A new methodology of using design of experiments as a precursor to neural networks for material processing: extrusion die design

Abstract

Extrusion die design and making is an art and a science. In present day extrusions using composites, polymers, and other new alloys, the product geometries are extremely complicated. The flow analysis inside an extrusion die using finite element analysis (FEA) is tedious and time consuming. To optimize the design of a die one needs to perform hundreds of runs, requiring several weeks or months of computer time. In the past researchers have used neural networks (NN) to optimize the design and predict flow patterns for newly designed dies of similar geometries. But, even for NN it has been proven that one needs a few thousand runs to train a network and accurately predict the flow. This paper shows a new methodology of using design of experiments (DOE) as a precursor to identify the importance of some variables and thus reduce the data set needed for training a NN. Based on the DOE results, a neural network training set is generated with more variations for the most significant inputs. A comparison of design using only NN versus using DOE and then NN is shown. The results indicate a significant reduction in the size of the training set, the time required for training and improvement in accuracy of the predicted results. To reduce the analysis time, a newly developed upper bound technique was used for generating the training set. The DOE model is extremely fast and can be used for real time (online) control of the process.