Batch constrained multi-objective Bayesian optimization using the example of ultrasonic wire bonding

Setting optimum process parameters for complex manufacturing processes such as ultrasonic wire bonding is already challenging for one target variable. Due to numerous influencing physical factors, such processes often lack the necessary detailed physical models. Due to the lack of these models, such processes cannot be adjusted to an optimum using classic optimization methods. In manufacturing processes in particular, the process must be optimized with regard to several objectives such as process time or quality. A popular method for optimizing such processes without models is multi-objective Bayesian optimization. For this purpose, surrogate models in the form of Gaussian processes are used in combination with a multi-objective acquisition function. In real processes, parallel function evaluation offers advantages in terms of trial efficiency and scalability. In this paper, we present a new algorithm for batch-constrained Bayesian multi-objective optimization. With this algorithm, an arbitrary number of function evaluations per iteration can be specified, whereby a more efficient determination of the Pareto front approximation can be achieved in real applications. For this purpose, the Expected Hypervolume Improvement is extended by a term to consider a quality criterion. Using the ultrasonic wire bonding process as an example, we experimentally prove that the proposed framework is able to approximate the Pareto front of the system with only a few function evaluations. For this purpose, the process capability index should be maximized and the process time minimized.