsymbol{\phi}_{t}\,\mathrm{d}s=0 \tag*{(20)} \end{aligned}

for all (ϕ,q)∈Vh×Ph(\boldsymbol{\phi},q)\in V_h\times P_h,

where Vh⊂{v∈H1(Ωh),v=0 on Γin}V_h\subset\{\mathbf{v}\in H^{1}(\Omega_h),\mathbf{v}=0\text{ on }\Gamma_{\mathrm{in}}\}, Ph⊂L2(Ωh)P_h\subset L^{2}(\Omega_h) and vin∗\mathbf{v}_{\mathrm{in}}^{*} is the finite element extension of vin\mathbf{v}_{\mathrm{in}} to the whole Ω‾h\overline{\Omega}_h. The unit boundary normal n\mathbf{n} is computed as the continuous vertex normal as defined in [10].

The discrete VhV_h, PhP_h spaces are defined using either the MINI or P1/P1P_1/P_1 finite elements. The advantage of these elements is that there are significantly fewer degrees of freedom than, for example, the Taylor–Hood element used in [10, 11]. On the other hand, the P1/P1P_1/P_1 element is not inf-sup stable by itself, so we have to add appropriate stabilization. In this case, we use the interior penalty stabilization [8], which is achieved by adding the following terms to the weak formulation:

αv∑K∈F∫Kρh2[∇v]⋅[∇ϕ] ds+αp∑K∈F∫Kh2ρ[∇p]⋅[∇q] ds,(21)\alpha_v\sum_{K\in\mathcal{F}}\int_K \rho h^2[\nabla\mathbf{v}]\cdot[\nabla\boldsymbol{\phi}]\,\mathrm{d}s+\alpha_p\sum_{K\in\mathcal{F}}\int_K\frac{h^2}{\rho}[\nabla p]\cdot[\nabla q]\,\mathrm{d}s, \tag*{(21)}

where F\mathcal{F} denotes the set of all interior faces in Ωh\Omega_h and [⋅][\cdot] denotes the jump of any quantity across given face. The interior penalty stabilization is used because of the simplicity of the implementation compared to, for example, GLS or SUPG stabilization. The effect of stabilization in optimal control was previously studied in [1] for GLS stabilization and in [12] for SUPG stabilization. The adjoint equations for the weak formulation (20) with stabilization (21) are included in the appendix.

Setup of the numerical experiments

The numerical experiments are designed to test whether the method can robustly reconstruct the control variables vin\mathbf{v}_{\mathrm{in}} and θ\theta from given subsampled, noisy data in various geometries.

Geometries

Three artificial 3D geometries, shown in Figure 4, are created using GMSH [16] and VMTK [4] libraries to test the performance of the method discussed in section 2. The first geometry is a straight narrowing tube with a circular cross-section. The width and length of the geometry were selected to approximately agree with the dimensions of a real descending aorta. The second geometry is a slightly bent, extended version of the first geometry. The third geometry consists of a tube with a constant radius with a 180-degree bend mimicking the shape of the aortic arch.

Figure 4. Schemes of the three geometries created to test the method’s performance are presented in section 2. The length of the straight tube is 0.12 m, and its inlet and outlet radii are 9 mm and 8 mm, respectively. The bent tube measures 0.14 mm in length, with inlet and outlet radii set to 10 mm and 8 mm, respectively. Both tube geometries include a 10 mm long region at each end with a constant radius. The arch geometry has a constant radius of 10 mm and straight 20 mm long sections at each end of the 180-degree bend.

Reference velocity fields

The ground truth data are obtained by solving for velocity and pressure fields for given (prescribed) θ\theta and vin\mathbf{v}_{\mathrm{in}}. To ensure good resolution of the results, the edge length hh of the meshes is set uniformly to 0.75 mm.

Using the generated meshes, the reference fields are computed for each of the following values of θ∈[0.2,0.5,0.8,1.0]\theta\in[0.2, 0.5, 0.8, 1.0]. The inlet velocity vin\mathbf{v}_{\mathrm{in}} is prescribed to be the analytical profile defined by (9) for each corresponding θ\theta, and the magnitude of the inlet velocity VV is set to 0.1 m · s−1^{-1}, which approximately corresponds to the diastolic velocity of blood in descending aorta. The MINI element without stabilization (21) is used to discretize the weak formulation (20) to compute the ground truth solution. The density ρ\rho and viscosity μ\mu are set to 1050 kg · m−3^{-3} and 3.8955⋅10−33.8955 \cdot10^{-3} kg · m−1^{-1} · s−1^{-1} respectively to agree with values used in literature for human blood, these parameters are summarized in Table 1. The parameter γ∗\gamma_{*} is set to 0.25 m2^{2} · s · kg−1^{-1} to maximize the boundary dissipation

