• Home
  • :
  • Community
  • :
  • Blogs
  • :
  • Silicon Signoff
  • :
  • Pattern Technology Applied to Machine Learning-based Hotspot…

Silicon Signoff Blogs

Philippe Hurat
Philippe Hurat
20 Feb 2019
Subscriptions

Get email delivery of the Cadence blog featured here

  • All Blog Categories
  • Breakfast Bytes
  • Cadence Academic Network
  • Cadence Support
  • Custom IC Design
  • カスタムIC/ミックスシグナル
  • 定制IC芯片设计
  • Digital Implementation
  • Functional Verification
  • IC Packaging and SiP Design
  • Life at Cadence
  • The India Circuit
  • Mixed-Signal Design
  • PCB Design
  • PCB設計/ICパッケージ設計
  • PCB、IC封装:设计与仿真分析
  • PCB解析/ICパッケージ解析
  • RF Design
  • RF /マイクロ波設計
  • Signal and Power Integrity (PCB/IC Packaging)
  • Silicon Signoff
  • Spotlight Taiwan
  • System Design and Verification
  • Tensilica and Design IP
  • Whiteboard Wednesdays
  • Archive
    • Cadence on the Beat
    • Industry Insights
    • Logic Design
    • Low Power
    • The Design Chronicles

Pattern Technology Applied to Machine Learning-based Hotspot Prediction

I have been working on DFM solutions for (too) many years and the objective hasn’t change: Detect or predict design-process weakpoints also known as hotspots, to limit systematic yield loss in semiconductor manufacturing.

Traditional methods, currently used in production and released by foundries to their customer designers, are;

  • Collection of known hotspots to build a weakpoint database and use it to build detection models (patterns) and CAD solutions to fix them during the design and mask tapeout cycles. This approach is limited to finding known weakpoints and cannot predict unknown hotspots.
  • Using process simulation(s) to predict such weakpoints. This method has predictive capabilities, but it is very compute intensive and cannot be easily deployed during the design signoff cycle. Also, it is very difficult to deliver such simulation kit without revealing the process information.

As an alternative to the time expensive process simulations, AI seems a good fit for hotspot prediction if it has the predictive characteristics without the prohibitive runtime and risk associated with process IP disclosure. A variety of machine learning-based techniques have been proposed. Research works have focused on two important aspects: (1) finding an efficient representation of layout features and (2) developing machine learning models using readily available production data such as diagnostic test data.

Mainstream layout representations include density-based feature, pixel-based feature (layout clip images), frequency domain feature, concentric circle sampling (CCS) etc. Most of these representations are either suffering from information loss (e.g. density-based feature, and CCS), or are not storage efficient (e.g. images). To address these problems, we propose a convolutional neural network called Squish-Net where the input pattern representation is in an adaptive squish form. Here, the squish pattern representation used by Cadence Pattern Analysis is modified to handle variations in the topological complexity across a pattern catalog, which still allows high data compression with no information loss.

At SPIE, the Cadence DFM team will present a combination of the adaptive squish representation and associated ML solutions. Join the SPIE session “Hotspot Detection Using Squish-Net” on Thursday, Feb. 28. 1:30 – 3:30, if you want to understand how adaptive squish representation is used to achieve best-in-class hotspot detection accuracy on ICCAD-2012 dataset by incorporating a medium-sized convolutional neural network.

Figure 1: Squish-Net convolutional neural network

Tags:
  • pattern analysis |
  • machine learning |
  • silicon learning |
  • signoff |
  • yield |
  • design for manufacturing |
  • DFM |