Wafer production method and system based on image segmentation

Abstract

The application provides a wafer production method and system based on image segmentation, and relates to the technical field of wafer production.The application constructs a multi-region space-time thermal response matrix through image segmentation, and automatically determines an optimal temperature disturbance quantity in combination with a thermal diffusion equation to realize accurate marking of a thermal diffusion abnormal region; subsequently, multi-wavelength measurement light is applied to the marked region and the light intensity is dynamically corrected to enhance the recognition sensitivity of defects at different depths and shapes; micro-vibration and light response data are introduced, a pre-trained defect prediction model is used to extract multi-modal features, and preliminary defect indicators are generated; finally, the polarization degree and phase delay of the polarized structured light are multiplied to correct, which can not only amplify the real defect signal, but also suppress noise interference, significantly improve the detection accuracy, distinguishability and robustness, and improve the production quality of wafers.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor defect detection, in particular to a wafer production method and system based on image segmentation. BACKGROUND

[0002] Image segmentation algorithms gradually penetrate into the field of thermal image processing, and through semantic segmentation or instance segmentation technology, thermal images are divided into several measurement units, providing fine spatial division for subsequent thermal diffusion analysis. At the same time, the application of thermal diffusion equation in defect detection makes it possible to quantitatively analyze temperature gradient and heat flux density. However, the above methods often rely on single modal data input, and lack dynamic adaptive mechanism in preset temperature disturbance and optical measurement light intensity selection, making it difficult to balance detection sensitivity and noise resistance. In addition, although micro-vibration measurement and polarized structured light application have been tried by a few studies, there is still no systematic detection process based on artificial intelligence and multi-modal fusion.

[0003] In the prior art, a wafer defect detection method with publication number CN112561849B can use automatic image recognition technology to obtain wafer mask images, reducing the consumption of manpower and time. The wafer defect detection method comprises the following steps: obtaining a wafer standard sample image; performing automatic segmentation processing on the wafer standard sample image to obtain a wafer mask image; obtaining a wafer image to be detected, and detecting wafer defects by comparing the wafer mask image and the wafer image to be detected.

[0004] The prior art has many deficiencies in wafer surface defect production detection:

[0005] Firstly, relying only on static thermal images or single-wavelength optical imaging makes it difficult to distinguish thermal diffusion abnormalities caused by material internal micro-cracks, uneven doping or stress and strain, resulting in high false positive and false negative rates; secondly, the parameter calibration of the thermal diffusion equation is mostly based on empirical calculation, and lacks optimal temperature disturbance selection for different defect types, making it impossible to maximize the signal-to-noise ratio of the detection process;

[0006] Thirdly, single modal light intensity setting and vibration response data acquisition cannot meet the dynamic needs of the defect prediction model, making it difficult to correct the optical measurement light intensity in real time;

[0007] Finally, although polarized structured light can enhance the contrast of defects in surface topography measurement, existing solutions usually fail to associate and correct parameters such as polarization degree and phase delay with preliminary defect measurement indicators, resulting in limited detection accuracy.

[0008] The above reasons have a great impact on the quality of wafer production, so a wafer production method and system based on image segmentation are needed to solve the above technical problems. SUMMARY

[0009] The present application aims to provide a wafer production method and system based on image segmentation to solve the problems raised in the background art.

[0010] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0011] The wafer production method based on image segmentation comprises the following specific steps:

[0012] Step S1: Collecting thermal image data of the wafer surface to be detected, and dividing the thermal image data into a plurality of fixed-size measurement regions based on an image segmentation algorithm; during the monitoring period, collecting thermal response data of each measurement region under the action of a preset temperature change amount to form a multi-region space-time thermal response matrix;

[0013] Step S2: Analyzing the thermal response data of each measurement region during the monitoring period using a heat diffusion equation to determine the temperature change and mark the heat diffusion abnormal region, and determine the optimal preset temperature change amount;

[0014] Step S3: Applying a multi-wavelength measurement light to the marked heat diffusion abnormal region in turn, and performing multiple linear analysis on the thermal response data of the heat diffusion abnormal region and the optimal preset temperature change amount to dynamically adjust the reference light intensity of the multi-wavelength measurement light, and obtaining the adjusted light intensity;

[0015] Step S4: Applying micro-vibration to each heat diffusion abnormal region and collecting corresponding vibration response data, and obtaining light response data under each heat diffusion abnormal region;

[0016] The vibration response data, thermal response data and light response data are input into a pre-trained defect prediction model, the defect prediction model performs automatic feature extraction and correlation analysis on the labeled data in the historical defect database to generate a preliminary defect measurement index reflecting the severity of the region defect;

[0017] Step S5: Based on the preliminary defect measurement index of each heat diffusion abnormal region, applying different combinations of polarized structured light to each heat diffusion abnormal region, and correcting the preliminary defect measurement index according to the degree of polarization and phase delay.

