Existence of weak martingale solutions to a stochastic fluid-structure interaction problem with a compressible viscous fluid

In this article, we consider a fluid-structure interaction (FSI) problem involving isentropic compressible fluid in a three-dimensional domain, where part of the fluid domain boundary consists of an elastic deformable structure, and where the system is perturbed by stochastic effects. The stochastic forcing is applied to both the momentum equations as a volumetric body force and the structure as an external load to the deformable fluid boundary. The noise coefficients depend on the structure displacement, the structure velocity, the density of the fluid, and its momentum. The isentropic compressible fluid flow is governed by the 3D Navier-Stokes equations whereas Koiter shell-type equations give the elastodynamics of the structure. The adiabatic constant describing the power law for the fluid pressure is assumed to be $>\frac{3}{2}$. The fluid interacts with the elastic structure at the fluid-structure interface via a two-way coupling, in which the fluid dynamics and elastodynamics of the structure mutually interact with each other. This two-way coupling ensures the continuity of velocities and of contact forces at the randomly moving fluid-structure interface. The main result of this manuscript is a proof of the existence of a local-in-time (until an almost surely positive stopping time) weak martingale solution to this nonlinearly coupled stochastic fluid-structure interaction problem. That is, we prove the existence of solutions that are weak in the analytical sense and in the probabilistic sense, thereby proving the robustness of the underlying deterministic benchmark fluid-structure interaction problem to external stochastic noise. To the best of our knowledge, this is the first result in the field of stochastic moving boundary problems involving compressible fluids, specifically when the random and time-dependent displacement of the fluid domain is not known a priori and is itself an unknown in the problem.

One of the fundamental difficulties in this problem is the random moving domain, which is a priori unknown, on which the fluid equations are posed. The fluid-structure interaction is described by two-way coupling conditions, namely, the dynamic coupling condition and the kinematic coupling condition, which results in a problem that is highly nonlinear and presents numerous mathematical challenges. In order to handle the dynamic coupling condition between the fluid and the structure, we use a splitting scheme approach in the spirit of [32], which separates the elastodynamics of the structure and the fluid after discretizing in time via an approximation parameter $\Delta t$. This involves running two subproblems, a fluid subproblem and a structure subproblem, over discretized time intervals of length $\Delta t$, such that the net total contribution of both subproblems on a single time interval, summed over all time intervals, is a discretized approximation of the limiting weak formulation that will converge as the parameter $\Delta t\to 0$. In the first (structure) subproblem, only the elastodynamics equations are updated while keeping the fluid entities the same as in the previous step. In the second (fluid) subproblem, which is posed on a maximal domain, the fluid density and velocity are updated, where the fluid subproblem is supplemented with a viscous regularization term for the density in the continuity equation via parameter $\varepsilon$ and an artificial pressure term in the momentum equations via parameter $\delta$. These two layers of approximations, inspired by the method of Feireisl [18], are intended to be compatible with the Galerkin approximation and to improve the integrability of the pressure.

We address here two additional issues that we come across in the construction of the second subproblem, caused by the stochasticity in the problem. Firstly, we note that the test functions for the fluid and the structure problems are required to satisfy the kinematic coupling condition, which is the essential no-slip boundary condition imposed at the random and time-varying fluid-structure interface. This condition, which is typically embedded in the test space, requires that we define this subproblem in terms of stochastic test functions. This is not amenable to the fact that we are constructing martingale solutions. We overcome this issue by decoupling the structure and the fluid equations along the kinematic coupling condition via a penalty method that essentially treats the structure as a semi-permeable material of thickness of some order of $\delta$. In particular, at the approximate level, the no-slip condition is not satisfied but it is recovered in the final limit passage of the approximate solutions. To be precise, we penalize the boundary behavior of the fluid and the structure velocities in a tubular neighborhood of size of some order of $\delta$ outside of the moving interface. This gives us a better control on the fluid density outside of the moving domain (caused by the seepage) that other Brinkman-type penalization terms fail to offer. We emphasize that the penalty term that we use is new and unique to our current problem. While this penalization term in the approximate weak formulations allows us to consider a decoupled pair of deterministic test functions, it causes further analytical issues, discussed below, as it allows for seepage of the fluid through the structure.

