TY - GEN
T1 - A validation neural network (VNN) metamodel for predicting the performance of deep generative designs
AU - Cunningham, James
AU - Tucker, Conrad S.
N1 - Publisher Copyright:
Copyright © 2018 ASME
PY - 2018
Y1 - 2018
N2 - This work presents a deep neural network method for approximating the performance of generated design concepts. This deep learning meta-modeling approach minimizes the need for costly simulations that test for design concept feasibility by discovering the visual features of a design that correlated to good and bad performance. These form-function relationships are discovered by simply observing the pixels of images of many candidate designs and their corresponding performance in a simulation environment. As opposed to existing metamodeling techniques, this evaluation is agnostic to the simulation environment and applicable to any design space in which form and function are closely linked. A case study is presented in which 2D sketches of boats generated from a deep generative model are evaluated in a simulation environment based on their ability to travel through water without sinking as well as their speed of travel. It is shown through simulation that 57.5% of the designs, which are validated according to their form during the generation process, fail in their intended function. Additionally, the trained VNN is able to classify designs it has never seen before as successful or failing with an accuracy of 86.6% and an F1-Score of 0.806.
AB - This work presents a deep neural network method for approximating the performance of generated design concepts. This deep learning meta-modeling approach minimizes the need for costly simulations that test for design concept feasibility by discovering the visual features of a design that correlated to good and bad performance. These form-function relationships are discovered by simply observing the pixels of images of many candidate designs and their corresponding performance in a simulation environment. As opposed to existing metamodeling techniques, this evaluation is agnostic to the simulation environment and applicable to any design space in which form and function are closely linked. A case study is presented in which 2D sketches of boats generated from a deep generative model are evaluated in a simulation environment based on their ability to travel through water without sinking as well as their speed of travel. It is shown through simulation that 57.5% of the designs, which are validated according to their form during the generation process, fail in their intended function. Additionally, the trained VNN is able to classify designs it has never seen before as successful or failing with an accuracy of 86.6% and an F1-Score of 0.806.
UR - http://www.scopus.com/inward/record.url?scp=85057020232&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85057020232&partnerID=8YFLogxK
U2 - 10.1115/DETC201886299
DO - 10.1115/DETC201886299
M3 - Conference contribution
AN - SCOPUS:85057020232
T3 - Proceedings of the ASME Design Engineering Technical Conference
BT - 44th Design Automation Conference
PB - American Society of Mechanical Engineers (ASME)
T2 - ASME 2018 International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, IDETC/CIE 2018
Y2 - 26 August 2018 through 29 August 2018
ER -