[0018] A wafer production system based on image segmentation, the system is used to execute the wafer production method based on image segmentation, comprising:

[0019] The heat response data generation module is configured to collect thermal image data of a wafer surface to be detected, and divide the thermal image data into a plurality of measurement regions of a fixed size based on an image segmentation algorithm; during a monitoring period, collect heat response data of each measurement region under a preset temperature variation to form a multi-region space-time heat response matrix;

[0020] The anomaly marking module is configured to analyze the heat response data of each measurement region during the monitoring period by using a heat diffusion equation, to determine temperature variation and mark heat diffusion abnormal regions, and to determine an optimal preset temperature variation;

[0021] The light intensity adjustment module is configured to sequentially apply multi-wavelength measurement light to the marked heat diffusion abnormal regions, and perform multi-element linear analysis on the heat response data of the heat diffusion abnormal regions and the optimal preset temperature variation to dynamically adjust the reference light intensity of the multi-wavelength measurement light, to obtain an adjusted light intensity;

[0022] The initial index generation module is configured to apply micro-vibration to each heat diffusion abnormal region, and collect corresponding vibration response data and light response data under each heat diffusion abnormal region;

[0023] The vibration response data, the heat response data and the light response data are input into a pre-trained defect prediction model, the defect prediction model performs automatic feature extraction and correlation analysis on labeled data in a historical defect database to generate a preliminary defect measurement index reflecting the severity of the region defects;

[0024] The correction module is configured to apply different combinations of polarized structured light to each heat diffusion abnormal region based on the preliminary defect measurement index of each heat diffusion abnormal region, and correct the preliminary defect measurement index according to the degree of polarization and phase delay.

[0025] Advantages: Compared with the prior art, the advantages of the present application are: a multi-region space-time heat response matrix is constructed by image segmentation, and the optimal temperature disturbance is automatically determined by combining the heat diffusion equation to realize accurate marking of the heat diffusion abnormal regions; subsequently, multi-wavelength measurement light is applied to the marked regions and the light intensity is dynamically corrected to enhance the recognition sensitivity of defects of different depths and topographies; micro-vibration and light response data are introduced, multi-modal features are extracted by using a pre-trained defect prediction model to generate a preliminary defect index; finally, the degree of polarization and phase delay of the polarized structured light are multiplied to correct, which can not only amplify the real defect signal, but also suppress noise interference, significantly improve the detection accuracy, discrimination and robustness, and improve the production quality of the wafer. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 The present application is a whole method flowchart;

[0027] Figure 2This is a block diagram of the overall system modules of the present invention. Detailed Implementation

[0028] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0029] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0030] Example 1:

[0031] Please see Figure 1 This invention provides a wafer fabrication method based on image segmentation, the specific steps of which include:

[0032] Step S1: Collect thermal image data of the wafer surface to be inspected, and divide the thermal image data into multiple fixed-size measurement areas based on the image segmentation algorithm. During the monitoring period, collect thermal response data of each measurement area under the action of a preset temperature change to form a multi-region spatiotemporal thermal response matrix.

[0033] Further explanation: A laser scanner with a resolution of 0.1 mm was used to perform a three-dimensional topographic scan of the wafer surface to be inspected, obtaining detailed three-dimensional coordinate data for each wafer region.

[0034] Based on the 3D topography scanning results, a threshold segmentation algorithm was used to divide the wafer surface corresponding to the thermal image data into 1mm segments. 2 The grid is divided into several measurement areas to ensure that each measurement area is the same size, thereby ensuring the balance of the data.

