Wafer Defect Research Articles

Optimized Convolutional Neural Network Towards Effective Wafer Defects Classification

Optimized Convolutional Neural Network Towards Effective Wafer Defects Classification

Wafer Defect Identification with Optimal Hyper-Parameter Tuning of Support Vector Machine Using the Deep Feature of ResNet 101

As semiconductor processing technologies continue to advance, semiconductor wafers are becoming more densely packed and intricate, resulting in a higher incidence of surface imperfections. Therefore, it is crucial to detect these defects early and accurately classify them to pinpoint the root causes of the defects in the manufacturing process, ultimately leading to improved yield. Therefore, defect detection is critical in the industrial production of monocrystalline silicon. This study employs deep learning techniques to propose a framework for detecting defects on silicon wafers, focusing on optimizing the hyperparameters of support vector machines (SVM). Three methods were utilized to fine-tune the SVM parameters: Bayesian optimization, grid search, and random search techniques. This study demonstrates how selecting optimal values for SVM parameters can lead to better classification. Additionally, real manufacturing data were utilized to evaluate the performance of the proposed SVM classifier, with a comparison to state-of-the-art techniques in the field. By using deep features from ResNet 101 and a support vector machine, this work achieves 74.5% accuracy in identifying wafer defects without employing any optimization technique. However, the performance of the model was further improved by utilizing the random search optimization technique, which yielded the best result among the three optimization techniques tested, with an accuracy of 88.1%.

The performance of Adsorbing and Removing gaseous acetic acid by activate carbon

Currently, gaseous molecular contaminants (acetic acid) is one of the main factors that cause defects in semiconductors and wafers, reduce product yields in high-tech process plants. Propylene Glycol Monomethyl Ethyl Ether Acetate (PGMEA) is a versatile solvent with many applications. Solvents used in the solder mask process in the manufacture of printed circuit boards (PCBs). The diluent in photoresist and edge bead removers used in semiconductor and wafer fabrication is PGMEA, which forms acetate contaminants through an acid-catalyzed hydrolysis reaction. The concentration standard of acetic acid pollutants in clean rooms is 1ppb.In order to protect the expensive optical components, the acetic acid generated by PGMEA needs to be controlled.For this purpose, the activated carbon filter in the intake filter prevents acetic acid from contaminating the components and corroding the material. In this study, we used the commercially available activated carbon fabricated filter to adsorption of acetic acid. The effect of adsorption under different face velocity (0.03/0.06/0.09 m/s) and initial acetic acid concentrations (4/6/8/10 ppm) were measured according to the principle of ASHRAE 145.1standard. The experimental data are also evaluated for adsorption performance, adsorption capacity, kinetic model of acetic acid adsorption process, and isothermal model of acetic acid adsorption process.The results of the acetic acid adsorption experiment show that MM3000 activated carbon has the highest adsorption capacity and the longest adsorption saturation time than KPL activated carbon. The higher the inlet acetic acid concentration at a fixed wind speed, the shorter the saturation time and the higher the adsorption capacity; the faster the wind speed at a fixed acetic acid concentration, the shorter the saturation time and the higher the adsorption capacity.The Langmuir (R2) of the MM3000 sample is higher than the Freundlich’s coefficient of determination (R2), which indicates that the adsorption phenomenon of the MM3000 sample is more suitable to be described by the Langmuir isothermal adsorption mode, which is the adsorption of a single molecular layer on a homogeneous surface. The Freundlich R2 of the KPL sample is higher than the Langmuir’s coefficient of determination (R2), which means that the adsorption phenomenon of this KPL sample is more suitable to be described by the Freundlich isothermal adsorption mode, and it is a case of adsorption of multiple layers on an inhomogeneous surface. The adsorption of MM3000 activated carbon and KPL activated carbon is more suitable to be described by the proposed second-order adsorption kinetic model, which is a phenomenon of chemisorption.

Characterization of Multimodal Spot Scanning Imaging System for Wafer Defect Inspection

Characterization of Multimodal Spot Scanning Imaging System for Wafer Defect Inspection

Class Imbalance Wafer Defect Pattern Recognition Based on Shared-Database Decentralized Federated Learning Framework

Class Imbalance Wafer Defect Pattern Recognition Based on Shared-Database Decentralized Federated Learning Framework

Unsupervised Wafer Defect Classification Model Based On Joint Reconstruction And Clustering

Defect detection of unpatterned wafer is very important for determining the causes of wafer defects, and it is also a significant way to improve production yield. At present, the defect detection model based on the deep learning method has been widely used and has shown promising performance. However, the labor cost of supervised learning method based on labeled samples is very expensive, so the wafer defect detection model based on unsupervised learning is a future research direction. In this paper, we propose an unsupervised wafer defect classification model(UWDDM), in which the reconstruction loss is used to train the feature extractor of wafer defects, and the classifier is trained based on the cross entropy loss of clustering results and classification results. The experimental results on the real dataset MixedWM38 show that the proposed model has higher recognition accuracy and less inference time than other unsupervised models. This unsupervised learning method has great potential for wafer defect detection. It can avoid the expensive manual labeling cost and provide an effective way to automatically detect wafer defects.

