Using compressed air in industrial processes is often accompanied by a poor cost-benefit ratio and a negative impact on the environmental footprint due to usual distribution inefficiencies. Compressed air-based systems are expensive regarding installation and lead to high running costs due to pricey maintenance requirements and low energy efficiency due to leakage. However, compressed air-based systems are indispensable for various industrial processes, like handling parts with Class A surface requirements such as outer skin sheets in automobile production. Most of those outer skin parts are solely handled by vacuum-based grippers to minimize any visible effect on the finished car. Fulfilling customer expectations and simultaneously reducing the running costs of decisive systems requires finding innovative strategies focused on using the precious resource of compressed air as efficiently as possible. This work presents a sim2real reinforcement learning approach to efficiently hold a workpiece attached to a vacuum suction cup cluster. In addition to pure energy-saving, reinforcement learning enables those agents to be trained without collecting extensive data beforehand. Furthermore, the sim2real approach makes it easy and parallelizable to examine numerous agents by training them in a simulation of the testing rig rather than at the testing rig itself. The possibility to train various agents fast additionally facilitates focusing on the robustness and simplicity of the found agents instead of only searching for strategies that work, making training an intelligent system scalable and effective. The resulting agents reduce the amount of energy necessary to hold the workpiece attached by more than 15% compared to a reference strategy without machine learning and by more than 99% compared to a conventional strategy.