Optical Simulation for Pseudo-Random Defect Dataset Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The semiconductor industry faces challenges in obtaining sufficient valid datasets for defect detection algorithms due to the diversity of semiconductor and defect characteristics, making it difficult to efficiently support the development and innovation of machine learning algorithms.
Innovation Solution
A method is developed for image simulation and pseudo-random defect dataset generation, which includes constructing a simulated three-dimensional model and a Kohler illumination model, performing optical simulations, and generating pseudo-random defective images using graphical processing techniques.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If real defect datasets are collected from semiconductor manufacturing processes, then the dataset reflects actual defect characteristics, but the data acquisition is expensive, time-consuming, and limited in quantity and diversity
Solution Approach 1:
The patent creates virtual copies of real defect images through simulation. The simulation system generates synthetic defect images that replicate the appearance and characteristics of real defects, allowing unlimited generation without the cost and time constraints of collecting actual defect data from manufacturing processes.
Solution Approach 2:
The patent replaces the physical/mechanical process of collecting real defect images from semiconductor manufacturing with a computational simulation system. This substitution eliminates the need for physical wafer inspection and manual data collection, enabling rapid generation of diverse defect datasets through software-based optical simulation.
2Adaptability or versatility
If multiple defect array samples are prepared to increase dataset diversity, then more defect types can be studied, but the cost increases and the data remains limited
Solution Approach 1:
The simulation system allows dynamic adjustment of defect parameters such as size, shape, position, and optical properties without physical manufacturing. Users can modify these parameters to generate an unlimited variety of defect types and configurations, eliminating the need to prepare multiple physical defect arrays and significantly reducing preparation costs.
3Measurement precision
If computational simulation is performed with high sampling density to improve image accuracy, then simulation precision increases, but computational cost and time increase
Solution Approach 1:
The patent implements an iterative simulation process with convergence criteria. The system performs simulations at different sampling densities and uses feedback from image quality assessment to determine when sufficient accuracy has been achieved, allowing the system to stop computing once the desired precision is reached, thereby optimizing the balance between accuracy and computational cost.
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 method enables the rapid generation of patterned wafer defective image datasets, allowing for the training of defect inspection models and achieving direct detection of micro and nano defects in patterned wafers.
Implementation Method 1
performing an optical simulation based on the simulated three-dimensional model and the simulated Kohler illumination model, the optical simulation including simulating by a simulation environment constructed by combining the simulated three-dimensional model and the simulated Kohler illumination model and synthesizing an image by obtaining far-field electromagnetic field distribution data to synthesize an image
Data Source
AI summary
The embodiments of the present disclosure provide a method for an image simulation generation based on a near-field simulation of a computational electromagnetic field, the method comprising: constructing a simulated three-dimensional model based on model parameters; constructing a simulated Kohler illumination model based on a light source parameter; using a degree of similarity change in a synthesized image under incremental aperture diaphragm sampling points as a criterion for approximate convergence of the simulation to determine a count of samples to be used for a balancing combination of simulation cost and accuracy; an optical simulation is performed based on the simulated three-dimensional model and the simulated Kohler illumination model, and a far-field electromagnetic field distribution data is obtained to obtain a simulated image by synthesizing the image. In the generation of a large count of simulated images on the basis of pseudo-random defect dataset generation may be further realized, in the acquisition of a large count of datasets, the dataset of defect inspection model may be trained, in order to achieve a direct detection for patterned wafer defective images and solve the problem of difficult access to reference images in a process of patterned wafer defect detection.