[0035] The center coordinates of each measurement area are marked as follows: Where i∈{1,2,…,N}, i is the index of the measurement area, N is the total number of measurement areas, the reference temperature of the wafer surface to be tested is set to 25°C, and the preset temperature change is set to... That is, the temperature variation range of each measurement area is ;set up The value is selected from the range [0.1, 10]; this embodiment initially sets... .

[0036] The surface of the wafer under test was sequentially scanned using an infrared thermal imager at a reference temperature of 25°C, and temperature time-series data for each measurement area i were acquired during the monitoring period. Characterized thermal response data, in which To measure the temperature value of region i at time t; these temperature time series data... By integrating the data, a multi-regional spatiotemporal thermal response matrix is ​​formed.

[0037] It should be noted that infrared thermal imagers achieve optimal measurement accuracy and response speed at 25°C.

[0038] The infrared thermal imager in this embodiment has a resolution of 0.05°C and a response time of 50ms, enabling it to accurately capture temperature change information in each measurement area, ensuring the timeliness and accuracy of the data. The center coordinates of each measurement area and the thermal response data are uniformly stored in a database, forming a correlation between the center coordinates and the temperature time-series data; the coordinate table stored in the database is as follows. .

[0039] Step S2: Analyze the thermal response data of each measurement area during the monitoring period using the thermal diffusion equation to determine the temperature change, mark the abnormal thermal diffusion areas, and determine the optimal preset temperature change.

[0040] Further explanation: Regarding the preset temperature change... Multi-point simulation measurements were performed within a preset range [0.1, 10] to determine the optimal preset temperature change that maximizes the global temperature change. The specific logic includes:

[0041] Define the preset temperature change amount The candidate set is ; This is the temperature adjustment factor;

[0042] make A value is taken every 0.5°C within the range [0.1, 10] to balance measurement accuracy and computational efficiency, for each candidate preset temperature change. Perform multi-point simulations as follows, setting the upper and lower temperature limits of the wafer surface to... ,right Middle measurement area Perform a pre-scan to acquire the corresponding temperature time series data. In this embodiment, m=10, indicating that 10 measurement areas are randomly selected, and the measurement is performed under a given preset temperature change. Next, calculate the measurement area. The maximum temperature variation range during the monitoring period is calculated based on the difference between the highest and lowest temperature values ​​during the monitoring period.

[0043] Next, the average change amplitude corresponding to the candidate preset temperature change amounts is calculated. The average change amplitude is the average of the maximum temperature change range of 10 measurement areas. Among all candidate preset temperature change amounts, the one with the largest average change amplitude is selected as the optimal preset temperature change amount. , select Write the data into the "Optimal Temperature Parameter Table" in the database for use in subsequent steps.

[0044] Further explanation: The optimal preset temperature change... Under these conditions, the residuals are calculated using finite difference discretization based on the two-dimensional thermal diffusion equation, and measurement areas exceeding a first preset threshold are marked as thermal diffusion anomaly regions; the specific logic includes:

[0045] The two-dimensional thermal diffusion equation is selected as follows:

[0046]

[0047] in, To measure the temperature of region i at time t; The thermal diffusivity of the wafer to be tested; in this embodiment, a calibrated value is used. Grid spacing Time step The response time of the infrared thermal imager is consistent with that in step S1, and the temperature time series data of the measurement area i is... Calculate the approximate time derivative , For measuring region i in The range of temperature changes over time Corresponding to the thermal imager response time, the Laplace operator in the discretized space is calculated for the temperature values ​​of the measurement area i and its four neighboring regions. The Plas operator reflects the magnitude of the temperature difference between the measurement region i and its surrounding measurement regions.

[0048] In this embodiment, when calculating points on the boundary of the wafer to be detected, if there are no adjacent points in a certain direction, the Neumann boundary condition needs to be applied: setting the temperature derivative in that direction to zero, or the Dirichlet boundary condition: directly specifying the temperature in that direction as a preset constant.

[0049] For each measurement region i, calculate the residual of the discrete form of the heat diffusion equation. ,when An increase indicates a deviation between the actual temperature change and the ideal diffusion model. The mean residual value of all N measurement regions is calculated. with standard deviation Set an anomaly determination coefficient k, when At that time, the measurement area i is marked as a region of abnormal thermal diffusion. In this embodiment, k=3; It is the first preset threshold.

