Working up to NS

Current strategy (26-12-2023)

  • Prototype stable numerical traditional FE method using firedrake
  • Primary simulation interest is post-gather high-viscosity flow where flow is essentially irrotational. We will deal with domain overlap problems by remeshing.
  • This approach provides higher probability of experimentally determining a stable preconditioning method for dealing with complex constitutive equation of glass at immobile/quasi-solid temperatures.
  • The difficulty of the complex constitutive equation will probably have an approximate splitting of the hard elliptical problem from the lagrangian velocity term. The method pioneered in the work that started all this may be of use after all.

Currently I plan to use a Continuous Galerkin formulation for the full problem. As the numerical difficulty of this problem results from the elliptic terms in the problem rather than the hyperbolic ones (i.e. shocks will not predominate), I feel the additional complexity of DG methods aren't warranted.

Working through Finite Elements and Fast Iterative Solvers

In order to build my understanding of where the difficulties arise in solving the incompressible NS equations, I will be working through some of (Finite Elements and Fast Iterative Solvers With Applications in Incompressible Fluid Dynamics)[https://academic.oup.com/book/27915].

Poisson

The Poisson equation 2u=f -\nabla^2 u = f is the prototypyical elliptic PDE. Typical (Robin) boundary conditions can be expressed as αu+βun=g on Ω. \alpha u + \beta \nabla u \cdot \mathbf{n} = g \textrm{ on } \partial \Omega. Robin boundary conditions encompass Dirichlet (β=0\beta = 0) and Neumann (α=0\alpha = 0) boundary conditions.

In the Neumann case un=g \nabla u \cdot \mathbf{n} = g, the divergence theorem and the product rule for divergences gives us that ΩuudA=Ωudx=Ωfdx -\int_{\partial \Omega} \nabla u \cdot \mathbf{u} \intd{A} = - \int_\Omega \nabla \cdot \nabla u \intd{\mathbf{x}} = \int_\Omega f \intd{\mathbf{x}} but integrating the (strong version of) the Neumann boundary conditions, we find ΩundA=dΩg \int_{\partial \Omega} \nabla u \cdot \mathbf{n} \intd{A} = \intd_{\partial \Omega} g and we get a constraint that must be satisfied for the problem to be well-posed, namely Ωg+Ωf=0 \int_{\partial \Omega} g + \int_\Omega f = 0

Following the derivation we did in the other article here, under Neumann conditions we derive the weak form of the Poisson equation Ωuvdx=Ωvf+Ωv(un)dA=Ωvf+ΩvgdA. \int_\Omega \nabla u \cdot \nabla v \intd{\mathbf{x}} = \int_\Omega vf + \int_{\partial \Omega} v (\nabla u \cdot \mathbf{n}) \intd{A} = \int_\Omega vf + \int_{\partial \Omega} vg \intd{A}. The quantities necessary to formulate the above weak problem naturally live in the space H1(Ω)\mathcal{H}^1(\Omega), which is the space of functions uu which are weakly first differentiable .

A useful generic boundary problem that we will be implementing is 2u=f on Ωu=gD on ΩDun=gN on ΩNΩ=ΩDΩN \begin{align*} -\nabla^2 u &= f \textrm{ on } \Omega \\ u &= g_D \textrm{ on } \partial \Omega_D \\ \nabla u \cdot \mathbf{n} &= g_N \textrm{ on } \partial \Omega_N \\ \partial \Omega &= \partial \Omega_D \sqcup \Omega_N \end{align*}

Note that solutions exist in the affine subspace of H1\mathcal{H}^1 that is HS1={H1(Ω)u=gd on Ωd}\mathcal{H}^1_S = \{\mathcal{H}^1(\Omega) \mid u=g_d \textrm{ on } \partial \Omega_d \} and test functions are in HT1={H1(Ω)u=0 on Ωd}\mathcal{H}^1_T = \{\mathcal{H}^1(\Omega) \mid u=0 \textrm{ on } \partial \Omega_d \}. The test functions, while an affine subspace, are not a linear subspace. So we get the following weak formulation Ωuvdx=Ωvfdx+ΩNvgNdA for all vHT1 \int_\Omega \nabla u \cdot \nabla v \intd{\mathbf{x}} = \int_\Omega vf \intd{\mathbf{x}} + \int_{\partial \Omega_N} vg_N \intd{A} \textrm{ for all } v \in \mathcal{H}_T^1 where we see that the restriction of vv negates the contribution of the boundary integral on ΩD\partial \Omega_D.

The (Bubnov-) Galerkin Method with boundary

Construct a good subspace STHT1 \mathcal{S}_T \subset \mathcal{H}_T^1, with basis functions {ϕi}1in \{\phi_i\}_{1 \leq i \leq n}. Note that these basis functions satisfy the Dirichlet boundary condition. Then choose additional basis functions {ϕi}n+1in+n\{\phi_i\}_{n+1 \leq i \leq n+n_\partial} such that for some uSTu' \in \mathcal{S}_T we can augment this function intou=u+i=n+1n+nuiϕiu' = u + \sum_{i=n+1}^{n+n_\partial} u_i \phi_i to satisfy the Dirichlet boundary condition. We denote the set of uu' constructed thusly as SS\mathcal{S}_S . My current understanding is that since the basis functions of ST\mathcal{S}_T are zero on the boundary, the appropriate choice of nn_\partial will uniquely determine the combination uj u_j which I believe can be calculated offline. Therefore if the Dirichlet condition is fixed, these pseudo-basis functions do not increase the number of DOFs in the overall solution space. That is, for u=j=1nujϕj+j=n+1n+nujϕj u = \sum_{j=1}^n u_j \phi_j + \sum_{j=n+1}^{n+n_\partial} u_j \phi_j, the bilinear form ϕj,u\langle \phi_j, u \rangle defines a linear system which constrains only {ui}1in\{u_i\}_{1\leq i \leq n} . This justifies the definition of ST \mathcal{S}_T as the space of test functions. We can then write the finite-dimensional weak formulation which is to find uSSu \in \mathcal{S}_S such that Ωuvdx=Ωvfdx+ΩNvgNdA for all vST. \int_\Omega \nabla u \cdot \nabla v \intd{\mathbf{x}} = \int_{\Omega} vf \intd{\mathbf{x}} + \int_{\partial \Omega_N} v g_N \intd{A} \textrm{ for all } v \in \mathcal{S}_T. If we construct the basis functions such that the above equation is satisfied, then for a given test function ϕi\phi_i, we get `$$$

$$$`

Parent post: