Application of the fast automatic differentiation technique for solving inverse coefficient problems (Q2206404)

Inverse coefficient problems for the heat equation are important in materials science. They may arise while controlling such processes as welding or metal solidification, etc. One of such inverse problems is to find the temperature-dependent thermal conductivity coefficients \(K(T)\) in the non-stationary heat equation basing on temperatures inside the domain of interest and on heat fluxes at the domain boundaries. Both temperatures and fluxes are known (e.g., measured) at some certain moments of time. This problem may be reformulated as a variational and is supposed to be solved numerically. Basing on initial and boundary conditions the range of temperature variation is found. Then certain values \(T_{m}\) in this range are mapped with the unknown \(k_{m}=K(T_{m})\), and the cost functional must be minimized with respect to the set of \(k_{m}\). The authors obtain an analytical expression for the cost functional gradient. As it is very complicated, a fast automatic differentiation (FAD) technique is used. FAD significantly outperforms symbolic differentiation and results in expressions for the gradient of the cost functional that are exact for the chosen approximation of the problem being solved. Finally, for the unknowns, a system of algebraic equations is obtained that is solved by iteration methods. At the final part of the paper, the authors discuss their experience of inverse problems solution in one and two dimensions. They also point out some numerical differentiation schemes that allow to reduce the computational cost of the inverse 3D-problem by splitting it in three orthogonal directions in space.