[0050] For all measurement areas marked as thermal diffusion anomalies Number them and their corresponding residual values Stored in the "Abnormal Area Table" of the database. This "Abnormal Area Table" corresponds to the center coordinates of the measured area in the database. Correlation. It should be noted that the anomaly determination coefficient k is determined based on the "Three Sigma Rule" in statistics and a trade-off between actual detection sensitivity and false alarm rate. Specifically: Three Sigma Rule: Residual... It approximately follows a normal distribution; according to the properties of the normal distribution: 68.3% of the data falls within... Within; 95.5% of the data falls within Within; 99.7% of the data falls within Therefore, taking k=3 means that only when the residual of the measurement area exceeds the normal fluctuation range of 99.7% will it be marked as abnormal, thus effectively reducing the "false anomalies" caused by measurement noise or small fluctuations to below 0.3%.

[0051] Step S3: Apply multi-wavelength measurement light to the marked thermal diffusion anomaly areas in sequence, and perform multivariate linear analysis on the thermal response data of the thermal diffusion anomaly areas and the optimal preset temperature change to dynamically adjust the reference light intensity of the multi-wavelength measurement light and obtain the adjusted light intensity.

[0052] Further explanation: For each region with an abnormal thermal diffusion Two adjustable light sources of different wavelengths are applied sequentially to dynamically adjust the reference light intensity of the multi-wavelength measurement light, resulting in the adjusted light intensity. The specific logic includes:

[0053] Two light sources were selected: a 1.5µm laser and an 850nm LED; to cover the response differences of different depths and surface features, and to balance measurement sensitivity and depth information acquisition.

[0054] The reference light intensity for both wavelength light sources was set to be... ;

[0055] Center wavelength 1.5µm, beam diameter ≤0.5mm, reference light intensity ;

[0056] LED module: center wavelength 850nm, divergence angle 30°, reference luminous intensity is the same. .

[0057] It should be noted that a wavelength of 1.5µm can penetrate the surface material of a wafer to obtain deeper information. This wavelength of laser is beneficial for detecting deeper defects or non-surface problems in materials, thus responding to defects or deformations deep within the wafer. 850nm LEDs belong to the near-infrared (NIR) spectrum, which is sensitive to surface features such as surface defects, microcracks, contamination, and oxide layers. The 850nm wavelength is suitable for detecting surface defects, especially in semiconductor manufacturing, such as microcracks, corrosion, and dirt.

[0058] Based on reference light intensity Introducing anomaly regions of thermal diffusion Temperature variation range and the optimal preset temperature change Multivariate linear analysis was performed to identify the regions of thermal diffusion anomalies. Adjusted light intensity Adjusted light intensity The formula is as follows:

[0059]

[0060] in, Region with abnormal heat diffusion Adjusted light intensity; Region with abnormal heat diffusion The range of temperature change; The optimal preset temperature change k1 is the gain coefficient; in this embodiment, k1=3.

[0061] when When it increases, Linear increase to enhance response to high-temperature spots; when When decreasing, The exposure level should be reduced accordingly to prevent excessive irradiation in each area with abnormal heat diffusion. Two light sources were used for irradiation, in a fixed order: 1.5µm laser → 850nm LED, for a duration of [duration missing]. During the irradiation process, the sampling rate of the light response signal is... Record the area of ​​thermal diffusion anomaly Light intensity time series data This embodiment ; .

[0062] The control system drives the light source and high-speed photodetector to be triggered synchronously, and collects light intensity time-series data. And it is stored in real time. Database table structure: The "Light Response Table" field includes: Index of thermal diffusion anomaly areas. Light source type (1.5µm / 850nm), adjusted light intensity Light intensity time series data irradiation duration Sampling rate .

[0063] Indexing by thermal diffusion anomaly regions Associated with the "Abnormal Area Table" and "Coordinate Table", this provides input data for the multiple regression analysis in step S4.

[0064] Step S4: Apply micro-vibration to each thermal diffusion anomalous region and collect the corresponding vibration response data, while simultaneously acquiring the optical response data under each thermal diffusion anomalous region.

[0065] Vibration response data, thermal response data, and optical response data are input into a pre-trained defect prediction model. The defect prediction model uses labeled data from a historical defect database to automatically extract features and perform correlation analysis, generating preliminary defect measurement indicators that reflect the severity of defects in the area.

