Input convex neural networks as surrogates in mathematical optimisation