Fully Convolutional Neural Network for Semiconductor Image Simulation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for generating simulated images from input images in semiconductor manufacturing are inefficient, requiring extensive computation and involving cropping and reconstructing steps, which complicate software implementation and limit the size of input images.

Innovation Solution

A system utilizing fully convolutional deep generative models enables the generation of simulated images from arbitrarily sized images, eliminating the need for cropping and reconstructing steps, and significantly reducing computational time by 100x to 1000x, using neural networks with convolutional layers instead of fully connected layers.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If current methods are used to generate simulated images from input images, then the images can be processed, but extensive computation is required and the size of input images is limited

Engineering Contradiction:
Improvecomputational efficiencyVSAvoidsoftware implementation complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent applies segmentation by dividing the input image into smaller patches that can be processed independently by the neural network. The system processes multiple patches in parallel and then reconstructs the complete simulated image from these patches, enabling efficient handling of large images while maintaining computational feasibility

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent uses copying by creating multiple simulated images from the same input image through the neural network. The system generates synthetic images that replicate the characteristics of real inspection images, allowing extensive computational experiments and training without requiring additional physical specimens

Inventive Principle:
Principle #26Copying

2Adaptability or versatility

If current methods are used to generate simulated images, then images can be transformed, but cropping and reconstructing steps are required which complicate software implementation

Engineering Contradiction:
Improveinput image size flexibilityVSAvoidsoftware implementation complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent segments the image processing into patch extraction, individual patch processing through the neural network, and reconstruction of the complete image. This segmentation enables the system to handle arbitrarily sized input images by processing them in manageable chunks without requiring complex cropping and reconstructing steps

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The neural network is designed with universal functionality to process images of any size through the patch-based approach. The same network architecture and processing pipeline can handle different input dimensions without requiring separate implementations or complex adaptive logic

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Measurement precision

If current methods are used to generate simulated images, then transformation can be achieved, but computational time is excessive

Engineering Contradiction:
Improveimage transformation accuracyVSAvoidcomputational time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-training the neural network on large datasets of real inspection images and their corresponding simulated images. This pre-training establishes the foundational transformation capabilities, allowing the network to generate accurate simulated images rapidly during actual use without requiring extensive computation at runtime

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3465174B1Generating simulated images from input images for semiconductor applications
Publication Date: 2022.03.09 KLA CORP
  • EP3465174B1 patent drawingFigure 1
  • EP3465174B1 patent drawingFigure 1a
  • EP3465174B1 patent drawingFigure 2

AI summary

Methods and systems for generating a simulated image from an input image are provided. One system includes one or more computer subsystems and one or more components executed by the one or more computer subsystems. The one or more components include a neural network that includes two or more encoder layers configured for determining features of an image for a specimen. The neural network also includes two or more decoder layers configured for generating one or more simulated images from the determined features. The neural network does not include a fully connected layer thereby eliminating constraints on size of the image input to the two or more encoder layers.