[0066] Further explanation: The defect prediction model uses deep neural networks or machine learning algorithms to identify areas with abnormal heat diffusion. The preliminary defect measurement indicators are marked as The piezoelectric ceramic micro-vibration stage was placed in a region of abnormal thermal diffusion. Above, to collect raw vibration signals Characterized vibration response data; vibration amplitude a = 0.5g, vibration frequency... Vibration duration The original vibration signal in this embodiment Using a resolution of 0.001g, sampling rate The accelerometer; the vibration amplitude a is limited to the range of 0.1g-1.0g.

[0067] For the original vibration signal Perform a 4th-order Butterworth bandpass filter to obtain the filtered signal. In this embodiment, the bandwidth of the fourth-order Butterworth bandpass filter is 50Hz–5,000Hz, based on the original vibration signal. Calculate the thermal diffusion anomaly regions respectively. Root mean square characteristic and peak characteristics The specific calculation formula is existing technology and will not be elaborated here.

[0068] Obtaining the thermal diffusion anomaly region Light intensity time series data Based on light intensity time series data Calculate the thermal diffusion anomaly region Average light response and standard deviation of light response , the temperature variation amplitude in the thermal response data of the thermal diffusion abnormal region , the root mean square value feature , the peak value feature , the average light response , and the light response standard deviation , and the temperature variation amplitude are taken as input features of the defect prediction model, preliminary defect measurement indicators are taken as outputs of the defect prediction model.

[0069] The architecture of the defect prediction model is set as follows in this embodiment: the input layer has 5 nodes, and the hidden layers are as follows in sequence: hidden layer 1: 128 nodes, ReLU activation; hidden layer 2: 64 nodes, ReLU activation; hidden layer 3: 32 nodes, ReLU activation; and output layer: 1 node, linear output. .

[0070] The normalized input vector of the thermal diffusion abnormal region is input into the trained network to obtain the preliminary defect measurement indicators of the thermal diffusion abnormal region .

[0071] According to the label distribution of the historical training set, the mean and the standard deviation of the training set corresponding to the defect prediction model are calculated, and a second preset threshold is set. , The calculation method is , if , the thermal diffusion abnormal region is marked as “suspected defect” in the “preliminary defect table”, and the output value and the corresponding are recorded at the same time.

[0072] Step S5: Based on the preliminary defect measurement indicators of each thermal diffusion abnormal region, different combinations of polarized structured light are applied to each thermal diffusion abnormal region, and the preliminary defect measurement indicators are corrected according to the polarization degree and phase delay.

[0073] Further explanation: In this embodiment, a digital micromirror device (DMD) is used to generate structured light with a determined spatial frequency and sequentially apply four linear polarization states to obtain reflection intensity for polarization feature calculation.

[0074] The different combinations of polarized structured light include four linear polarization states, which are 0°, 45°, 90°, and 135° in sequence.

[0075] ​The four polarization states of this embodiment cover two sets of orthogonal and 45° angle, which can completely extract Stokes parameters. The four polarization states are switched in sequence, and the corresponding reflected light intensity is collected 、 、 、 .

[0076] The DMD (model: Texas Instruments DLP4500) of this embodiment generates sinusoidal structured light with a spatial frequency of 0.5 cycles / mm to balance the clarity of the stripes and the phase sensitivity. An electro-optic modulator (EOM) is used to output 0° / 45° / 90° / 135° linear polarization in sequence; the model of the electro-optic modulator used in this embodiment is Thorlabs EO-AM-NR-C4; the specific operation steps are as follows:

[0077] Fix the wafer to be detected on the sample stage, and align the light path center Set the EOM polarization angle to 0° in sequence, start the DMD to project structured light, and trigger the camera (Basler acA2040-90um, exposure time 50µs) to collect the reflected intensity at the same time, record as Switch the polarization angle to 45°, 90°, and 135° in sequence, repeat the above collection, and record 、 、 respectively. If the polarization switching angle accuracy changes by ±1°, the error of the corresponding collected intensity value does not exceed 2%. Based on the collected reflected light intensity 、 、 、 , the polarization degree and the phase delay of the thermal diffusion abnormal area are calculated and the phase delay are Z-score normalized according to the training set statistics, and a multiplication correction model is constructed according to the normalization result to output the final defect measurement index of each thermal diffusion abnormal area .