Optical emission spectroscopy of fiber laser sustained Xe plasma

Laser-sustained plasma (LSP), a unique method of obtaining high brightness broadband light sources, has been used for rapid optical inspection of wafer defects and has thus attracted great interest. Then Xe plasma sustained under high pressure by a continuous fiber laser was investigated using optical emission spectroscopy (OES). Emission spectra of spatial integration diagnosis have shown that with the increase of laser power and the decrease of Gaussian beam focusing number F, the electron temperature fitted by Boltzmann plot method increased from 0.45 eV to 0.85 eV, the corresponding electron density was roughly an order of magnitude of 1016-1017 cm-3, and the calculated absorption coefficient of plasma was distributed in 0.9–1.5 cm-1. To further analyze the spatial distribution of plasma parameters, the spatial resolved OES of LSP was acquired, and the two-dimensional electron temperature and electron density were determined. The non-uniform distribution of the plasma parameters was analyzed due to the existence of thermal gravity convection and forced convection. The serious self-reversal of emission lines of 881.94 nm and 823.16 nm was observed and analyzed.

Wafer defect identification with optimal hyper-parameter tuning of support vector machine using the deep feature of ResNet 101

Wafer defect identification with optimal hyper-parameter tuning of support vector machine using the deep feature of ResNet 101

Formation and evolution of multiple-core structures in laser-sustained plasmas

We report on the formation of multiple-core structures in laser-sustained plasmas (LSPs) through the utilization of an improved two-dimensional laser-thermal-hydrodynamically coupled fluid model. Our findings reveal that steady-state LSPs exhibit various temperature structures, including single, double, and triple cores, as the input laser power increases incrementally. The spatiotemporal distributions of these multiple-core structures are demonstrated, and the presence of an LSP core is identified when the rate of enthalpy change becomes positive. This behavior is predominantly influenced by factors such as laser power absorption, radiation, and diffusion along the laser path. The obtained results provide explicit insights into the temperature structures within LSPs, thereby potentially contributing to the advancement of LSP-based light sources for optical wafer defect inspection in the semiconductor industry.

Mixed-type wafer defect detection based on multi-branch feature enhanced residual module

Wafer testing is crucial in semiconductor manufacturing. Accurate identification of wafer defects enables precise localization of manufacturing issues, thereby enhancing chip production yield. Mixed-type wafer defect patterns are complex, resulting from the combination of different single-type defects, exhibiting variations in shape, size, and distribution. Currently, most deep learning models for classifying mixed-type wafer defect patterns are complex and require further refinement to enhance their effectiveness. Therefore, this study proposes a deep learning model called MER-Net (Multi-branch Feature Enhanced Residual Network). By adjusting deformable convolution and Ghost convolution, a multi-branch feature enhanced residual module is crafted. This module serves as the foundation for a deep learning model designed to classify mixed-type wafer defects. To validate the effectiveness of the proposed model, experiments were conducted using a dataset comprising 38 types of mixed-type wafer defects. The experimental results reveal that this method achieves a recognition accuracy of 98.84% on the mixed-type wafer defect dataset. In comparison to existing deep learning models, this method demonstrates higher accuracy and superior performance. These findings further underscore the feasibility and effectiveness of the approach presented in this study.

A Copula network deconvolution-based direct correlation disentangling framework for explainable fault detection in semiconductor wafer fabrication

Wafer fabrication is a highly complex manufacturing system. Using complex network models to portray the correlation between parameters is an effective tool for finding the key influencing factors associated with the cause of wafer defects. However, the complex network in a wafer fabrication system has many nonlinearities, non-normal distributions, coupling, and spurious correlations. Quantitative clarification of direct and indirect correlations in the complex network is the prerequisite for achieving explainable fault detection in smart manufacturing. Therefore, a Copula network deconvolution-based direct correlation disentangling framework is proposed in the context of explainable fault detection in semiconductor wafer fabrication. Firstly, The complex network correlation diagram of the parameters in the wafer fabrication process is constructed with each parameter as a node and the correlation coefficients between parameters as connected edges. Then, a nonlinear correlation metric model based on adaptive Copula function selection is designed to deal with the nonlinear correlation of complex network models. The adaptive selection of the Copula function is realized by the goodness-of-fit test method based on Euclidean distance. After that, considering the coupling and spurious correlation in the complex network model, a network deconvolution-based fault detection method is designed to identify direct correlation. Finally, a case study using the actual data from semiconductor wafer fabrication systems is conducted to compare the existing mainstream fault detection methods and validate the effectiveness of the proposed method.