θγ∗(1−θ)∫Γwall∣v∣2 ds(22)\frac{\theta}{\gamma_{*}(1-\theta)}\int_{\Gamma_{\mathrm{wall}}} \lvert\mathbf{v} \rvert^{2}\,\mathrm{d}s \tag*{(22)}
SymbolNameValueUnit
μ\mudynamic viscosity3.8955×10−33.8955 \times10^{-3}kg⋅m−1s−1\mathrm{kg} \cdot\mathrm{m}^{-1}\mathrm{s}^{-1}
ρ\rhodensity1,050kg⋅m−1\mathrm{kg} \cdot\mathrm{m}^{-1}
VVmean inlet velocity0.1m⋅s−1\mathrm{m} \cdot\mathrm{s}^{-1}
γ∗\gamma^{*}-0.25m2⋅s⋅kg−1\mathrm{m}^{2} \cdot\mathrm{s} \cdot\mathrm{kg}^{-1}
β∗\beta^{*}Nitsche penalty parameter1,000-
θ\thetaslip parameter0.0 – 1.0-

caption: Parameters used to generate reference velocity and pressure fields.

:::

for θ=0.5\theta= 0.5 in the case of Poiseuille flow. The effect of γ∗\gamma^{*} has been discussed in more detail in our previous work [10].

Preparing data for assimilation

A shorter segment of each geometry was created, as shown in Figure 5, to eliminate the effect of inlet and outlet from the reference velocity fields. These segments are meshed with an edge length set uniformly to 0.8 mm, 1 mm and 1.5 mm to acquire meshes with different densities, in order to test the sensitivity of the assimilation process. Using coarser meshes also helps to mimic reality, where the computational domain and real domain are not identical. Moreover, it also corresponds better to the real-patient cases where the typical voxel size of the 4D PC-MRI data is usually between 1.5 and 2.5 mm. The numbers of degrees of freedom of the generated meshes are provided in Table 2, Table 3 and Table 4.

GeometryVerticesFacesCellsMINI dofsP1/P1P_1/P_1 dofs
Straight tube14,790169,52683,227308,84159,160
Bent tube24,482287,316141,664522,92097,928
Arch26,513315,267155,762573,338106,052

Table 2. Table of degrees of freedom of the shorter meshes with h=1.5h = 1.5 mm.

GeometryVerticesFacesCellsMINI dofsP1/P1P_1/P_1 dofs
Bent tube80,716974,956483,0181,771,918322,864
Arch89,5541,088,000539,5611,976,899358,216

Table 3. Table of degrees of freedom of the shorter meshes with h=1h = 1 mm.

GeometryVerticesFacesCellsMINI dofsP1/P1P_1/P_1 dofs
Bent tube155,7721,901,386943,6863,454,146623,088
Arch164,7452,028,4191,008,1393,683,397658,980

Table 4. Table of degrees of freedom of the shorter meshes with h=0.8h = 0.8 mm.

Image not extracted

Figure 5. Shorter segments of the geometries used for the assimilation with edge length set to 1.5 mm.

The corresponding reference velocity fields for each θ\theta are then interpolated to the coarser meshes, and an additional Gaussian noise is added based on the given signal-to-noise ratio (SNR). The noise is scaled by vmax⁡/SNRv_{\max}/\mathrm{SNR} and its standard deviation is set to

σ=0.45vmax⁡SNR,(23)\sigma= 0.45 \frac{v_{\max}}{\mathrm{SNR}}, \tag*{(23)}

where vmax⁡v_{\max} is a maximal characteristic velocity (inspired by velocity encoding parameter for MRI), which represents an upper bound for the velocity components of dMRI for each of the geometries. The relation (23) is used for the same purpose in [23].

Solver setup

The finite element implementation of the problem is done using the FEniCS framework [2]. The forward discrete problem is solved by the Newton method with line search as implemented in the PETSc library [5, 14]. The linear subproblems are solved by sparse direct LU factorization applied to the full system using the MUMPS library [3].

The optimization step is done using SciPy [34] implementation of the L-BFGS-B algorithm and the dolfin-adjoint library [27], which generates the adjoint equation automatically from the weak formulation of the forward problem (20)–(21). In each optimization iteration, the solution from the previous iteration is used as the starting point for the nonlinear solver to reduce computational time. The stopping criteria ftol and gtol for the L-BFGS-B algorithm are set to 10−710^{-7} and 10−510^{-5}, respectively, where ftol measures the relative decrease of the functional and gtol controls the norm of the gradient of the functional.

