Learning to Optimize: Applications in Physical Designs and Manufacturing

Passcode: 817134

Abstract:

Engineering and physical science involve designing and manufacturing physical devices, which often require an iterative trial-and-error process conducted by human experts before achieving the desired performance. Recently, researchers have successfully applied machine learning to learn models for solving optimization problems. However, optimizing physical designs and manufacturing processes is not trivial due to the unique challenges associated with complicated real-world optimization problems, including 1) large design space, 2) non-uniqueness of global optima, and 3) costly data collection.

Based on the optical inverse design and the manufacturing optimization problems, this dissertation investigates methods for addressing the issues mentioned above. We propose efficient learning-to-optimize methods based on 1) reinforcement learning and 2) hybrid machine learning and optimization methods for optical multilayer thin-film designs. The discovered designs can outperform the best designs found by human experts. In addition, we propose an offline meta-reinforcement learning algorithm to address sample complexity issues associated with learning-to-optimize algorithms. The proposed algorithm can efficiently adapt to unseen tasks with a handful of experience by learning from the logged observational data from related manufacturing tasks.

Chair: Professor L. Jay Guo