Learning to Optimize: Applications in Physical Designs and Manufacturing

Abstract

Engineering and physical science often involve the design and manufacturing of physical devices. Conventionally, optimizing the physical design and the manufacturing process heavily relies on domain expertise and requires an iterative trial-and-error process conducted by human experts before achieving desired performance. Though numerical optimization methods have been developed for assisting domain experts, they often rely on heuristics that could be sub-optimal for the tasks of interest. Additionally, the performance of conventional optimization methods does not improve as more tasks are solved. This dissertation frames optimization as a learning problem, i.e., learning-to-optimize, where machine learning models are trained to solve optimization problems.

We propose three methods for solving practical optical inverse design and manufacturing problems. Our first proposed method OML-PPO treats optical multilayer thin films design tasks as sequence generation problems. Sequence generation networks that can discover optimal designs corresponding to user-specified optical properties are trained by reinforcement learning. The proposed method has been used to design a perfect broadband absorber with reflectance higher than 99%, an incandescent light bulb filter that can enhance the brightness by 16.3 times, and chrome replacement coatings with a close appearance to chrome films. Instead of targeting generic optical design tasks, our second method NEUTRON is a hybrid machine learning and optimization approach for efficiently designing optical multilayer thin films for structural color applications. By modeling the structural color inverse design as a bi-level optimization problem, NEUTRON applies machine learning models for fast, approximate material selection and particle swarm optimization for an exact search of the optimal thickness. We applied NEUTRON to both the chrome replacement coating and image reconstruction tasks. The results show that NEUTRON can achieve more accurate designs than machine learning or optimization alone. Thanks to the high efficiency of NEUTRON, we can reconstruct images with more than 200,000 pixels within a few hours. Our third method M2BOP addresses the costly data collection problem common in manufacturing problems by combining meta-learning and model-based offline reinforcement learning. By learning a meta environment model using offline data collected from relevant tasks, M2BOP can solve new tasks efficiently with a handful of data. On robot locomotion control tasks, M2BOP outperforms baseline approaches, especially on offline datasets that contain sub-optimal demonstrations.