Additive manufacturing: directed energy deposition process parameters optimization via machine learning

• Main Author: Chen, Minyang
• Other Authors: Li Hua
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering; Engineering::Mechanical engineering
• Online Access: https://hdl.handle.net/10356/168284
• Institution: Nanyang Technological University

Additive manufacturing (AM) is a rapidly growing industry that creates intricate industrial parts. One of the methods employed by AM is directed energy deposition (DED), which involves melting metal powder with a laser beam to form various components. The process of 3D printing is known to be complex and involves a multitude of parameters, making the optimization of the process a significant challenge. Therefore, the application of machine learning (ML) is being considered as a potential solution. ML is especially useful in analyzing structured data and is ideal for tasks such as regression. In the case of AM, ML has proven to be valuable in optimizing design and manufacturing efficiency. Despite the growing interest in the DED process, there is a dearth of research on the optimization of DED process parameters for mechanical properties via ML. Adjusting process parameters can enhance the mechanical properties of printed components. However, generalizing the results of a single experiment to other scenarios can be difficult. As a result, many researchers have had to repeat their experiments to optimize the production of parts with mechanical properties beyond the initial experimental range. To address the need for efficient and effective techniques to optimize DED process parameters for predictions outside of the original experiment design range, a potential solution is to employ ML. For this project, Response Surface Methodology (RSM) and 9 different ML models were utilized to predict the ultimate tensile strength, yield strength, and average hardness of stainless steel 316L (SS316L) outside of the RSM design range. The RSM model was fitted on real data points while the machine learning models were trained with a combination of synthetic data points generated from the RSM model and real data points. The best performing model was chosen to predict the mechanical properties, and a Python script was created to help identify the optimal process parameters for achieving the desired mechanical properties. This project revealed that there is a highly complex and nonlinear correlation between the input process parameters and the resulting mechanical properties. The Keras Neural Network model demonstrated superior performance among the tested models, owing to its fine-tuning capability. The Scikit learn library models also performed well, closely trailing the Keras Neural Network model. Conversely, the Minitab RSM model exhibited lower performance on average when compared to the Keras Neural Network model and the Scikit learn library models. The use of a Python script for identifying optimal process parameters in the DED process via ML allowed for increased efficiency and the production of parts with better mechanical properties outside of the RSM design range.