Results of numerical experiments

The first test is performed on the straight tube mesh with edge length h=1.5h = 1.5 mm. Data with SNR=2\mathrm{SNR} = 2 were used for the assimilation with the regularization weights α\alpha and β\beta set to 0.001 and 0.1, respectively. A few different initial guesses were tested, and their choice did not significantly influence the optimal θ\theta and vin\mathbf{v}_{\mathrm{in}}. However, a good initial guess helped reduce the number of optimization iterations and, in some cases, reduce the risk of divergence of the nonlinear solver. The comparison of the data and the assimilation results for the MINI element and stabilized P1/P1P_1/P_1 element with stabilization weights αv=0.01\alpha_v = 0.01 and αp=0.01\alpha_p = 0.01 are shown in Figure 6 and the differences between the assimilation results and ground truth velocity are shown in Figure 7. We observe that the MINI element performed better in terms of recovering the value of θ\theta and the ground truth velocity field compared to the stabilized P1/P1P_1/P_1 element. This is expected since the stabilized P1/P1P_1/P_1 element is less accurate and has significantly fewer degrees of freedom than the MINI element on the same mesh, as shown in Table 2. It is also important to point out that the MINI element has been used to generate the ground truth data and, therefore, has an advantage in the reconstruction process. Figure 7 also shows that the error for stabilized P1/P1P_1/P_1 element has a similar character to the error for stabilized MINI element, suggesting that the stabilization might be the main cause of the inaccuracies.

Image not extracted

Figure 6. Comparison of noisy data on straight tube geometry (h=1.5h = 1.5 mm) with SNR = 2 (first row) with the assimilation velocity results using MINI element (second row) and stabilized P1/P1P_1/P_1 element (third row) for multiple values of θ\theta.

Image not extracted

Figure 7. Visualization of the difference between reference data (without noise) and computed velocity fields on straight tube geometry (h=1.5h = 1.5 mm) using MINI element (first row), stabilized MINI element (second row) and stabilized P1/P1P_1/P_1 element (third row) for multiple values of θ\theta.

In the following subsection, we present and comment on various aspects of the assimilation results for the bent and arch geometries.

Finite elements and stabilization

The method was tested for both the MINI element and the P1/P1P_1/P_1 element. The MINI element was used without and with stabilization using weights αv=0.01\alpha_v = 0.01 and αp=0\alpha_p = 0. Since the P1/P1P_1/P_1 element is not inf-sup stable by itself, we tested it only with stabilization using weights αv=αp=0.01\alpha_v = \alpha_p = 0.01. The results for bent tube geometry and arch geometry with edge length h=1.5h = 1.5 mm are present in Figure 8 and Figure 10, respectively. The errors with respect to the ground truth velocities are shown in Figure 9 and Figure 11. The results show that the stabilization influences θopt\theta_{\mathrm{opt}} especially close to no-slip because it introduces additional diffusion. However, it must be included with the P1/P1P_1/P_1 element to ensure its stability.

Image not extracted

Figure 8. Comparison of data without noise with assimilation velocity results in bent tube geometry with edge length h=1.5h = 1.5 mm using MINI element, stabilized MINI element (αv=0.01\alpha_v = 0.01, αp=0\alpha_p = 0) and stabilized P1/P1P_1/P_1 element (αv=αp=0.01\alpha_v = \alpha_p = 0.01) for multiple values of θ\theta.

Image not extracted

Figure 9. Visulatization of the difference between data and computed velocity fields on bent geometry (h=1.5h = 1.5 mm) using MINI element, stabilized MINI element and stabilized P1/P1P_1/P_1 element for multiple values of θ\theta.

Image not extracted

Figure 10. Comparison of data without noise with assimilation velocity results in arch geometry with edge length h=1.5h = 1.5 mm using MINI element, stabilized MINI element (αv=0.01,αp=0\alpha_v = 0.01,\alpha_p = 0) and stabilized P1/P1P_1/P_1 element (αv=αp=0.01\alpha_v = \alpha_p = 0.01) for multiple values of θ\theta.

Image not extracted

Figure 11. Visualization of the difference between data and computed velocity fields on arch geometry (h=1.5h = 1.5 mm) using MINI element, stabilized MINI element and stabilized P1/P1P_1/P_1 element for multiple values of θ\theta.

Mesh coarseness

Three meshes with different levels of refinement were used to assess the convergence of results computed using stabilized P1/P1P_1/P_1 element with αv=αp=0.01\alpha_v = \alpha_p = 0.01. The edge lengths of the meshes were selected as h=1.5h = 1.5 mm, h=1h = 1 mm and h=0.8h = 0.8 mm. Therefore, the finest mesh has approximately the same number of degrees of freedom for the P1/P1P_1/P_1 element as the coarsest mesh using the MINI element. The comparisons of the results are shown in Figure 12 and Figure 14, and the errors with respect to the ground truth velocities are shown in Figure 13 and Figure 15. We can see that the velocity field and the slip parameter θopt\theta_{\mathrm{opt}} get closer to the ground truth with the increasing number of degrees of freedom. Of course, it is important to point out that the problems involving finer meshes are much more computationally expensive and often more prone to divergence of the nonlinear solver.

Comparison of data and assimilation velocity results across mesh resolutions and values of $\theta$

Figure 12. Comparison of data without noise with assimilation velocity results in the bent tube geometry using stabilized P1/P1P_1/P_1 element with αp=αv=0.01\alpha_p = \alpha_v = 0.01 for the coarser mesh (h=1.5h = 1.5 mm, number of DOFs = 97,928), the finer mesh (h=1h = 1 mm, number of DOFs = 322,864), the finest mesh (h=0.8h = 0.8 mm, number of DOFs = 623,088) and multiple values of θ\theta.