The second issue is associated with the fact that the fluid domain can degenerate in a random fashion i.e. when the top compliant boundary comes in contact with the bottom rigid boundary. We handle this no-contact condition by introducing appropriate stopped processes that we call “artificial structure displacements”. The construction of these processes provides a deterministic upper bound for the $H^{s}$-norm, for a fixed $s<2$, of the structure displacement, and ensures that self-interaction of the artificial structure does not occur at any time by maintaining a minimum distance of $\alpha>0$ between the lateral walls. The discrepancy between the artificial and the original structure displacement is eventually resolved by using a stopping time argument, as in [40], which provides the time-length of existence of the solution. That is, we prove that the two displacements coincide until an almost surely positive stopping time. These deterministic bounds further enable us to work with a maximal domain that contains all the artificial structure displacements. We are thus able to extend the viscosity coefficients with respect to the artificial structure displacements and pose the second (momentum) subproblem on a fixed maximal domain containing all associated artificial fluid domains. The method of extending the fluid problem to a maximal domain via extension of the viscosity coefficients has been previously employed in the study of deterministic compressible Navier-Stokes equations on time-dependent domains [15, 19, 25] and deterministic compressible fluid-structure interaction, for example in [32].

We then derive tightness results for the laws of the approximate solutions, in appropriate phase spaces, by employing compactness arguments. Our noise coefficient depends on the structure displacement, structure velocity, fluid momentum, and fluid density. Passage of the approximation parameters to their limits and the structure of our noise coefficient require almost sure convergence of the structure velocity strongly in $L^{2}_{t}L^{2}_{x}$. For that purpose, we derive a novel regularity result for the structure velocity which gives uniform bounds for the fractional time derivative, of some order strictly less than $\frac{1}{2}$. This result combined with the Aubin-Lions theorem gives us the tightness of the laws of the structure velocity in $L^{2}_{t}L^{2}_{x}$. This is the first temporal regularity result for the structure in this setting of FSI involving a compressible fluid. Next, due to the nature of compressible Navier-Stokes equations, we obtain tightness of the laws of the fluid entities in the weak topology of their respective phase spaces. Thus, upgrading to almost sure convergence on these non-Polish phase spaces requires a variant of the Skorohod Representation theorem which is obtained by a composition of the results in [35] and [24], that provides the existence of a sequence of new random variables (on a new probability space) with the same laws as the original variables that converge almost surely in the topology of the phase space, where the new probability space can be taken canonically to be the same probability space $[0,1]^{2}$ with the Borel sigma algebra and Lebesgue measure, and where the new Brownian motions on the same probability space are the same (i.e. parameter independent). This will be important in order to construct filtrations which are compatible with the process of taking the limit, in the final limit $\delta\to 0$, where we have to construct and work with random test processes so that the penalty term drops out. More precisely, the approximate random test functions will depend on both the $\delta$-approximate and the limiting structure displacements. Since the stochastic force appears in the momentum equations as a volumetric body force, we must ensure that these test processes are adapted to the filtration that we construct. Hence we require the filtrations for the approximate problems at each $\delta>0$ approximation level on the new probability space to contain the filtration generated by the limiting solution too. However, constructing a filtration by enlarging the natural filtrations of the approximate solutions by that of the limiting solutions may give us a filtration which is not non-anticipative with respect to the Brownian motions i.e. that their increments in time may not be independent with respect to this enlarged filtration. Having the same Brownian motion on the new probability space which is independent of $\delta$ fixes this issue and allows us to rigorously pass to the limit as $\delta\to 0$ using random test functions.

We note here that the usual use of the Bogovski operator to obtain higher integrability of the pressure fails due to the fact that our structure displacement is not Lipschitz continuous in space. Hence, by constructing an appropriate random test function, we provide better pressure estimates in the interior of the moving domain away from the fluid-structure interface. We then prove that the pressure does not concentrate at the moving interface by using the ideas given in [28] to deal with rough boundaries and then adapted to the moving domain (deterministic) case in [11].

Using these almost sure convergence results we are able to pass the approximation parameters to their respective limits in the following order: first we pass the time-discretization step $\Delta t\to 0$ in Section 5, then the Galerkin parameter to $N\to\infty$ in Section 6, then the viscous regularization parameter to $\varepsilon\to 0$ in Section 7, and finally, in Section 8, the pressure regularization, the penalty and the extension parameter to $\delta\to 0$.

We must also address that even though we have extended the approximate fluid subproblem to a maximal domain, the dynamics, in the limit, really are occurring in only the physical domain. In the deterministic case, this is accomplished by a vanishing of density result (see Lemma 3.1 in [32] and Lemma 4.1 in [19]), which shows that the vacuum outside of the initial domain is transported outside of the physical domain at all times. In contrast to the case of deterministic analysis, we do not have that our $\delta$-level approximate densities are zero on the part of the maximal domain that is outside of the physical domain in the stochastic regime, since we cannot handle the random kinematic coupling of test functions until the final limit passage. Thus, we must develop new estimates that show that the total mass of the fluid in the maximal domain that is outside of the physical domain, while not identically zero necessarily, converges to zero, as some order of $\delta$, in the limit as $\delta\to 0$, see Proposition 8.4. These estimates are important in particular, for considering the limit of the advection term, which at the approximate level is posed on the entire maximal domain.

