Lattice design recommendation for multi-loading structure using machine learning

• Main Author: Toh, Chiesa Jun Ee
• Other Authors: Yeong Wai Yee
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/167781
• Institution: Nanyang Technological University

The drive to reduce material usage and improve properties of structures has led to additive manufacturing techniques being used to develop lattice structures. However, designing the optimal lattice for specific applications is challenging due to high computational resources to evaluate and this prompted the exploration of machine learning methods for lattice design. In this study, a machine learning based approach will be implemented, leveraging on datasets of lattice unit cells of different geometry and strut thickness and their corresponding mechanical properties. Gaussian Process Regression outperformed Random Forest Regression in the interpolated and extrapolated prediction of maximum total deformation in their default parameters using cubic lattice unit cells dataset with an absolute percentage error difference of 0.54% and 8.39% compared to 11.40% and 36.54% respectively. Further analysis was done on the Gaussian Process Regression to show that RQ kernel performed better than the RBF kernel for prediction of maximum total deformation, maximum equivalent stress, maximum shear stress, maximum shear elastic strain for cubic lattice using two features of pressure loads where the mean absolute percentage difference does not exceed 1%. However, the prediction of three features for the cubic, cubic with bcc, octa and octet lattice unit cells which included the strut thickness was not optimal. To improve the performance, data augmentation techniques were implemented in conjunction with grid search to optimize the hyperparameters. The results showed improvements in the generalization errors of the predictions. Further studies are needed to explore the potential of other machine learning models or techniques to enhance performance of the Gaussian Process Regression model. However, the findings suggest that machine learning techniques hold potential in the utilization of optimizing lattice designs.