Figure 13. Visualization of the difference between data and computed velocity fields on bent geometry using stabilized P1/P1P_1/P_1 element with αp=αv=0.01\alpha_p = \alpha_v = 0.01 for the coarser mesh (h=1.5h = 1.5 mm, number of DOFs = 97,928), the finer mesh (h=1h = 1 mm, number of DOFs = 322,864), the finest mesh (h=0.8h = 0.8 mm, number of DOFs = 623,088) and multiple values of θ\theta.

Comparison of data without noise with assimilation velocity results for multiple values of $\theta$ on arch geometry

Figure 14. Comparison of data without noise with assimilation velocity results in the arch geometry using stabilized P1/P1P_1/P_1 element with αp=αv=0.01\alpha_p = \alpha_v = 0.01 for the coarser mesh (h=1.5h = 1.5 mm, number of DOFs = 106, 052), the finer mesh (h=1h = 1 mm, number of DOFs = 358, 216), the finest mesh (h=0.8h = 0.8 mm, number of DOFs = 658, 980) and multiple values of θ\theta.

Figure 15. Visualization of the difference between data and computed velocity fields on arch geometry using stabilized P1/P1P_1/P_1 element with αp=αv=0.01\alpha_p = \alpha_v = 0.01 for the coarser mesh (h=1.5h = 1.5 mm, number of DOFs = 106, 052), the finer mesh (h=1h = 1 mm, number of DOFs = 358, 216), the finest mesh (h=0.8h = 0.8 mm, number of DOFs = 658, 980) and multiple values of θ\theta.

Amount of noise

The robustness of the method with respect to the signal-to-noise ratio (SNR) was tested for various values of the slip parameter θ\theta. In Figure 16 and Figure 17, we present the comparison of the results for θ=0.5\theta= 0.5 and θ=0.8\theta= 0.8. The value of SNR ranges from ∞\infty to 0.5, corresponding to no noise and noise with a

Comparison of data with various amounts of noise and assimilation velocity results on the bent tube geometry

Figure 16. Comparison of data with various amounts of noise with assimilation velocity results on the bent tube geometry with edge length h=1.5h = 1.5 mm using stabilized P1/P1P_1/P_1 element with αv=αp=0.01\alpha_v = \alpha_p = 0.01 for θ=0.5\theta= 0.5 and θ=0.8\theta= 0.8.

Comparison of data and assimilation velocity results for SNR = $\infty$, SNR = 2, SNR = 1, and SNR = 0.5, with $\theta = 0.5$ and $\theta = 0.8$

Figure 17. Comparison of data with various amounts of noise with assimilation velocity results on the arch geometry with edge length h=1.5h = 1.5 mm using stabilized P1/P1P_1/P_1 element with αv=αp=0.01\alpha_v = \alpha_p = 0.01 for θ=0.5\theta= 0.5 and θ=0.8\theta= 0.8.

maximum twice the size of maximal velocity, respectively. We observed that the amount of Gaussian noise does not have a significant effect on the reconstructed velocity field and value of the slip parameter θopt\theta_{\mathrm{opt}}. This was true also for all the elements and values of θ\theta we have tested.

Effect of regularization

The regularization weights in the error functional (10) can significantly impact the performance and results of the assimilation. In Figure 18, we plotted the L-curve for each regularization weight we used to determine which regularization weights have the best trade-off between the accuracy and the problem stability. The curves show the relation between J(v)\mathcal{J}(\mathbf{v}) and R(m)\mathcal{R}(\mathbf{m}) for varying values of regularization weights in the log-log scale. Ideally, the plotted quantities should be balanced to avoid over- and under-regularization. This should be possible to achieve by picking the highest regularization weights for which J(v)\mathcal{J}(\mathbf{v}) is still reasonably small. This approach for selecting the regulation weights was inspired by a similar problem in discrete inverse problems, see for example [20]. Figure 19 shows how the regularization affects the velocity field. If the regularization weights α\alpha and β\beta are not high enough, the inlet velocity profile is not smooth. On the other hand, if the regularization weights are too high, the inlet profile does not agree with the data at all, and θopt\theta_{\mathrm{opt}} is reconstructed poorly.

Graphs showing the relations between Tikhonov regularization and the distance of the velocity from noisy data

Figure 18. The graphs show the relations between the Tikhonov regularization and the distance of the velocity from the noisy data for varying regularization weights in the log-log scale. In both cases, the other regularization weight was kept constant at the value we selected as optimal, i.e. α=0.001\alpha= 0.001, β=0.1\beta= 0.1. This type of graph is sometimes called the L-curve.

Comparison of noisy data and assimilation velocity results for multiple values of $\theta$

Figure 19. Comparison of noisy data (SNR = 2.0) with assimilation velocity results computed using stabilized P1_1/P1_1 element with αv=αp=0.01\alpha_v = \alpha_p = 0.01 on the arch tube geometry with edge length h=1.5h = 1.5 mm with various amounts of regularization for multiple values of θ\theta.

