Data-driven models for surface enhancement processes

• Main Author: Lee, Lukas Yi Liang
• Other Authors: Upadrasta Ramamurty
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence; Engineering::Materials::Material testing and characterization; Engineering::Mechanical engineering
• Online Access: https://hdl.handle.net/10356/150411
• Institution: Nanyang Technological University

Shot Peening (SP) and Laser Shock Peening (LSP) are surface enhancement processes that help improve fatigue life. However, these processes are complex as there are many input parameters and each input will affect the results differently. Thus, there was a need for a simpler and faster way of conducting these experiments. Finite Element Method (FEM) is widely used to simulate the complex peening processes as it is a cheaper alternative to doing the experiment physically and also allows the user to simulate many different scenarios with different input settings. Even though FEM simulations are faster than experiments, performing repeated calculations can still be time consuming. For example, optimizing a process could require running hundreds or thousands of simulations, which remains challenging. To address this concern, machine learning algorithms that have already been used to optimize different processes were adopted in this project. With machine learning, tests can be done quickly as it takes a few minutes for the program to run and produce a result compared to FEM simulations which can take hours. In this study, regression techniques were explored, using Python, and implemented to the mechanical processes mentioned above. Machine learning was used to make a data-driven model that acts as a surrogate to help with processing the results from the FEM simulations. The few regression methods that were explored were Linear Regression, Gradient Boosting Regression (GBR), Gaussian Process Regression (GPR) and Multilayer Perceptron (MLP). It was found that the linear regression method could not produce very accurate results compared to the other methods due to the non linear nature of the data (e.g. how the material responds to the different inputs). On the other hand, results from the other 3 methods were comparable to each other with GBR being able to produce the most accurate results as well as present more information for analysis through the use of error bars. Particle Swarm Optimisation (PSO) and error propagation were also used to further analyse the data from the FEM simulations.