Spectral Wafer Inspection Using Noise-Aware Deep Learning

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing technologies face challenges in efficiently monitoring and improving the fine pattern formation operations in semiconductor manufacturing, particularly in achieving high inspection speed and accuracy for complex structures and fine patterns with line widths of 10 nm or less.

Innovation Solution

A method involving a spectral optical system to measure the spectrum of a patterned wafer, combined with a deep learning model trained based on domain knowledge, including noise inducing factors, to analyze and predict pattern characteristics, thereby evaluating the pattern quality effectively.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If a spectral optical system is used to measure the spectrum of the patterned wafer, then the inspection speed is improved, but the measurement precision deteriorates due to noise sensitivity

Engineering Contradiction:

Improveinspection speedVSAvoidpattern measurement precision

Core Design Contradiction:

SpeedVSMeasurement precision

Solution Approach 1:

The deep learning model is trained in advance with augmented spectra that include noise inducing factors, so that when actual measurement is performed, the model has already learned to compensate for these noise factors, maintaining precision while using fast spectral optical measurement

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent converts the harmful noise factors into a benefit by deliberately incorporating them during the training phase through spectrum augmentation. The model learns to recognize and compensate for these noise patterns, transforming what would be measurement errors into training data that improves robustness

Inventive Principle:

Principle #22Blessing in disguise (Convert harm into benefit)

2Reliability

If domain knowledge including noise inducing factors is incorporated into training, then the reliability is improved, but the device complexity increases

Engineering Contradiction:

Improvepattern evaluation reliabilityVSAvoidtraining system complexity

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The patent changes the parameters of the training data by incorporating noise inducing factors such as wavelength shifts, intensity variations, and spectral distortions. This transforms the training process to produce a more robust model without requiring changes to the physical measurement system

Inventive Principle:

Principle #35Parameter changes

Applied Scientific Principles

This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.

Function Achieved in This Case

This approach enhances the reliability and accuracy of pattern evaluation, reduces sensitivity to noise, and improves the productivity of semiconductor manufacturing by enabling faster and more precise monitoring of fine pattern formation operations.

Implementation Method 1

measuring, with a spectral optical system, a spectrum of the patterned wafer

Methodology Applied

Scientific EffectSpectroscopy: Absorption Spectroscopy

Data Source

PatentUS12205040B2Method of training deep learning model for predicting pattern characteristics and method of manufacturing semiconductor device

Publication Date: 2025.01.21 SAMSUNG ELECTRONICS CO LTD

AI summary

A method of manufacturing a semiconductor device includes forming a pattern on a wafer, measuring a spectrum of the pattern on the wafer, with a spectral optical system, performing an analysis of the spectrum through a deep learning model for predicting pattern characteristics, the deep learning model being trained based on a domain knowledge, and evaluating the pattern on the wafer based on the analysis of the spectrum, wherein the domain knowledge includes a noise inducing factor of the spectral optical system.