Higher velocities

Since the motivation of this work is to apply the method to magnetic resonance images, it was tested for multiple velocities typical for blood flow in the descending aorta. Using the parameters from Table 1, we can compute the Reynolds number as

Re=LρVμ,\mathrm{Re} = \frac{L\rho V}{\mu},

where LL is the characteristic length of the geometry and VV is the characteristic flow speed. Table 5 lists Reynolds numbers corresponding to different points in the cardiac pulse cycle. The assimilation results for the higher velocities are shown in Figure 20 and Figure 21. When increasing the velocity, the main obstacle is the convergence of the nonlinear solver, especially for more complicated geometries, where more intricate flow patterns might appear. In such a case, using more suitable finite element stabilization and more robust algorithms for nonlinear problems based on continuation or Picard’s iterations might be necessary. Apart from the convergence issues, we observed similar phenomena as described in the previous sections.

diastoleaveragepeak systole
VV0.1 ms0.1\,\frac{\mathrm{m}}{\mathrm{s}}0.25 ms0.25\,\frac{\mathrm{m}}{\mathrm{s}}0.8 ms0.8\,\frac{\mathrm{m}}{\mathrm{s}}
Re\mathrm{Re}2706742156

Table 5. Reynolds numbers for different inflow velocities

Comparison of data and assimilation velocity results for higher velocities at multiple values of $\theta$

Figure 20. Comparison of data without noise with assimilation velocity results for higher velocities using the stabilized P1/P1P_1/P_1 element with αv=αp=0.01\alpha_v = \alpha_p = 0.01 on the bent tube geometry for multiple values of θ\theta.

Comparison of data without noise with assimilation velocity results for multiple values of $\theta$

Figure 21. Comparison of data without noise with assimilation velocity results for V=0.25V = 0.25 using the stabilized P1/P1P_1/P_1 element with αv=αp=0.01\alpha_v = \alpha_p = 0.01 on the arch geometry for multiple values of θ\theta.

Pressure reconstruction

Since the assimilation uses only the velocity field data in the error functional J\mathcal{J}, this method can be used to reconstruct the pressure field as a byproduct. We compared the reconstructed pressure fields with the ground truth pressure obtained in section 3.2 and interpolated to the shorter mesh. The results are presented in Figure 22, where we can observe that the method was able to reconstruct the pressure field very close to the ground truth. This suggests that this method could be used as an alternative for pressure reconstruction from velocity measurements as in [32].

Figure 22. Comparison of the reference and reconstructed pressure field using stabilized P1/P1P_1/P_1 element with αv=αp=0.01\alpha_v = \alpha_p = 0.01 on the arch geometry with edge length h=1.5h = 1.5 mm for multiple values of θ\theta. The error of pressure reconstruction is measured using a quantity PerrP_{\mathrm{err}} defined as ∥popt−pref∥L2(Ω)∥pref∥L2(Ω)\frac{\lVert p_{\mathrm{opt}}-p_{\mathrm{ref}}\rVert_{L_2(\Omega)}}{\lVert p_{\mathrm{ref}}\rVert_{L_2(\Omega)}}.

Analytical tests

The numerical settings used in this work were also tested using the methodology proposed by Málek and Rajagopal in [26]. Using a numerical solution of (20) computed on a 0.10.1 m long cylinder with radius R=0.01R = 0.01 m, the volumetric flow rate QQ and negative pressure gradient cc were computed using the following relations:

Q=∫Γoutv⋅n ds,c=1∣Ω∣(∫Γoutp ds−∫Γinp ds).(24)Q = \int_{\Gamma_{\mathrm{out}}} \mathbf{v} \cdot\mathbf{n}\,\mathrm{d}s,\qquad c = \frac{1}{\lvert\Omega\rvert}\left(\int_{\Gamma_{\mathrm{out}}} p\,\mathrm{d}s - \int_{\Gamma_{\mathrm{in}}} p\,\mathrm{d}s\right). \tag*{(24)}

As derived in [26, Sect. III] for Poiseuille flow in a pipe, assuming Navier’s slip condition on the impermeable wall, the Navier-Stokes fluid exhibits slip along the wall if the following criterion is fulfilled:

Q>−cπ8μR4.(25)Q > -\frac{c\pi}{8\mu}R^4. \tag*{(25)}
numerical settingθ=0.0\theta= 0.0θ=0.2\theta= 0.2θ=0.5\theta= 0.5θ=0.8\theta= 0.8θ=1.0\theta= 1.0
MINI, h=1.5h = 1.5 mm−2.32×10−3-2.32 \times10^{-3}0.200.490.790.99
MINI, h=1h = 1 mm−1.26×10−3-1.26 \times10^{-3}0.200.490.790.99
stab. P1/P1P_1/P_1, h=1.5h = 1.5 mm−2.41×10−2-2.41 \times10^{-2}0.140.410.781.00
stab. P1/P1P_1/P_1, h=1h = 1 mm−1.00×10−2-1.00 \times10^{-2}0.170.460.780.99
stab. P1/P1P_1/P_1, h=0.8h = 0.8 mm−6.54×10−3-6.54 \times10^{-3}0.180.470.770.97

caption: The comparison of the prescribed slip parameter θ\theta and its value computed using (26) on a straight tube using different numerical settings.