A Deep Learning Model for Wafer Defect Map Classification: Perspective on Classification Performance and Computational Volume

Wafer defect map images are generated by performing electrical tests on each chip on wafer. These images demonstrate specific failure patterns occurred from semiconductor manufacturing process. It is crucial for engineers to classify what kind of defect patterns on this wafer is early. In an attempt to automate the classification of wafer defect maps, which are currently manual dependent, various machine learning and deep learning models have been introduced. However, it has not been successfully applied to real‐world mass production environments due to problems such as classification performance and computational volume. Therefore, the deep learning model integrating the Inception module and the skip connection module for wafer defect map classification is proposed here. This model has a fast training and inference speed with a small number of parameters, it is highly practical when processing large amounts of real‐time test data in a semiconductor manufacturing environment because of its small computational volume and high classification performance. This method is applied on the real‐field wafer test data, and the result shows that the proposed model takes a significant improvement on inference time by over 59% with high performance compared to the baseline model.

Application of 1D ResNet for Multivariate Fault Detection on Semiconductor Manufacturing Equipment.

Amid the ongoing emphasis on reducing manufacturing costs and enhancing productivity, one of the crucial objectives when manufacturing is to maintain process tools in optimal operating conditions. With advancements in sensing technologies, large amounts of data are collected during manufacturing processes, and the challenge today is to utilize these massive data efficiently. Some of these data are used for fault detection and classification (FDC) to evaluate the general condition of production machinery. The distinctive characteristics of semiconductor manufacturing, such as interdependent parameters, fluctuating behaviors over time, and frequently changing operating conditions, pose a major challenge in identifying defective wafers during the manufacturing process. To address this challenge, a multivariate fault detection method based on a 1D ResNet algorithm is introduced in this study. The aim is to identify anomalous wafers by analyzing the raw time-series data collected from multiple sensors throughout the semiconductor manufacturing process. To achieve this objective, a set of features is chosen from specified tools in the process chain to characterize the status of the wafers. Tests on the available data confirm that the gradient vanishing problem faced by very deep networks starts to occur with the plain 1D Convolutional Neural Network (CNN)-based method when the size of the network is deeper than 11 layers. To address this, a 1D Residual Network (ResNet)-based method is used. The experimental results show that the proposed method works more effectively and accurately compared to techniques using a plain 1D CNN and can thus be used for detecting abnormal wafers in the semiconductor manufacturing industry.

Multi-detector simultaneous sampling design in the spot-scanning defects detection system

With the decreases in semiconductor critical dimension and increases in circuit intricacy, the impact of wafer defects on the yield and manufacturing cost of integrated circuits is becoming increasingly significant. In this paper, a multimodal spot-scanning defect detection system designed for unpatterned wafer surfaces is introduced. The defect detection system's mechanics, electronics, and software are introduced, focusing on presenting a multi-detector simultaneous sampling design. The signal sampling is based on the position triggering of the rotary stage encoder. The focused illumination spot is scanned in a helical trajectory relative to the wafer in coordination with the motion of the linear stage and rotary stage. The signal detection is carried out synchronously, and the signal sampling is triggered by rotary stage encoder position pulses to achieve equally spaced sampling. The detector signals are trans-impedance amplified and transmitted differentially to transfer the analogue signals from the optical head to the main control board. Under FPGA control, the main control board completes the simultaneous sampling and signal processing through nine high-precision 16-bit SAR(successive-approximation-register) ADCs(analogue-to-digital converter) under the trigger signal. The defect detection system is designed with a pixel resolution of 4 × 4 μm, requiring a sampling position error of less than 1 pixel and an inter-channel position deviation of less than 0.1 pixel. The rotary stage prototype has a design speed of 120 rpm, and the positioning accuracy of the encoder is ±8 arcsec. Therefore, the maximum sampling position error is 1.94 μm for a 4-inch wafer sample. The maximum phase deviation between the nine channels is 32.5 ns, and the maximum position deviation is 20.4 nm.