[0078] The polarization degree and the phase delay of all samples in the historical defect database are taken out, the mean and standard deviation of the polarization degree, and the mean and standard deviation of the phase delay are calculated. After the polarization degree and the phase delay are normalized, the relative correction coefficient is constructed by introducing the weight determined in the training stage, and the final defect index is generated by multiplying the preliminary index.

[0079] The multiplication correction model is constructed as follows:

[0080]

[0081] wherein , are the weight proportions of the corresponding parameters; and are the normalized values of the polarization degree and the phase delay , , are determined by cross-validation, and the values of the present embodiment are 0.3 and 0.2 respectively, when , , if the value of is -0.5, the contribution of this term can suppress the false increase caused by noise. Again, the second preset threshold is used, if , the thermal diffusion abnormal area is labeled as “key re-inspection”; otherwise, it is labeled as “routine inspection”.

[0082] It should be noted that: the present embodiment uses a digital micromirror device (DMD) to generate a structured light with a determined spatial frequency and sequentially applies four kinds of linear polarization states to obtain the reflection intensity for polarization feature calculation;

[0083] “Key re-inspection” represents high-risk areas: these areas show higher defect severity and have a higher probability of existing potential defects such as scratches, holes or cracks.

[0084] Further in-depth detection and analysis are required, such as high-resolution imaging, microscopic examination or other advanced detection techniques, to confirm the specific type and severity of the defects.

[0085] More detection resources and manpower need to be allocated to these high-risk areas in priority to ensure timely discovery and handling of potential defects and prevent defects from expanding or affecting product quality.

[0086] Low-risk areas are labeled as “routine inspection” areas, which have lower defect risk, and the preliminary defect measurement indicators indicate that the defect degree is within an acceptable range or the probability of defects is small;

[0087] Subsequent routine monitoring and random sampling are required to ensure that these areas remain in a normal state and do not need to be immediately subjected to in-depth inspection. Reducing the detection frequency and resource input for these areas can optimize the overall detection process and improve detection efficiency.

[0088] Classifying the detection areas into "key re-inspection" and "regular inspection" has the following effects: improving detection efficiency: concentrating resources and efforts on high-risk areas to ensure the quality and reliability of critical parts.

[0089] Optimizing resource allocation: avoiding excessive investment in low-risk areas, reducing overall detection costs;

[0090] Improving product quality: ensuring timely discovery and processing of critical defects, reducing product defect rates, and improving the market competitiveness of the final product.

[0091] Please refer to Figure 2 A wafer production system based on image segmentation, said system is used to perform the wafer production method based on image segmentation, comprising:

[0092] A thermal response data generation module: for collecting thermal image data of the wafer surface to be detected, and dividing the thermal image data into a plurality of fixed-size measurement areas based on an image segmentation algorithm. During the monitoring period, collect the thermal response data of each measurement area under the action of the preset temperature change to form a multi-region spatiotemporal thermal response matrix;

[0093] An anomaly marking module: for analyzing the thermal response data of each measurement area during the monitoring period using a heat diffusion equation to determine the temperature change and mark the heat diffusion abnormal area, while determining the optimal preset temperature change;

[0094] An optical intensity adjustment module: for sequentially applying multi-wavelength measurement light to the marked heat diffusion abnormal area, introducing the thermal response data of the heat diffusion abnormal area and the optimal preset temperature change for multiple linear analysis to dynamically adjust the reference light intensity of the multi-wavelength measurement light, and obtaining the adjusted light intensity;

[0095] An initial index generation module: for applying micro-vibration to each heat diffusion abnormal area and collecting the corresponding vibration response data, while obtaining the light response data under each heat diffusion abnormal area;

[0096] Input the vibration response data, thermal response data and light response data into the pre-trained defect prediction model. The defect prediction model uses the labeled data in the historical defect database for automatic feature extraction and correlation analysis to generate preliminary defect measurement indexes reflecting the severity of the area defects;

[0097] A correction module: for applying different combinations of polarized structured light to each heat diffusion abnormal area based on the preliminary defect measurement index of each heat diffusion abnormal area, and correcting the preliminary defect measurement index according to the polarization degree and phase delay.