:::

In addition, see [26, formula (3.7)], the appropriate slip parameter θ\theta can be then computed using the relations

θγ∗(1−θ)=κ=12−cRπR2+c8μR2.(26)\frac{\theta}{\gamma_* (1-\theta)} = \kappa= \frac{1}{2} - \frac{cR}{\pi R^2 + \frac{c}{8\mu}R^2}. \tag*{(26)}

The comparison of the prescribed slip parameter θ\theta and its value computed using (26) is presented in Table 6. The results show that the selected numerical solutions can capture the relations between the macroscopic quantities well. We also observed that the amount of the finite element stabilization can significantly influence the computed pressure field up to the point where 26 no longer holds.

Conclusion and future work

We developed and implemented a variational assimilation method to reconstruct the boundary condition parameters vinv_{\mathrm{in}} and θ\theta from 3D measured velocity data for Navier–Stokes fluid with Navier’s slip boundary condition on the wall. The method was tested on three artificial geometries that correspond to real-patient geometries that we intend to study. Several experiments were performed to study the impact of the choice of finite element, amount of regularization and stabilization, mesh density, amount of noise and velocity magnitude. The aim of this study was to assess the robustness of the method with respect to the numerical setting for future use on real patient magnetic resonance data.

We observed that using the interior penalty stabilization resulted in a worse fit of the slip parameter due to the additional diffusion introduced by the stabilization, which affected the cases where θ\theta is closer to 1 in particular. On the other hand, stabilized P1/P1P_1/P_1 element proved to be significantly cheaper and more stable computationally and, therefore, more suitable for larger and more complicated problems for which element like MINI or Taylor Hood might be too computationally expensive. We also observed that the fit of θ\theta improved with increased mesh density, enabling the stabilized P1/P1P_1/P_1 element to recover more details in the flow pattern. The experiments also showed that the method is not very sensitive to the amount of Gaussian noise in the data. The effects of Tikhonov regularization R(m)\mathcal{R}(\mathbf{m}) were discussed, and a way to choose the regularization weights was proposed. The method was also tested for higher velocities, which brought complications connected with the higher Reynolds number. However, we observed similar effects as in the lower velocity regime. We also discussed the pressure reconstruction capability of the method. The results showed that the pressure was reconstructed well for artificially generated data even with stabilized P1/P1P_{1}/P_{1} elements and a relatively coarse mesh if the amount of stabilization is sufficiently small.

One of the limitations of the method is that the nonlinear solver is not guaranteed to converge, especially in the early stage of the optimization process. This can be partially salvaged by providing a suitable initial guess for the control variables or by using continuation and Picard’s iteration instead. Especially for the higher flow velocities it could be improved further by incorporating a more suitable stabilization such as SUPG/PSPG.

Appendix

The adjoint equations have to be assembled in order to compute the gradient of JRJ_{R} with respect to the control variables θ\theta and vin\mathbf{v}_{\mathrm{in}}. This is done automatically using the dolfin–adjoint library [27]. However, we derived the adjoint equations by hand as well. The adjoint equations for (20) including the stabilization terms (21) are as follows:

