Comparison of design of experiment methods for modeling injection molding experiments using artificial neural networks

Different design of experiments (DoE) strategies for generation of injection molding process data and later usage as training database, e. g. for an artificial neural network substitute model, are investigated and compared. The objective is to find the most efficient and effective strategies for the modeling. The effects of six injection molding parameters on the length, width and weight of polypropylene plate specimen are simulated with different DoE methods from the following categories: full and fractional factorial designs, central composite designs, orthogonal Taguchi arrays, o-optimal designs and a space-filling design. The prediction performance, e. g. of the artificial neural networks, for unknown test data is evaluated. A 2(6-3) fractional factorial design including a center point is very efficient for this modeling task, whereas an inscribed central composite design is most effective. The artificial neural network with the latter experimental design as training data achieves a coefficient of determination R-2 of 0.930 without a hyperparameter tuning to the specific data set.

Automated summary

Different design of experiments strategies for injection molding process data are investigated, with 2(6-3) fractional factorial design and an inscribed central composite design found to be the most effective for modeling tasks.