Hi @tmigot, can the objective functional have finite support (point evaluations, rather than integrals over the entire domain)? And how is speed?
It sounds like this might be a solution for the problem I described here Computing sensitivity to Dirichlet data in Gridap.jl