A Dual Neural Network Approach to Topology Optimization for Thermal-Electromagnetic Device Design

Abstract

Topology optimization for engineering problems often requires multiphysics (dual objective functions) and multi-timescale considerations to be coupled with manufacturing constraints across a range of target values. We present a dual neural network approach to topology optimization to optimize a 3-dimensional thermal-electromagnetic device (optical shutter) for maximum temperature rise across a range of extinction ratios while also considering manufacturing tolerances. One neural network performs the topology optimization, allocating material to each sub-pixel within a repeating unit cell. The size of each sub-pixel is selected with manufacturing considerations in mind. The other neural network is trained to predict performance of the device using extinction ratio and temperature rise over a given time period. Training data is generated using a finite element model for both the electromagnetic wave frequency domain and thermal time domain problems. Optimized designs across a range of targets are shown.