SUBJECT: Ph.D. Proposal Presentation
   
BY: Zhichao Wang
   
TIME: Wednesday, July 13, 2022, 9:00 a.m.
   
PLACE: https://bit.ly/3ORVGQH, N/A
   
TITLE: Generative Design using Deep Learning Methods for Functionality and Manufacturability
   
COMMITTEE: Dr. Rosen, David, Chair (ME)
Dr. Melkote, Shreyes (ME)
Dr. Choi, Seung-Kyum (ME)
Dr. Wang, Yang (CEE)
Dr. Gombolay, Matthew (COC)
 

SUMMARY

With the rapid development of web-based and cloud-based technologies, cyber manufacturing has emerged as a discipline to facilitate management and improve the productivity of geographically distributed manufacturing operations. However, conventional design methods are unsuitable for cyber manufacturing, which delays product development. Recently, generative design has begun to offer new opportunities to facilitate cyber manufacturing. The generative design can assist human labor in exploring design space in two directions: design for functionality (DFF) and design for manufacturability (DFM). Iterations between DFF and DFM will also be conducted to realize a balance between DFF and DFM. Although deep learning can be a good solution for speeding up the design procedure, previous work on a generative design by deep learning was limited, and more work is required before it can be embedded into cyber manufacturing.
Several subproblems are involved in the generative design process with deep learning, and these lead to the various issues that will be addressed in this research. To understand generative design with deep learning more clearly, four subprocesses need to be accomplished: (a) manufacturing processes selection in initial design; (b) design for functionality; (c) design for manufacturability; (d) iterations between DFF and DFM. To address the four subprocesses, three research questions will be investigated by this research:
•How do we select shape manufacturing processes with or without non-shape information like materials and tolerances?
•How to generate novel functionality design (fd) from manufacturable design (md) under unsupervised learning and different boundary conditions and load conditions?
•How to generate novel manufacturable design (md) from functional design (fd), especially when the design shapes are complex?
The first question will be answered by designing a neural network to select suitable manufacturable processes for the given design with and without non-shape information. The second question will be solved by embedding topology optimization into the neural network so that the output will have good functionality. We will deal with the last question by representing different manufacturing rules as different neural networks and applying them sequentially to generate manufacturable designs.