Heuristics and performance bounds for automated design

arxiv.org

11 points·LolWolf··5 comments

Hey all! Author here; I'd be happy to answer any questions about this or the general idea of physical design! (i.e., the idea of automatically designing physical devices by optimization.)

great job, can you give me some resources to get in touch with the field of automated design because I want really to see the potential of using technics such as AI and DL for creation or optimization tasks.

Thanks!

This is a good question; AI and DL has been used for a little bit in physical design (in a scalable way starting roughly the past few years or so). The paper has some references at the top of page 26), but [0] and [1] have decent overviews as well!

-----

[0] https://www.degruyter.com/view/journals/nanoph/9/5/article-p...

[1] https://pubs.acs.org/doi/10.1021/acsphotonics.7b01377 (an arXiv link is found here: https://arxiv.org/pdf/1710.04724.pdf)

thanx lot. because I get fascinated by the idea of generative design and how we can use it to create a very optimized product using only computational methods. I am doing a master's degree in AI and I want to involve more in this area.

For sure! Another thing to note is that, at the moment, the usual ML/AI approaches fall very short of the classical optimization methods, so there is still plenty of work to be done.

For example, classical optimization methods can approximately optimize a design with ~10 million parameters in a few hours on a decent machine, while the current AI state of the art (notably using DL and some GANs) can only optimize designs with 10-100s of parameters, maybe 1000s and take many days to train.

Reply on news.ycombinator.com