In addition, the final limit passage as $\delta\to 0$ requires careful consideration of moving domains. Because of the way that we have extended the viscosity coefficients for example, we only have uniform estimates of the gradient of the fluid velocity on the moving domain and not on the entire fixed maximal domain. Throughout the limit passage as $\delta\to 0$, we require careful consideration of how indicator functions of the moving domain and the exterior tubular neighborhood on which the penalty term is defined, affect the analysis. The convergence of stochastic integrals in the limit as $\delta\to 0$ in Lemma 8.5 is in the spirit of convergence results for stochastic compressible fluid dynamics on fixed domains in Chapter 4 of [7], which involve rigorously identifying how having uniform bounds on approximate solutions allows for weak convergence of compositions of Carathéodory functions (satisfying appropriate growth conditions) with the approximate solutions, see Theorem 8.1 for the statement of this result in the context of our current stochastic moving boundary problem. This is combined with deterministic convergence lemmas for functions defined on moving domains from [10] in order to properly identify the weak limits of random nonlinear quantities on random moving domains, see Lemma 8.3. To handle the fact that we only have uniform bounds on the fluid velocity on moving domains rather than the entire maximal domain, we also prove a new extension result in Theorem 8.2, which states that the extension by zero of an $H^{s}(\mathcal{O})$ function, for a domain ${\mathcal{O}}$ which is a hypograph of an $\alpha$-Hölder continuous function, is in $H^{s\alpha}({\mathbb{R}}^{3})$ for all $0\leq s<\alpha/2$. Since domains with boundaries that are $\alpha$-Hölder continuous but not Lipschitz are found often in fluid-structure interaction problems, we believe that such an extension by zero regularity result is also of independent interest.

One of the final challenges in the last limit passage as $\delta\to 0$ is the fact that we must construct random test functions for the approximate and limiting weak formulations with appropriate measurability and convergence properties and boundary behaviour. Namely, the random test pair for the approximate weak formulation at this stage must be adapted, must satisfy the kinematic coupling condition so that the penalty term drops out, and it must converge to the test pair for the limiting weak formulation in appropriately regular spaces as $\delta\to 0$. One cannot find these approximate test functions by using usual tricks involving restricting or transforming, via the Arbitrary Lagrangian-Eulerian map, the limiting fluid test function on the approximate fluid domains as these methods inherit the regularity of the structure displacement which, in our case, is only Hölder continuous in space. Hence, we provide a detailed construction for such a test pair in Section 8.3, previously missing from existing literature, that satisfies the requirements mentioned above while possessing appropriate measurability properties (adaptedness) as necessitated by the stochastic integral defined on moving domains.

We will consider a fluid-structure interaction model for a compressible fluid in a three-dimensional domain, where part of the boundary is an elastic deformable boundary, and where the system is perturbed by stochastic effects, involving stochastic perturbations in the balance of momentum of the compressible fluid and in the elastodynamics of the structure.

We will consider the fluid flow in a periodic channel interacting with a complaint structure that sits atop the fluid domain. We will define the reference configuration of the elastic structure located on the top boundary of the box $\mathcal{O}$ by

The fluid reference domain, is given by

The bottom part of the boundary $\partial{\mathcal{O}}$ will be denoted by

We will assume that the elastic structure at the top boundary of $\mathcal{O}$ displaces in only the transverse $z$ direction so that $\eta$ will be the time-dependent (scalar) displacement of the elastic structure from its reference configuration $\Gamma$ in the $z$ direction. Given a particular function $\eta(t,x,y)$ describing the scalar transverse displacement of the elastic structure from $\Gamma$, we can define the time-dependent fluid domain at time $t\in[0,T]$ by (see Fig. 1):

and the top boundary of $\mathcal{O}_{\eta}(t)$ will be the physical time-dependent configuration of the elastic structure at time $t$, given by

We will assume that a compressible, isentropic, viscous fluid occupies the three-dimensional domain $\mathcal{O}_{\eta}(t)$ at time $t$ and that the dynamics of the fluid and the structure influence each other.

