Data-Driven Problem Classification and Algorithm Selection for Injection Molding Optimization

Injection Molding (IM) is an established and essential production process. Dynamic process changes according to market demands, and advances in additive manufacturing technologies enable both chances and challenges. New molds can be produced more quickly and cheaper, but the demand for ongoing parameter optimization due to the changes in the production process rises. Digital-Twin-based process simulation assists process engineers in estimating optimal parameter settings. Due to its complexity, manual optimization is inefficient. Selecting an appropriate optimization algorithm for this task is challenging. There is a need for automated data science pipelines, especially for small and medium-sized enterprises (SMEs) that often lack algorithm engineering and data science capabilities. This paper addresses the data-driven classification of different IM optimization problems retrieved by IM simulation. As these IM simulations are computationally expensive, we apply Gaussian Process Simulations (GPS) on a small number of samples of the IM simulations to compute cheap to evaluate test problems. We compare a single and a 2-stage variant of GPS, both in their conditional and unconditional forms. The single variant tends to preserve the global characteristics of the problem. The 2-stage variant adds a second simulation model based on the residuals of the first model to simulate the finer local structure of the problem. The retrieved test problems are classified using Exploratory Landscape Analysis (ELA) applied to several minimization objectives of the IM process: max. volume shrinkage [%], avg. volume shrinkage [%], required cooling time [s], max. warpage [mm], max. deformation [mm], and avg. shrinkage [mm]. Finally, benchmark experiments will be conducted to compare optimization algorithms on the test problems and relate the results to the retrieved ELA features. These results build the foundation for algorithm selection and online parameter optimization.