Automatic Layout Generation with Applications in Machine Learning Engine Evaluation

Haoyu Yang, Wen Chen, Piyush Pathak, Frank Gennari, Ya-Chieh Lai, Bei Yu

TL;DR

This paper introduces a novel toolkit for automatically generating complex lithography layout patterns, specifically for metal and via layers, under design rule constraints. They then use this toolkit to create a new benchmark dataset to evaluate the robustness of existing machine learning-based hotspot detectors, revealing significant performance degradation on these more complex, generated patterns.

What Problem Does This Paper Address?

Machine learning (ML) models have shown great promise in detecting lithography hotspots, crucial for ensuring chip manufacturing yield. However, most existing research and benchmarks focus on well-established, simpler patterns from legacy technology nodes. This leaves a critical question unanswered: how well do these ML models generalize to the more complex and varied patterns encountered in modern chip designs? The authors aim to bridge this gap by creating a tool to generate diverse, realistic layout patterns and using them to rigorously test the robustness and generalization capabilities of state-of-the-art hotspot detection models.

Methods

The core of the paper is a layout generation toolkit built using Python and the KLayout platform. This toolkit can synthesize:

  • Unidirectional Metal Layers: It generates metal wire shapes based on a set of design constraints including wire width, track pitch, minimum/maximum line-end to line-end distance, and minimum/maximum wire length. The generation process is track-by-track, allowing for parallel processing.
  • Via Layers: This module generates vias (connections between metal layers) by identifying potential overlap regions between metal layers. A key innovation here is the use of "assist wires" to simplify the process of satisfying enclosure constraints and then filtering candidate via locations based on these constraints and via pitch requirements. The generation is guided by parameters like via size, density, enclosure distances, and via pitch.

The generated layout patterns are then used as a case study to evaluate existing ML-based hotspot detectors. The authors specifically examine models employing techniques like Batch Biased Learning (BBL), Smooth Boosting, and AUC optimization.

Datasets

The paper introduces a new benchmark dataset generated by their layout generation toolkit. This dataset consists of "DRC-clean via layout clips" specifically designed to challenge existing hotspot detection models. The generation process allows for control over various layout properties, creating a more complex and diverse set of patterns compared to traditional public benchmarks. The authors do not explicitly list a separate training dataset, but imply that the evaluated models were trained on existing public benchmarks.

Key Insights and Ideas

  • The "Real World" is Complex: Existing ML hotspot detectors, while performing well on simplified benchmarks, struggle significantly when faced with the more intricate and varied patterns generated by the toolkit. This highlights a critical gap between lab performance and real-world applicability.
  • Robustness is Key: The study underscores the necessity of developing ML models that are not only accurate on known patterns but also robust and generalizable to a wider range of complex layouts.
  • Assist Wires Simplify Via Generation: The novel use of "assist wires" in the via generation process simplifies the handling of enclosure constraints, offering a more practical approach to generating valid via candidates.
  • Via Density Matters: The generated via patterns, particularly those with varying via densities, proved to be a challenging test case for the evaluated ML models.

Key Results

The evaluation experiments demonstrated performance degradation across several representative ML-based hotspot detectors when tested on the generated complex via patterns. While specific quantitative results are presented in Table IV, the general trend shows that accuracy can drop significantly, and false alarm rates can increase substantially, for some models on these challenging datasets. For instance, the average accuracy across tested models on the generated data was often much lower than what might be expected from their performance on simpler benchmarks. The paper highlights that some models exhibit unstable performance, with high variance across different experimental runs.

Limitations and Future Work

  • Limited Scope of Generation: The current toolkit primarily supports unidirectional metal layers and via/contact generation. Future work could extend this to more complex multi-directional metal layers and other layout structures.
  • Focus on Via Layers: The case study specifically focused on via layer layouts. Expanding the evaluation to complex metal layer patterns would provide a more comprehensive assessment.
  • Need for More Robust Models: The paper strongly suggests that continuous research is needed to improve the robustness and generality of ML models for hotspot detection, enabling their reliable integration into real-world Design for Manufacturability (DFM) flows. This implies developing new architectures, training strategies, or feature extraction methods that are less sensitive to pattern complexity.

Summary updated

Abstract

Machine learning-based lithography hotspot detection has been deeply studied recently, from varies feature extraction techniques to efficient learning models. It has been observed that such machine learning-based frameworks are providing satisfactory metal layer hotspot prediction results on known public metal layer benchmarks. In this work, we seek to evaluate how these machine learning-based hotspot detectors generalize to complicated patterns. We first introduce a automatic layout generation tool that can synthesize varies layout patterns given a set of design rules. The tool currently supports both metal layer and via layer generation. As a case study, we conduct hotspot detection on the generated via layer layouts with representative machine learning-based hotspot detectors, which shows that continuous study on model robustness and generality is necessary to prototype and integrate the learning engines in DFM flows. The source code of the layout generation tool will be available at https://github.com/phdyang007/layout-generation.
BibTeX
@misc{yang2019automatic,
  author = {Haoyu Yang and Wen Chen and Piyush Pathak and Frank Gennari and Ya-Chieh Lai and Bei Yu},
  title = {{Automatic Layout Generation with Applications in Machine Learning Engine Evaluation}},
  year = {2019},
  eprint = {1912.05796},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1912.05796},
}

Loading related content...

Figure