Physics-Based Image Perturbation for Semiconductor Defect Detection

Solution Overview

Problem

Current methods for identifying defects in semiconductor devices are inefficient due to the reliance on simulations and machine-learning algorithms that require substantial data, making it challenging to accurately determine defects in wafers with few or infrequent defects.

Innovation Solution

A system and method that uses physics-based image perturbations to generate augmented images of defects, which are then used to create a machine learning classifier for efficient defect identification in semiconductor specimens.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If machine-learning algorithms are used to determine defects, then defect identification accuracy can be improved, but the requirement for substantial training data increases, making it ineffective for wafers with few or infrequent defects

Engineering Contradiction:

Improvedefect identification accuracyVSAvoidtraining data quantity

Core Design Contradiction:

Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent creates synthetic copies of defect images through physics-based simulations. These simulated defect images are generated by modeling how defects would appear under various imaging conditions, effectively copying and multiplying the limited real defect data into a substantial training dataset for machine learning algorithms.

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The patent performs preliminary defect simulation and image generation before actual defect detection. By pre-generating augmented defect images through physics-based models, the system prepares sufficient training data in advance, eliminating the need for large quantities of real defect samples during the actual detection process.

Inventive Principle:

Principle #10Preliminary action

2Measurement precision

If traditional simulation methods are used to determine defect characteristics, then defect location and structure can be identified, but the process is time-consuming and reduces productivity

Engineering Contradiction:

Improvedefect characterization accuracyVSAvoiddefect detection throughput

Core Design Contradiction:

Measurement precisionVSProductivity

Solution Approach 1:

The patent replaces traditional computational simulation methods with physics-based image perturbation models. Instead of running complex simulations to generate defect images, the system uses analytical physics models that rapidly generate augmented images by applying mathematical transformations to real defect images, significantly reducing computation time while maintaining accuracy.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent changes the approach from simulating physical defect formation to transforming existing defect images through parameter-based perturbations. By adjusting image parameters such as brightness, contrast, and geometric transformations based on physics models, the system rapidly generates diverse training images without time-consuming simulations.

Inventive Principle:

Principle #35Parameter changes

3Reliability

If more real defect data is collected to train machine learning algorithms, then model reliability improves, but the time and resources required for data collection increase

Engineering Contradiction:

Improvemodel reliabilityVSAvoiddata collection time

Core Design Contradiction:

ReliabilityVSLoss of time

Solution Approach 1:

The patent creates synthetic copies of defect images through physics-based simulations. These simulated defect images are generated by modeling how defects would appear under various imaging conditions, effectively copying and multiplying the limited real defect data into a substantial training dataset for machine learning algorithms.

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The patent performs preliminary defect simulation and image generation before actual defect detection. By pre-generating augmented defect images through physics-based models, the system prepares sufficient training data in advance, eliminating the need for large quantities of real defect samples during the actual detection process.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS11676264B2System and method for determining defects using physics-based image perturbations

Publication Date: 2023.06.13 KLA CORP

AI summary

A system for characterizing a specimen is disclosed. In one embodiment, the system includes a characterization sub-system configured to acquire one or more images a specimen, and a controller communicatively coupled to the characterization sub-system. The controller may be configured to: receive from the characterization sub-system one or more training images of one or more defects of a training specimen; generate one or more augmented images of the one or more defects of the training specimen; generate a machine learning classifier based on the one or more augmented images of the one or more defects of the training specimen; receive from the characterization sub-system one or more target images of one or more target features of a target specimen; and determine one or more defects of the one or more target features with the machine learning classifier.