A novel approach for wafer defect pattern classification based on topological data analysis

A novel approach for wafer defect pattern classification based on topological data analysis

Wafer Defect Identification with Optimal Hyper-Parameter Tuning of Support Vector Machine using the Deep Feature of ResNet 101

As semiconductor processing technologies continue to advance, semiconductor wafers are becoming more densely packed and intricate, resulting in a higher incidence of surface imperfections. Therefore, it is crucial to detect these defects early and accurately classify them to pinpoint the root causes of the defects in the manufacturing process, ultimately leading to improved yield. Therefore, defect detection is critical in the industrial production of monocrystalline silicon. This study employs deep learning techniques to propose a framework for detecting defects on silicon wafers, focusing on optimizing the hyperparameters of support vector machines (SVM). Three methods were utilized to fine-tune the SVM parameters: Bayesian optimization, grid search, and random search techniques. This study demonstrates how selecting optimal values for SVM parameters can lead to better classification. Additionally, real manufacturing data were utilized to evaluate the performance of the proposed SVM classifier, with a comparison to state-of-the-art techniques in the field. By using deep features from ResNet 101 and a support vector machine, this work achieves 74.5% accuracy in identifying wafer defects without employing any optimization technique. However, the performance of the model was further improved by utilizing the random search optimization technique, which yielded the best result among the three optimization techniques tested, with an accuracy of 88.1%.

Rectangular Amplitude Mask-Based Auto-Focus Method with a Large Range and High Precision for a Micro-LED Wafer Defects Detection System.

Auto-focus technology plays an important role in the Micro-LED wafer defects detection system. How to accurately measure the defocus amount and the defocus direction of the Micro-LED wafer sample in a large linear range is one of the keys to realizing wafer defects detection. In this paper, a large range and high-precision auto-focus method based on a rectangular amplitude mask is proposed. A rectangular amplitude mask without a long edge is used to modulate the shape of the incident laser beams so that the spot shape distribution of the reflected laser beam on the sensor changes with the defocus amount of the wafer sample. By calculating the shape of the light spots, the defocus amount and the defocus direction can be obtained at the same time. The experimental results show that under the 20× microscopy objective, the linear range of the auto-focus system is 480 μm and the accuracy can reach 1 μm. It can be seen that the automatic focusing method proposed in this paper has the advantages of large linear range, high accuracy, and compact structure, which can meet the requirements of the Micro-LED wafer defects detection equipment.

Application of Hydrosol to Aerosol Based Metrology to Predict Wafer Defects from Process Chemical Contamination

A method for continuous measurement of particle and particle precursor material in high purity liquids using aerosolization and condensation particle courting (A+CPC) is presented. A process for correlating wafter defects with the concentration values measured using A+CPC using Sequential Spin Coating + Surface Enhanced Particle Sizing (SSC + SEPS) is described. The method was applied to monitor the rinsate of an ion exchange resin following the SEMI C93 guide. The results showed correlation between the methods and the concentration values were used to calculate a deposition factor (liquid concentration/surface concentration) on the order of 1E5 (#/ml)/(#/cm2).

Efficient mixed-type wafer defect pattern recognition using a novel joint attention skipped graph convolutional neural network

For wafer foundries to maintain high yield and quality, errors in manufacturing wafers must be identified and corrected. To find the source of problems and increase overall yield, defect pattern recognition (DPR) is essential. Nevertheless, mixed-type DPR, which includes identifying numerous types of flaws at once, poses more difficulties because of the variety of spatial features, the unpredictability of flaw patterns, and the fluctuating amount of flaws present. So, a novel joint attention skipped graph convolutional neural network (JAS-GCNN) technique is suggested in this paper to analyse the difficulty of mixed-type DPR. The JAS- GCNN effectively captures locally as well as globally interconnections among faults on a wafer by fusing the strength of GCNN along with focused processes. The JAS-GCNN facilitates the learning of complicated fault patterns by allowing information to propagate across multiple layers by utilising skip connections. We do trials on the MixedWM38 WaferMap dataset, which contains a variety of defect types, to assess the proposed technique. The outcomes show that when compared to cutting-edge techniques, the JAS-GCNN performs better. It successfully detects various faults and performs defect classification jobs with high accuracy.

Structural Design and Simulation of a Multi-Channel and Dual Working Condition Wafer Defect Inspection Prototype.

Detecting and classifying defects on unpatterned wafers is a key part of wafer front-end inspection. Defect inspection schemes vary depending on the type and location of the defects. In this paper, the structure of the prototype is designed to meet the requirements of wafer surface and edge defect inspection. This prototype has four inspection channels: scattering, reflection, phase, and contour, with two working conditions: surface and edge inspection. The key structure of the prototype was simulated using Ansys. The simulation results show that the maximum deformation of the optical detection subsystem is 19.5 μm and the fundamental frequency of the prototype is 96.9 Hz; thus, these results meet the requirements of optical performance stability and structural design. The experimental results show that the prototype meets the requirements of the inspection sensitivity better than 200 nm equivalent PSL spherical defects.