Injection mold design of reverse engineering using injection molding analysis and machine learning

Abstract

Plastic composites are used in vehicle components to improve fuel efficiency. Thus, the warpage of injection-molded plastic parts has become a quality issue. Factors, such as product shape and thickness, resin, and other injection molding conditions, can be modified to improve the warpage problem. However, if these factors are set with no possible adjustments, reverse engineering may be required. Reverse engineering is a difficult process that requires many trials and errors; thus, it is only used as a last resort. With respect to the warpage issue, reverse engineering considers the following: (1) Predicting and (2) modeling the warpage in opposite directions. Autodesk Moldflow Insight accommodates these key considerations, but many researchers are reluctant to use it. Although existing injectionmolding analysis programs are mainly used to predict qualitative results, computer-aided engineering (CAE) for reverse engineering requires quantitative analysis. Hence, the considerations are different from the existing analyses. An error in warpage prediction may lead to a costly mold modification because of the molds' complex structures. Quantitative warpage prediction for reverse engineering depends on process variables; thus, understanding how warpages are affected by uncertain process variables is important to improve the reliability of reverse engineering. Moreover, even if appropriate process variables are set, they cannot be applied due to tolerance in lengths. For this reason, mold shrinkage must be identified before designing a mold. This study conducted injection molding analysis for a radiator tank that uses glass fiber-reinforced plastic using Autodesk Moldflow Insight 2018.2. Data for warpage prediction were generated in accordance with five process variables to identify the relationship between the level of warpage and process variables. CAE also showed the level of mold shrinkage that can reduce warpage. In addition, a predictive model was created using the multilayer perceptron (MLP)- supervised learning technique, which is a deep learning method for artificial neural networks. The predictive model was compared with typical regression models, such as polynomial regression (also known as response surface model), EDT and RBF, to determine the optimal approximation model. The real modeling time for a radiator tank product is 1 h, but the MLP approximation model required only 1 min and 8 s to perform 8530 iterations with a similar reliability.