Semiconductor Pattern Alignment Using Deep-Learned Image Reconstruction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional semiconductor alignment methods often fail to accurately align patterns due to mismatched tendencies between optical images and CAD images, leading to insufficient alignment accuracy.

Innovation Solution

A semiconductor inspection method utilizing a learning model based on deep learning to reconstruct a high-resolution image from a low-resolution image, allowing for accurate alignment by preferentially using regions with high certainty in the reconstructed image.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional alignment methods are used to align optical images with CAD images, then alignment can be performed, but the alignment accuracy is insufficient due to mismatched pattern tendencies between the two image types

Engineering Contradiction:
Improvealignment accuracyVSAvoidpattern matching reliability
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent transforms the low-resolution optical image into a high-resolution image by changing the resolution parameter. This allows the optical image to match the CAD image in terms of detail level and pattern tendency, thereby improving alignment accuracy. The transformation process adjusts the image parameters (resolution, clarity) to make them comparable with CAD images.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If deep learning-based image transformation is applied to convert low-resolution images to high-resolution images, then alignment accuracy is improved, but computational complexity and processing time increase

Engineering Contradiction:
Improvealignment accuracyVSAvoidimage processing complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent performs preliminary actions by pre-processing the optical image through deep learning-based transformation before the alignment process. The low-resolution image is transformed into a high-resolution image in advance, and the transformation results are stored or cached. This preliminary transformation simplifies the subsequent alignment process by providing pre-enhanced images that require less complex processing during actual alignment operations.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3958210B1Method for inspecting semiconductor and semiconductor inspecting device
Publication Date: 2025.05.28 HAMAMATSU PHOTONICS KK
  • EP3958210B1 patent drawingFigure 1
  • EP3958210B1 patent drawingFigure 2
  • EP3958210B1 patent drawingFigure 3(a)~3(c)

AI summary

A semiconductor inspection method by an observation system 1 includes a step of acquiring a first pattern image showing a pattern of a semiconductor device S, a step of acquiring a second pattern image showing a pattern of the semiconductor device S and having a different resolution from a resolution of the first pattern image, a step of learning a reconstruction process of the second pattern image using the first pattern image as training data by machine learning, and reconstructing the second pattern image into a reconstructed image having a different resolution from a resolution of the second pattern image by the reconstruction process based on a result of the learning, and a step of performing alignment based on a region calculated to have a high degree of certainty by the reconstruction process in the reconstructed image and the first pattern image.