Machine Learning Derived Graded Lattice Structures

Herein, we propose a new lattice generation strategy that is computationally cheaper and produces high-quality geometric definition based on Machine Learning (ML) when compared to traditional methods. To achieve the design of high-performance unit cells, firstly, the optimal mechanical property for each cell region is derived according to the loading condition and the reference density obtained utilising a conventional topology optimisation result. Next, a Neural Network (NN) is employed as an inverse generator which is responsible for predicting the cell pattern for the optimal mechanical property. Training data (~ 500) were collected from Finite Element (FE) analysis with varied cell parameters and then fed to the NN. With the help of ML, the time spent in building the inverse generator is significantly reduced. Furthermore, the ML-based inverse generator can handle different cell types rather than one specific type which facilitates the diversity and optimality of lattices.