TOPOLOGY OPTIMIZATION THROUGH DEEP NEURAL NETWORK FOR DIFFERENT MECHANICAL AND THERMOMECHANICAL PROBLEMS

In this present work, a neural network (NN) is trained to deal with the optimization process of topology optimization and generate optimized structures. The NN's activation functions are used to represent the popular Solid Isotropic Material with Penalization (SIMP) density field. Fourier projection based length scale control technique has been implemented to govern the minimum and maximum feature length in the optimized domain for manufacturability. And high-performance automatic differentiation (AD) library JAX has been used to build an end-to-end differentiable NN model. We demonstrate the application of this framework by solving several mechanical and thermomechanical compliance minimization problems. The results show that optimized designs obtained from the proposed NN based approach are comparable with the current mathematical programming based optimization approach.