Find (ξ,η)∈Vh×Ph(\xi,\eta) \in V_{h} \times P_{h} such that

∫Ωρ((∇v)ϕ+(∇ϕ)v)⋅ξ dx+∫ΩT(ϕ,q):∇ξ dx+∫Ωηdiv⁡ϕ dx+∫Γwallθγ∗(1−θ)ϕt⋅ξt ds−∫Γwall(T(ϕ,q)n)n⋅ξn ds+∫Γwallϕn⋅(T(ξ,η)n)n ds+β∗μh∫Γwallϕn⋅ξn ds−∫Γoutρ(v⋅n)−(ϕ⋅n)(ξ⋅n) ds−∫Γout(T(ϕ,q)n)t⋅ξt ds+∫Γoutϕt⋅(T(ξ,η)n)t ds+β∗μh∫Γoutϕt⋅ξt ds+αv∑K∈F∫Kρh2[∇ϕ]⋅[∇ξ] ds+αp∑K∈F∫Kh2ρ[∇q]⋅[∇η] ds=1ℓ3∫Ω(v−dMRI)⋅ϕ dx(27)\begin{aligned} &\int_{\Omega} \rho\left((\nabla\mathbf{v})\phi+ (\nabla\phi)\mathbf{v}\right)\cdot\xi\,\mathrm{d}x + \int_{\Omega} \mathbb{T}(\phi,q) : \nabla\xi\,\mathrm{d}x + \int_{\Omega} \eta\operatorname{div}\phi\,\mathrm{d}x \\ &+ \int_{\Gamma_{\mathrm{wall}}} \frac{\theta}{\gamma_{*}(1-\theta)}\phi_{t}\cdot\xi_{t}\,\mathrm{d}s - \int_{\Gamma_{\mathrm{wall}}} \left(\mathbb{T}(\phi,q)\mathbf{n}\right)_{n}\cdot\xi_{n}\,\mathrm{d}s \\ &+ \int_{\Gamma_{\mathrm{wall}}} \phi_{n}\cdot\left(\mathbb{T}(\xi,\eta)\mathbf{n}\right)_{n}\,\mathrm{d}s + \frac{\beta_{*}\mu}{h}\int_{\Gamma_{\mathrm{wall}}}\phi_{n}\cdot\xi_{n}\,\mathrm{d}s \\ &- \int_{\Gamma_{\mathrm{out}}} \rho(\mathbf{v}\cdot\mathbf{n})_{-}(\phi\cdot\mathbf{n})(\xi\cdot\mathbf{n})\,\mathrm{d}s \\ &- \int_{\Gamma_{\mathrm{out}}}\left(\mathbb{T}(\phi,q)\mathbf{n}\right)_{t}\cdot\xi_{t}\,\mathrm{d}s + \int_{\Gamma_{\mathrm{out}}}\phi_{t}\cdot\left(\mathbb{T}(\xi,\eta)\mathbf{n}\right)_{t}\,\mathrm{d}s \\ &+ \frac{\beta_{*}\mu}{h}\int_{\Gamma_{\mathrm{out}}}\phi_{t}\cdot\xi_{t}\,\mathrm{d}s + \alpha_{v}\sum_{K\in\mathcal{F}}\int_{K}\rho h^{2}[\nabla\phi]\cdot[\nabla\xi]\,\mathrm{d}s \\ &+ \alpha_{p}\sum_{K\in\mathcal{F}}\int_{K}\frac{h^{2}}{\rho}[\nabla q]\cdot[\nabla\eta]\,\mathrm{d}s = \frac{1}{\ell^{3}}\int_{\Omega}(\mathbf{v}-\mathbf{d}_{\mathrm{MRI}})\cdot\phi\,\mathrm{d}x \tag*{(27)} \end{aligned}

for all (ϕ,q)∈Vh×Ph(\phi,q) \in V_{h} \times P_{h},

where Vh⊂{v∈H1(Ωh),v=0 on Γin}V_{h} \subset\{\mathbf{v} \in H^{1}(\Omega_{h}), \mathbf{v}=0 \text{ on } \Gamma_{\mathrm{in}}\} and Ph⊂L2(Ωh)P_{h} \subset L^{2}(\Omega_{h}).

