Punch-bending is an industrial, pivotal process that allows to rapidly manufacture large quantities of products by bending semi-finished products, such as metal sheets or -wires, into target geometries. To ensure that the desired geometry can be reliably produced, a suitable process configuration, such as the used tool geometries and material properties, is required. Unlike traditional expensive trial-and-error approaches, we propose a machine learning surrogate model for the punch-bending process and subsequent plausible counterfactual explanations, which suggest process configurations for a desired target geometry. For this to work, the surrogate model must exhibit realistic generalization behavior. We analyze the generalization behavior of a surrogate model for a punch-bending scenario by generating plausible counterfactuals, and show that these also yield better process configurations compared to a baseline.