A brand new technical paper titled “An Adversarial Energetic Sampling-based Information Augmentation Framework for Manufacturable Chip Design” was printed by researchers on the College of Texas at Austin, Nvidia, and the California Institute of Know-how.
“Lithography modeling is a vital drawback in chip design to make sure a chip design masks is manufacturable. It requires rigorous simulations of optical and chemical fashions which might be computationally costly. Current developments in machine studying have supplied various options in changing the time-consuming lithography simulations with deep neural networks. Nonetheless, the appreciable accuracy drop nonetheless impedes its industrial adoption. Most significantly, the standard and amount of the coaching dataset immediately have an effect on the mannequin efficiency. To deal with this drawback, we suggest a litho-aware knowledge augmentation (LADA) framework to resolve the dilemma of restricted knowledge and enhance the machine studying mannequin efficiency. First, we pretrain the neural networks for lithography modeling and a gradient-friendly StyleGAN2 generator. We then carry out adversarial lively sampling to generate informative and artificial in-distribution masks designs. These artificial masks pictures will increase the unique restricted coaching dataset used to finetune the lithography mannequin for improved efficiency. Experimental outcomes show that LADA can efficiently exploits the neural community capability by narrowing down the efficiency hole between the coaching and testing knowledge cases.”
Discover the technical paper here. Printed October 2022 (preprint).
Authors: Mingjie Liu, Haoyu Yang, Zongyi Li, Kumara Sastry, Saumyadip Mukhopadhyay, Selim Dogru, Anima Anandkumar, David Z. Pan, Brucek Khailany, Haoxing Ren. arXiv:2210.15765v1