I study fluids that have internal structure, mostly when the fluid lives on a curved surface: lipid membranes whose molecules separate into domains, thin shells of liquid crystal whose defects wander around and annihilate each other, or, back in flat space, ferrofluids that respond to magnets. For each model I ask three questions. Does it have a solution, and how smooth is it? Can we build a finite element method that respects its physics, such as conservation of mass and dissipation of energy? Can we prove that the method converges, and how fast? The first question is PDE theory and the other two are numerical analysis. Often the answer to the first is what makes the third possible, so I work on both.

Membranes and films are thin enough to be modeled as surfaces, but then the geometry enters the equations in unexpected ways. To compute on a surface, you can mesh the surface itself (parametric FEM) or use a fixed mesh of the surrounding space and take traces of its functions on the surface (TraceFEM). I work with both.

Click on a topic below to read more.

Surface fluids without the saddle point

With Ricardo H. Nochetto (University of Maryland).

In the Stokes equations the pressure enforces incompressibility, which makes the problem a saddle point problem. Discretizations must then satisfy the discrete inf-sup condition, which rules out the most natural choice of equal-order elements. On a closed surface there is a way out. Taking the divergence of the momentum equation gives a Poisson equation for the pressure, and since the surface has no boundary, it needs no boundary condition. The geometry enters only through the Ricci curvature \(\mathbf{R}\) (the Gauss curvature in 2D):

\[-2\mu\,\Delta_S \mathbf{u} + \nabla_\Gamma p = \mathbf{f}, \qquad -\Delta_\Gamma p + 2\mu\,\mathrm{div}_\Gamma(\mathbf{R}\mathbf{u}) = -\mathrm{div}_\Gamma \mathbf{f} - 2\mu\,\Delta_\Gamma g.\]

This system is elliptic, so velocity and pressure can be discretized with any finite element spaces, with no stabilization parameter to tune. We prove stability and optimal error estimates for this approach. In a follow-up paper we show that the reformulation is just one step of Gaussian elimination applied to the original system. As a result, it is well-posed if and only if the original problem is, for the stationary and time-dependent Stokes and Navier–Stokes equations.

Tangential Navier–Stokes equations on a torus solved using a (non inf-sup stable) $P_k/P_k$ Lagrange velocity-pressure pair.

Papers:

How stable can TraceFEM be?

With Lucas Bouck (Flatiron Institute), Ricardo H. Nochetto (University of Maryland) and Vladimir Yushutin (University of Tennessee, Knoxville).

For time-dependent problems such as the heat equation, the strongest form of stability is inf-sup stability: the discrete problem is as well-posed as the continuous one, uniformly in the mesh size. We give necessary and sufficient conditions for inf-sup stability of TraceFEM, which hold no matter how the surface cuts through the mesh. Many results that were open for TraceFEM follow from this: best approximation, convergence for rough data, optimal error estimates and optimal conditioning. The project was motivated by the liquid crystal model below.

Paper:

Analysis on surfaces that aren't so smooth

With Gonzalo A. Benavides (University of Maryland) and Ricardo H. Nochetto (University of Maryland).

If the data are in \(L^p\), are the second derivatives of the solution in \(L^p\) too? Classical answers on surfaces assume the surface is infinitely smooth, but real surfaces are not. We prove \(W^{k,p}\) regularity for the Laplace–Beltrami, vector Laplace, Stokes and Navier–Stokes problems on surfaces that are only a few times differentiable, with sharp smoothness assumptions and elementary proofs. These results are used throughout my other work.

Papers:

Liquid crystal films

With Gonzalo A. Benavides (University of Maryland) and Ricardo H. Nochetto (University of Maryland).

Nematic liquid crystals are made of rod-like molecules that flow like a liquid but line up like a crystal. In thin curved shells, the curvature affects how their defects move and annihilate. We prove that the surface Beris–Edwards model, which couples surface Navier–Stokes flow to a tensor field describing the orientation of the molecules, has weak solutions. To our knowledge, this is the first solvability result for a hydrodynamic liquid crystal model on surfaces.

Paper:

Phase separation on membranes

With Maxim Olshanskii (University of Houston).

Membranes made of several kinds of lipids separate into domains, which are then carried by the flow in the membrane. We study the surface Stokes–Cahn–Hilliard model of this process. One twist has no counterpart in flat domains: some surfaces can rotate within themselves (think of a spinning sphere), and the fluid equations cannot see these Killing fields, even though they change how the phases move. We prove that the model is well-posed and design a TraceFEM that handles Killing fields, conserves mass and dissipates energy exactly. In a second paper we prove that this TraceFEM converges, with error estimates that required new \(L^p\)-based tools for TraceFEM.

Simulation of phase separation on a liposome produced with our method. Arrows describe the fluid flow and colormap indicates the phase.

Papers:

A detour into flat land: ferrofluids

With Gonzalo A. Benavides (University of Maryland) and Ricardo H. Nochetto (University of Maryland).

Ferrofluids are liquids with tiny magnetic particles suspended in them. In the standard model, the magnetization is carried by the flow with no diffusion, which makes the magnetic force very hard to control. Previous convergence proofs added an artificial diffusion term, and their guarantees fall apart when that term is removed. We derive the model from first principles and design Galerkin schemes for it without any added diffusion. To the best of our knowledge, these schemes come with the first error estimates in this setting.

Paper:

Codes

  • Stabilized TraceFEM codes for various PDEs on surfaces, implemented in the ngsxfem add-on to NGSolve; see GitHub.
  • Companion code for Surface Stokes Without Inf-Sup Condition, implemented in NGSolve; see GitHub.

Research profiles