The identification of optimal production settings holds large potential for improving operational efficiency, reducing costs, increasing productivity, and enhancing product quality. A group of researchers from Fraunhofer IPT, IME and ILT developed a framework with which AI methods and optimization algorithms can assist production experts in optimal process design.

With limited data and the ability to perform new experiments, Bayesian optimization allows for sequential and highly efficient process improvements. We show that, de-pending on the complexity of the process, the optimal parameter set can be found after about ten experiments. If, however, only historical data is available, it is favorable to first train machine learning models and apply black-box optimization algorithms afterwards. This can be less efficient and accurate but does not require con-ducting expensive experiments.

To also use the software for sustainability improvements of production processes, the team is currently working on multi-objective optimization and on better integration of expert knowledge.