For the sake of the rest of the question, I'm interested in the porous medium equation

$$S\frac{\partial\phi}{\partial t} = \nabla\cdot K\phi\,\nabla\phi$$

where $S$ and $K$ are spatially-variable coefficients and $\phi$ (the hydraulic head) is the field to be solved for. This is a parabolic PDE, but it's degenerate since the "diffusivity" $K\phi$ of the medium can go to zero whenever $\phi \to 0$. As a consequence, the $\phi = 0$ contour propagates with finite speed in contrast to the usual infinite propagation speed of disturbances for non-degenerate parabolic PDE.

There are well-posedness results for the porous medium equation and convergent discretization schemes that don't regularize away the degeneracy. So in practice the weirdness of a diffusion coefficient going to zero can be overcome.

Now suppose we want to estimate the hydraulic conductivity $K$ from observational data. In order to do that, we'll probably use the adjoint method. The adjoint of the linearization of the PDE is

$$S\frac{\partial\psi}{\partial t} = -\nabla\cdot K\phi\,\nabla\psi + K\nabla\phi\cdot\nabla\psi + \ldots$$

where the ellipses denote some source terms that don't particularly matter. This is solved for the adjoint state $\psi$ backwards on some interval $[0, T]$ with a final condition imposed on $\psi|_{t = T}$. The adjoint linearization is an advection-diffusion equation, again with a diffusion coefficient ($K\phi$) that can and probably does go to zero somewhere.

Does the degeneracy of the adjoint equation pose a problem in practice? I've seen papers on overcoming the degeneracy of the forward problem, but I haven't seen anyone discuss the impact (if any) on the inverse problem. I'd add that degenerate problems show up in more fields than just hydrology -- you get similar equations for ice flow, surface water runoff, algae growth in ocean columns, and wildfire spread.

  • $\begingroup$ I haven't worked with porous media equations, but I have worked with ones that are spatially varying where the diffusivity can be zero (or even infinity). If the diffusivity is zero this just means $\partial_t \phi = 0$, i.e. $\phi$ remains fixed at that location. In the problems that I solve this is equivalent to prescribing a Dirichlet condition (since it doesn't evolve). There's nothing too worrying about it, except maybe that in 2D and higher, if this occurs at a single point you get a logarithmic singularity. $\endgroup$
    – lightxbulb
    Dec 4, 2023 at 18:26


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.