References

  1. [1]F. Abraham, M. Behr, and M. Heinkenschloss. “The effect of stabilization in finite element methods for the optimal boundary control of the Oseen equations”. In: Finite Elements in Analysis and Design 41.3 (2004), pp. 229–251. DOI: http://doi.org/10.1016/j.finel.2004.06.001.
  2. [2]M. Alnæs et al. “The FEniCS Project Version 1.5”. In: Archive of Numerical Software 3.100 (2015). DOI: http://doi.org/10.11588/ans.2015.100.20553.
  3. [3]P. Amestoy, I. S. Duff, J. Koster, and J.-Y. L’Excellent. “A Fully Asynchronous Multifontal Solver Using Distributed Dynamic Scheduling”. In: SIAM Journal on Matrix Analysis and Applications 23.1 (2001), pp. 15–41. DOI: http://doi.org/10.1137/S0895479899358194.
  4. [4]L. Antiga, M. Piccinelli, L. Botti, B. Ene-Iordache, A. Remuzzi, and D. A. Steinman. “An image-based modeling framework for patient-specific computational hemodynamics”. In: Medical & Biological Engineering & Computing 46.11 (2008), pp. 1097–1112. DOI: http://doi.org/10.1007/s11517-008-0420-1.
  5. [5]S. Balay, W. D. Gropp, L. C. McInnes, and B. F. Smith. “Efficient Management of Parallelism in Object-Oriented Numerical Software Libraries”. In: Modern Software Tools for Scientific Computing. Birkhäuser Boston, 1997, pp. 163–202. DOI: http://doi.org/10.1007/978-1-4612-1986-6_8.
  6. [6]L. Bertagna, M. D’Elia, M. Perego, and A. Veneziani. “Data Assimilation in Cardiovascular Fluid–Structure Interaction Problems: An Introduction”. In: Fluid-Structure Interaction and Biomedical Applications. Ed. by T. Bodnár, G. P. Galdi, and Š. Nečasová. Basel: Springer Basel, 2014, pp. 395–481. DOI: http://doi.org/10.1007/978-3-0348-0822-4_6.
  7. [7]E. Burman. “A Penalty-Free Nonsymmetric Nitsche-Type Method for the Weak Imposition of Boundary Conditions”. In: SIAM Journal on Numerical Analysis 50.4 (2012), pp. 1959–1981. DOI: http://doi.org/10.1137/10081784x.
  8. [8]E. Burman. “Interior penalty variational multiscale method for the incompressible Navier–Stokes equation: Monitoring artificial dissipation”. In: Computer Methods in Applied Mechanics and Engineering 196.41-44 (2007), pp. 4045–4058. DOI: http://doi.org/10.1016/j.cma.2007.03.025.
  9. [9]R. Byrd, L. Peihuang, and J. Nocedal. A limited-memory algorithm for bound-constrained optimization. Mar. 1996. DOI: http://doi.org/10.2172/204262.
  10. [10]R. Chabiniok, J. Hron, A. Jarolímová, J. Málek, K. R. Rajagopal, K. Rajagopal, H. Švihlová, and K. Tůma. “A benchmark problem to evaluate implementational issues for three-dimensional flows of incompressible fluids subject to slip boundary conditions”. In: Applications in Engineering Science (2021), p. 100038. DOI: http://doi.org/10.1016/j.apples.2021.100038.
  11. [11]R. Chabiniok, J. Hron, A. Jarolímová, J. Málek, K. R. Rajagopal, K. Rajagopal, H. Švihlová, and K. Tůma. “Three-dimensional flows of incompressible Navier–Stokes fluids in tubes containing a sinus, with varying slip conditions at the wall”. In: International Journal of Engineering Science 180 (2022), p. 103749. DOI: http://doi.org/10.1016/j.ijengsci.2022.103749.
  12. [12]S. S. Collis and M. Heinkenschloss. Analysis of the Streamline Upwind/Petrov Galerkin Method Applied to the Solution of Optimal Control Problems. CAAM TR02-01 108. Rice University, 2002.
  13. [13]M. D’Elia, M. Perego, and A. Veneziani. “A Variational Data Assimilation Procedure for the Incompressible Navier-Stokes Equations in Hemodynamics”. In: Journal of Scientific Computing 52 (2012), pp. 340–359. DOI: http://doi.org/10.1007/s10915-011-9547-6.
  14. [14]L. D. Dalcin, R. R. Paz, P. A. Kler, and A. Cosimo. “Parallel distributed computing using Python”. In: Advances in Water Resources 34.9 (2011). New Computational Methods and Software Tools, pp. 1124–1139. DOI: http://doi.org/10.1016/j.advwatres.2011.04.013.
  15. [15]S. W. Funke, M. Nordaas, Ø. Evju, M. S. Alnæs, and K. A. Mardal. “Variational data assimilation for transient blood flow simulations: Cerebral aneurysms as an illustrative example”. In: International Journal for Numerical Methods in Biomedical Engineering 35.1 (2019), e3152. DOI: http://doi.org/10.1002/cnm.3152.
  16. [16]C. Geuzaine and J.-F. Remacle. “Gmsh: A 3-D finite element mesh generator with built-in pre- and post-processing facilities”. In: International Journal for Numerical Methods in Engineering 79.11 (May 2009), pp. 1309–1331. DOI: http://doi.org/10.1002/nme.2579.
  17. [17]T. Guerra, C. Catarino, T. Mestre, S. Santos, J. Tiago, and A. Sequeira. “A data assimilation approach for non-Newtonian blood flow simulations in 3D geometries”. In: Applied Mathematics and Computation 321 (2018), pp. 176–194. DOI: http://doi.org/10.1016/j.amc.2017.10.029.
  18. [18]T. Guerra, J. Tiago, and A. Sequeira. “A sequential control in blood flow simulations”. In: International Journal of Non-Linear Mechanics 64 (2014), pp. 57–69. DOI: http://doi.org/10.1016/j.ijnonlinmec.2014.04.005.
  19. [19]M. Gunzburger. “Adjoint Equation-Based Methods for Control Problems in Incompressible, Viscous Flows”. In: Flow, Turbulence and Combustion 3-4 (2000), pp. 249–272. DOI: http://doi.org/10.1023/A:1011455900396.
  20. [20]P. C. Hansen. Discrete Inverse Problems: Insight and Algorithms. Society for Industrial and Applied Mathematics, 2010. DOI: http://doi.org/10.1137/1.9780898718836.
  21. [21]R. Herzog and K. Kunisch. “Algorithms for PDE-constrained optimization”. In: GAMM Mitteilungen 33.2 (2010), pp. 163–176. DOI: http://doi.org/10.1002/gamm.201010013.
  22. [22]J. Hron, C. Le Roux, J. Málek, and K. R. Rajagopal. “Flows of Incompressible Fluids subject to Navier’s slip on the boundary”. In: Computers & Mathematics with Applications 56.8 (Oct. 2008), pp. 2128–2143. DOI: http://doi.org/10.1016/j.camwa.2008.03.058.

Paper details

vlnc:W_anvpu5j5
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