First-order conditions for the optimal control of learning-informed nonsmooth PDEs Academic Article