Identification of the thermal conductivity coefficient using a given surface heat flux (Q2420929)

The authors propose a method for the resolution of an inverse problem for the 2D heat problem $C(T)\frac{\partial T}{\partial t}=\operatorname{div}(K(T)\nabla T)$ posed in $Q=S\times (0,\Theta )$ with $S=(0,L)\times (0,R)$, with the boundary conditions $T=w_{\Gamma }$ on $\Gamma \times (0,\Theta )$, the temperature starting from a given value $w_{0}$ at $t=0$. The authors assume that the solution belongs to $C^{2}(Q)\cap C^{1}(\overline{Q})$. To determine the thermal conductivity coefficient $K$, they introduce the cost functional $\Phi (KT))=\int_{0}^{\Theta}\int\int_{S}(T(s,t),-Y(s,t))^{2}\mu (s,t)\,dx\,dy\,dt+\int_{0}^{\Theta }\int_{\Gamma }\beta (s(l))(-K(T(s(l),t))\frac{\partial T}{\partial \overline{n}}(s,t))-P(s(l),t))^{2}\,dl\,dt$, where $\mu \geq 0$ and $\beta \geq 0$ are given weight functions, $Y$ is a given temperature field in $Q$, $P$ is a given heat flux on the boundary $\Gamma $ and $\frac{\partial T}{\partial \overline{n}}$ is the normal derivative of the temperature on this boundary. The authors write the problem above in a conservative form: $\frac{\partial E}{\partial t}(T)-\operatorname{div}(K(T)\nabla T)=0$ with $E(T)=\int_{0}^{T}C(\zeta )\,d\zeta $, they introduce the admissible class of control functions $G=\{K(z):K\in C^{1}([a,b]),$ $K>0$ in $[a,b]\}$ and the Lagrange functional $I=\Phi (KT))+\int_{0}^{\Theta }\int\int_{S}(p(s,t)(\frac{\partial E(T(s,t))}{\partial t}-\operatorname{div}(K(T(s,t))\nabla T(s,t)))\,dx\,dy\,dt$, where $p\in C^{2}Q)\cap C^{1}(\overline{Q})$ is an arbitrary function. They compute the first variation $\delta I$ of this functional $I$ with respect to a variation $\delta K$ of the control function, that they specialize when $p$ is determined by the conditions associated to the adjoint problem of the conservative problem. They introduce a partition $Q_{i}$ of disjoint subdomains of $Q$ such that each subdomain $Q_{i}$ does not contain two different isothermal surfaces corresponding to the same temperature and they introduce new independent variables associated to surfaces of constant temperature. They compute the variation $\delta I$ with respect to these new variables. The main part of the paper computes the gradient of the cost functional $\Phi $ through a fast automatic differentiation technique. The paper ends with the presentation of numerical results obtained in six examples.