A review of deep neuron network applications in extrusion die design

Ensuring the quality of extrusion product necessitates meticulous die design, typically achieved through simulation iterations and/or experimental trials. However, this process is not only time-consuming but also costly. Despite substantial research utilizing historical data and finite element analysis (FEA) to elucidate design guidelines and principles, and the existence of numerous empirical equations guiding die design, it remains more of an art reliant on the designer's experience. In contrast, Deep Neural Networks (DNNs) have the capability to capture design experience with appropriately defined inputs and outputs, transforming it into abstract features for further application. With the advancement of DNNs, the automatic generation of precise die designs has become achievable. Several research studies have been undertaken to enhance die design through the application of DNNs, particularly Convolutional Neural Networks (CNNs). CNNs, a machine learning method commonly applied to extract information from images, have been utilized due to the intricate nature of die design. Given the inherent characteristics of DNNs, a significant challenge in incorporating DNNs into die design lies in devising a scheme to abstract 3D die designs for defining inputs without loss of information. Various methods exist for handling 3D objects, such as point clouds or projecting 3D objects into 2D depth graphs. Nonetheless, most of these methods prove challenging to implement effectively in the realm of die design. Another challenge stems from the overall complexity of the extrusion die. While most research has focused on automatically designing specific features of the die, such as the location or shape of portholes, there have also been data-driven studies attempting to generate entire die designs using historical data. This paper aims to review the status of the application of DNNs in hot extrusion die design and explore the further potential in this field.