For the fluid subproblem, we model the fluid velocity ${\boldsymbol{u}}$ and the fluid density $\rho$ by the compressible Navier-Stokes equations for a viscous fluid:

where the Cauchy stress tensor $\boldsymbol{\sigma}(\rho,{\boldsymbol{u}})$ is defined by

where $\displaystyle\boldsymbol{D}({\boldsymbol{u}})=\frac{1}{2}(\nabla{\boldsymbol{u}}+\nabla^{t}{\boldsymbol{u}})$ is the symmetrized gradient of the velocity and $F_{f}$ is an external force, to be specified later, which will eventually represent stochastic perturbations of the fluid momentum. Moreover, we assume that

The first equation is the continuity equation, which represents the conservation of mass, and the second equation represents the balance of momentum. Here, $\mu,\lambda>0$ are the viscosity coefficients, and $p$ is the fluid pressure, which is given as a function of the fluid density $\rho$ through a constitutive relationship for the problem. In the current setting, we will assume that the viscous compressible fluid is isentropic so that the pressure law is given by a power law:

We prescribe periodic boundary condition for the fluid velocity in $x$ and $y$ direction and prescribe no-slip boundary condition on the bottom boundary i.e. ${\boldsymbol{u}}|_{\Gamma_{b}}=0$.

For the structure subproblem, we model the evolution of the structure displacement $\eta$ by the Koiter shell equation:

where we recall that $\eta$ is the transverse scalar displacement in the $z$ direction of the elastic structure from its reference configuration $\Gamma$. We emphasize that while the fluid equations are posed on the physical time-dependent domain, the elastodynamics of the structure are prescribed in the Eulerian coordinates on the reference domain. Here, the constant $\alpha$ represents a viscoelasticity parameter and we assume that $\alpha>0$ in the current manuscript to obtain an existence result. We supply this elastodynamics equation for the structure displacement $\eta$ with periodic boundary conditions.

• •

  The kinematic coupling condition describes the continuity of velocities via the no-slip condition along the moving (time-dependent) fluid-structure interface so that:

  (5)

  $${\boldsymbol{u}}=(\partial_{t}\eta)\boldsymbol{e}_{z}\qquad\text{ on }\Gamma_{\eta}(t),$$

  where $\boldsymbol{e}_{z}$ is the unit normal vector in the $z$ direction, and where we recall that the structure is assumed to displace in only the transverse $z$ direction.

• •

  The dynamic coupling condition describes the load of the fluid on the structure via the Cauchy stress tensor. This condition specifies the form of the external force on the structure as

  (6)

  $$F_{s}=-\boldsymbol{\sigma}(\rho,{\boldsymbol{u}})\boldsymbol{n}(t)\cdot\boldsymbol{e}_{z}+F_{s}^{stoch},$$

  where $\boldsymbol{n}(t)$ is the time-dependent normal vector to the moving fluid-structure interface $\Gamma(t)$, and where $F_{s}^{stoch}$ is an external stochastic forcing acting on the structure that will be specified later.

Motivated by [7], we consider nonlinear noise intensity operators $G$ and $\boldsymbol{F}$ such that $\boldsymbol{F}(\rho,\rho{\boldsymbol{u}}):\mathcal{U}_{0}\to{H}^{-l}({\mathbb{R}}^{3})$ for $l>\frac{3}{2}$ and $G{(\eta,v)}:\mathcal{U}_{0}\to L^{2}(\Gamma)$ are Hilbert-Schmidt.  We define,

For given functions $\rho:\mathcal{O}_{\alpha}\to\mathbb{R}$ and ${\boldsymbol{u}}:\mathcal{O}_{\alpha}\to\mathbb{R}^{3}$, we define

We assume the following noise structure: There exists a sequence of positive constants $\{c_{k}\}_{k=1}^{\infty}$ with $\displaystyle\sum_{k=1}^{\infty}c_{k}^{2}<\infty$ such that for all positive integers $k$, the functions $f_{k}:[0,\infty)\times{\mathbb{R}}^{3}\times{\mathbb{R}}^{3}\to{\mathbb{R}}$ and $g_{k}:{\mathbb{R}}\times{\mathbb{R}}\times\Gamma\to{\mathbb{R}}$, which are $C^{1}$ functions in all of the inputs, satisfy:

Note in particular that (7) implies that the intensity of the noise in the momentum equations is identically zero whenever the fluid density is equal to zero which is consistent with deterministic theory. Moreover, this allows us to naturally extend the noise $\boldsymbol{F}dW$ outside of the moving domain to the entire space ${\mathbb{R}}^{3}$ by assuming that the fluid density is equal to 0 outside of the moving domain.

We are now in position to state the main result of this paper.

Before presenting our splitting scheme in Section 4.3, we will present the necessary setup. We fix some parameter $\alpha>0$, which will be the parameter for which the approximate structure displacements we work with at all steps of the existence proof will satisfy $\|\eta\|_{H^{s}(\Gamma)}\leq\frac{1}{\alpha}$, for some $s\in(3/2,2)$. Since $s\in(3/2,2)$, by Sobolev embedding of $H^{s}(\Gamma)\subset C(\Gamma)$, this allows us to define a maximal domain $\mathcal{O}_{\alpha}$: