Author: Mohamed Baker Alawieh
Publisher:
Published: 2020
Total Pages: 0
ISBN-13:
DOWNLOAD EBOOK →
Consumer electronics have become an integral part of people’s life putting at their disposal immense computational power that provides numerous applications. This has been enabled by the ceaseless down scaling of integrated circuit (IC) technologies which keeps pushing the performance boundary. Such scaling continues to drive, as a byproduct, an up scale in the challenges associated with circuit design and manufacturability. Among the major challenges facing modern IC Computer Aided Design (CAD) are those related to manufacturing and yield which are manifested through: (1) expensive modeling and simulation (e.g. large and complex designs); (2) entangled design and manufacturability (e.g., yield sensitive to design patterns); and (3) strict design constraints (e.g., high yield). While these challenges associated with retaining the robustness of modern designs continue to exacerbate, Very Large-Scale Integration (VLSI) CAD is becoming more critical, yet more challenging. Parallel to these developments are the recent advancements in Machine Learning (ML) which have altered the perception of computing. This dissertation attempts to address the aforementioned challenges in VLSI CAD through machine learning techniques. Our research includes efficient analog modeling, learning-assisted physical design and yield analysis, and model adaptation schemes tailored to the ever-changing IC environment. With aggressive scaling, process variation manifests itself among the most prominent factors limiting the yield of analog and mixed-signal (AMS) circuits. In modern ICs, the expensive simulation cost is one of the challenges facing accurate modeling of this variation. Our study develops a novel semi-supervised learning framework for AMS design modeling that is capable of significantly reducing the modeling cost. In addition, a new perspective towards incorporating sparsity in the modeling task is proposed. At the lithography stage, resolution enhancement techniques in general, and Sub Resolution Assist Feature (SRAF) insertion in particular, have become indispensable given the ever shrinking feature size. While different approaches have been proposed for SRAF insertion, the trade-off between efficiency and accuracy is still the governing principle. To address this, we recast the SRAF insertion process as an image translation task and propose a deep learning-based approach for efficient SRAF insertion. Besides, with complex designs, challenges at the physical design stage have exacerbated. Therefore, across-layers information sharing has become imperative for timely design closure. Particularly, in modern Field Programmable Gate Array (FPGA) place and route flows, leveraging routing congestion information during placement has demonstrated imperative benefit. Our study develops a new congestion prediction approach for large-scale FPGA designs that achieves superior prediction accuracy. Moreover, during fabrication, a critical first step towards improving production yield is to identify the underlying factors that contribute most to yield loss. And for that, wafer map defect analysis is a key. We present a novel wafer map defect pattern classification framework using confidence-aware deep selective learning. The use of ML for CAD tasks has the promise of delivering better performance and efficiency. However, one of the main characteristics of the field is that it is evolving with a fast rate of change. Therefore, by the time enough data is available to train accurate models under a given environment, changes start to occur. In this sense, the frequent restarts limit the returns on developing ML models. To address this, we develop a framework for the fast migration of classification models across different environments. Our approaches are validated with extensive experiments where they proved capable of advancing the VLSI CAD flow
