Feel++

Laplacian nlopt

Here we propose to solve the following problem, finding the parameters b,c,d solution of the following homogeneous heat equation

\begin{cases} \begin{aligned} -\nabla\cdot(\kappa\nabla u) &=0 \quad \text{on}\;\Omega \\ u &= g \quad \text{on}\;\partial\Omega \end{aligned} \end{cases}

with \kappa\rightarrow\kappa(\kappa_0,\kappa_1,\kappa_2) and where \Omega denotes the whole domain, \partial\Omega denotes the boundary of the domain such that these parameters verify the observation u_\text{obs} .

In other words, the goal is to minimize an objective function J(\kappa_0,\kappa_1,\kappa_2)

\min_{\kappa_0,\kappa_1,\kappa_2\in\mathbb{R}} J = \frac{1}{2}\int_{U} (u-u_\text{obs})^2

with U\in\Omega an open subset of all observable measures.

In practice, observations are often boundary measurement taken on a subset of the domain boundary U\subseteq\partial\Omega . Thus arise concerns about existence and uniqueness of solution for the inverse problem, especially if observations does not cover the whole boundary.

For the following example, we take U=\Omega and we define

\kappa=1+{\kappa_0}_{\vert_{I_0}} + {\kappa_1}_{\vert_{I_1}} + {\kappa_2}_{\vert_{I_2}}

where I_0, I_1, I_2 are different subset (inclusions) of \Omega .

Adjoint method

We add a small perturbation to the given parameter b, c, d denoted

\begin{aligned} \kappa_0^\delta = \kappa_0 + \alpha\delta \kappa_0 \\ \kappa_1^\delta = \kappa_1 + \alpha\delta \kappa_1 \\ \kappa_2^\delta = \kappa_2 + \alpha\delta \kappa_2 \end{aligned}

for \alpha\in\mathbb{R} . Then u^\delta=u+\alpha\delta u is the perturbed solution of the equation

\begin{cases} \begin{aligned} \displaystyle -\nabla\cdot(\kappa^\delta \nabla u^\delta) &= 0 \quad \text{on}\;\Omega \\ u^\delta &= 0 \quad \text{on}\;\partial\Omega \end{aligned} \end{cases}

where the diffusion coefficient for the perturbation \kappa^\delta \rightarrow \kappa( \kappa_0^\delta, \kappa_1^\delta, \kappa_2^\delta )

Lets note \delta J the shift for the cost function J such that

\begin{aligned} \delta J &=J(\kappa_0^\delta, \kappa_1^\delta, \kappa_2^\delta) - J(\kappa_0,\kappa_1,\kappa_2) \\ &=\frac{1}{2}\int_U (u^\delta - u_\text{obs})^2 - (u-u_\text{obs})^2 \\ &=\frac{1}{2}\int_U (u^\delta - u)(u^\delta+u-2 u_\text{obs}) \\ &=\frac{1}{2}\int_U (u^\delta - u)( (u^\delta-u_\text{objs}) + (u - u_\text{obs}) ) \\ \end{aligned} If we divide by \alpha and look when \alpha\rightarrow 0 , then we have

\delta J \approx \int_U \delta u(u - u_\text{obs})

Now we desire a method to compute the gradient. We consider the following linear tangent model deduced from two previous system of equations. We have

-\nabla\cdot(\kappa^\delta \nabla u^\delta) = -\nabla\cdot(\kappa \nabla u)

We introduce a state variable p (the adjoint) chosen appropriately as we will see further. We multiply the previous model by this variable and integrate on the domain.

-\int_\Omega \nabla\cdot(\kappa^\delta \nabla u^\delta) p = -\int_\Omega \nabla\cdot(\kappa \nabla u) p

Then writing the weak form we have

\int_\Omega \kappa^\delta \nabla u^\delta \nabla p -\int_\Omega\kappa^\delta \frac{\partial u^\delta}{\partial\mathbf n}p = \int_\Omega \kappa\nabla u\nabla p -\int_\Omega\kappa\frac{\partial u}{\partial\mathbf n}p

we expand the perturbed diffusion coefficient as \kappa^\delta=\kappa +\alpha\delta\kappa and rearrange terms by u^\delta -u ( =\alpha\delta u ). We divide by \alpha as previously and makes \alpha\rightarrow 0 , we have

\int_\Omega \kappa \nabla \delta u \nabla p -\int_{\partial\Omega} \kappa \frac{\partial \delta u }{\partial\mathbf n}p \approx -\int_\Omega \delta\kappa\nabla u \nabla p +\int_{\partial\Omega} \delta\kappa \frac{\partial u}{\partial\mathbf n}p

We integrate by part a second time

-\int_\Omega \delta u \nabla\cdot(\kappa \nabla p) +\int_{\partial\Omega} \left( \kappa \frac{\partial p }{\partial\mathbf n}\delta u - \kappa \frac{\partial \delta u }{\partial\mathbf n}p \right) \approx \int_\Omega u \nabla\cdot(\delta\kappa \nabla p) -\int_{\partial\Omega} \left( \delta\kappa \frac{\partial p}{\partial\mathbf n}u +\delta\kappa \frac{\partial u}{\partial\mathbf n}p \right)

we obtain finally for p=0 on the boundary

-\int_\Omega \delta u \nabla\cdot(\kappa \nabla p) \approx \int_\Omega u\nabla\cdot(\delta\kappa \nabla p)

Now to compute the gradient, we recall the previous equation for J and we deduce from build from the previous equation the following model (adjoint equation) for the state variable p

\begin{cases} \begin{aligned} \displaystyle -\nabla\cdot(\kappa\nabla p) &= (u - u_\text{objs}) \\ u &= 0 \quad \text{on}\;\partial\Omega \end{aligned} \end{cases}

If we multiply this equation by u^\delta-u and integrate over the domain, we obtain

\begin{aligned} -\int_\Omega \nabla\cdot(\kappa\nabla p)(u^\delta-u) &= \int_\Omega (u - u_\text{objs})(u^\delta - u) \\ &\approx \int_\Omega (u - u_\text{objs})\delta u \\ &\approx \delta J \end{aligned}

Now if we write the weak form of this equation, we have

\delta J \approx\int_\Omega \kappa\nabla p\nabla \delta u -\int_{\partial\Omega} \kappa\frac{\partial p}{\partial\mathbf n} \delta u

We apply the boundary condition

\begin{aligned} \delta J &=\int_\Omega \kappa\nabla p\nabla\delta u \\ &=-\int_\Omega \delta\kappa\nabla p\nabla u \end{aligned}

We recall \kappa chosen as

\kappa=1+{k_0}_{\vert_{I_0}} + {k_1}_{\vert_{I_1}} + {k_2}_{\vert_{I_2}}

Then the we deduce the gradient from the derivatives

\nabla J(\kappa_0,\kappa_1,\kappa_2) = \left( -\int_{I_0} \nabla p\nabla u, -\int_{I_1} \nabla p\nabla u, -\int_{I_2} \nabla p\nabla u \right)

So now we have a method to compute the gradient.