Efficiency comparison of design of experiments and machine learning techniques with small dataset

Abstract

Design of Experiments (DOE) is a powerful methodology that allows for the extraction of significant information with a minimal number of experiments. DOE is particularly valuable when the cost of experiments is high and the feasibility of conducting trials is complex. This paper aims to compare the predictive performance of a regression model obtained through DOE with three artificial intelligence methods: Random Forest, Extreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM), specifically in the context of the injection molding process for predicting the quality defect of shrinkage. In the first step, data from 64 experiments is generated using a factorial design of experiments within the injection molding process. The second step involves modeling this data through regression analysis and machine learning models. The results indicate that th e Random Forest model, with an R-squared of 95.74%, outperforms XGBoost, the DOE-based regression model, and SVM. The results underscore the potential of artificial intelligence methods to efficiently handle small datasets and achieve greater predictive accuracy than traditional DOE approaches in the context of complex industrial settings.