Session: DAC-25-01: AI-Driven Design Innovation

Paper Number: 117830

117830 - A Generic Gan Model for Multi-Criteria Generative Design 

Recently, Generative Adversarial Networks (GANs) have been the subject of tremendous success in generating designs for various applications. They are, however,  lacking certain key characteristics which are needed for non-optimization generative design tasks such as the early stages of visual design. This paper presents a novel GAN-based design concept recommender which augments the evaluation network of StyleGAN2, an state-of-the-art GAN model, to address the inherent lack of certain characteristics necessary for creatively demanding design tasks generative design. The proposed approach regularizes the loss function using multiple design-specific measures, namely Structural Similarity Index (SSIM) to preserve the geometrical constraints in the solution given a defined silhouette, Local Outlier Factor (LOF) to provide guidance in terms of finding novel solutions, and Deep Multimodal Design Evaluation (DMDE) to increase the diversity and desirability of the generated solutions. To test the efficacy of the proposed model, a large set of sneakers is used as the dataset and both visual and quantitative results are reported. The quantitative results suggest a 9% improvement in terms of compatibility with the provided geometrical constraint and a 29% improvement in terms of novelty. The results demonstrate that the proposed platform outperforms StyleGAN2 as the baseline, thereby confirming its potential for generative design tasks.

Presenting Author: Parisa Ghasemi Northeastern University

Presenting Author Biography: Parisa Ghasemi is a PhD student at the Mechanical and Industrial Engineering Department at Northeastern University. Her research background is on developing Generative Machine Learning Models as well as adapting them for real-world applications. Currently, Parisa’s focus is on product design automation using Generative Design methods and especially, Generative Adversarial Networks to fortify design teams with AI tools capable of producing novel design concepts.

Authors:

Parisa Ghasemi Northeastern University
Mohsen Moghaddam Northeastern University