[0098] It should be noted that all the calculation formulas in the present application file use regression analysis including but not limited to machine learning algorithms to deeply analyze the collected relevant parameters, identify their natural trends and mutual relationships. Professional software such as Python's Scikit-learn library or R language is used to automatically generate mathematical models matching the data. Then, the performance of the model is objectively evaluated by methods such as cross-validation, and combined with continuous feedback and optimization to ensure that the created formula truly reflects the inherent law of the data, thereby ensuring its effectiveness and accuracy. In all the calculation formulas in the present application, the parameters in each formula are processed by consistent range of dimensionless to ensure that different physical quantities are compared on the same scale; the dimensionless technique includes but is not limited to Min-Max normalization, Z-Score standardization;

[0099] The technical solutions of the present application can be embodied in the form of a software product, which can be stored in a computer-readable storage medium such as a floppy disk, a read-only memory (ROM), a random access memory (RAM), a FLASH, a hard disk or an optical disk, etc., including a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the method of each embodiment of the present application.

[0100] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered a list of executable instructions for implementing logical functions, and can be specifically embodied in any computer-readable medium for use by an instruction execution system, apparatus or device, such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, apparatus or device and execute the instructions, or in conjunction with these instruction execution systems, apparatus or devices. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transport programs for use by an instruction execution system, apparatus or device, or in conjunction with these instruction execution systems, apparatus or devices.

[0101] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and all should be covered in the scope of the claims of the present application.

Claims

1. A wafer fabrication method based on image segmentation, characterized in that, The specific steps include: Step S1: Collect thermal image data of the surface of the wafer to be inspected, and divide the thermal image data into multiple fixed-size measurement areas based on the image segmentation algorithm. During the monitoring period, collect thermal response data of each measurement area under the action of a preset temperature change to form a multi-region spatiotemporal thermal response matrix. Step S2: Analyze the thermal response data of each measurement area during the monitoring period using the thermal diffusion equation to determine the temperature change, mark the abnormal thermal diffusion areas, and determine the optimal preset temperature change amount. Step S3: Apply multi-wavelength measurement light to the marked thermal diffusion anomaly areas in sequence, and perform multivariate linear analysis on the thermal response data of the thermal diffusion anomaly areas and the optimal preset temperature change to dynamically adjust the reference light intensity of the multi-wavelength measurement light and obtain the adjusted light intensity. Step S4: Apply micro-vibration to each thermal diffusion anomalous region and collect the corresponding vibration response data, while simultaneously acquiring the optical response data under each thermal diffusion anomalous region; Vibration response data, thermal response data, and optical response data are input into a pre-trained defect prediction model. The defect prediction model uses labeled data from a historical defect database to automatically extract features and perform correlation analysis, generating preliminary defect measurement indicators that reflect the severity of defects in the area. Step S5: Based on the preliminary defect measurement indicators of each thermal diffusion anomaly region, apply different combinations of polarized structured light to each thermal diffusion anomaly region, and correct the preliminary defect measurement indicators according to the degree of polarization and phase delay. It also includes performing multi-point simulation measurements of the preset temperature change within a preset range to determine the optimal preset temperature change that maximizes the global temperature change, specifically including: m measurement regions are randomly selected from multiple measurement regions; a candidate set of preset temperature change amounts is defined; multi-point simulation is performed on each candidate preset temperature change amount; the multi-point simulation first sets the upper and lower limits of the wafer surface temperature; then the m measurement regions are pre-scanned to obtain the corresponding temperature time series data; Given a preset temperature change, calculate the maximum temperature change range of the sampled measurement area during the monitoring period; and calculate the average change magnitude corresponding to the candidate preset temperature change. Among all candidate preset temperature changes, the preset temperature change with the largest average change amplitude is selected as the optimal preset temperature change. It also includes calculating the residuals based on the finite difference method of the two-dimensional thermal diffusion equation under the optimal preset temperature change condition, and marking the measurement area exceeding the first preset threshold as a thermal diffusion anomaly area, specifically including: For the temperature time series data of measurement area i, calculate the approximate time derivative; and determine the temperature change amplitude of measurement area i in the approximate time derivative; For the temperature values ​​of the measurement region i and its four neighboring regions, calculate the Laplace operator in the discretized space; For each measurement region i, calculate the residual of the discrete form of the heat diffusion equation. ; Calculate the mean residual of all measured areas with standard deviation Set the anomaly determination coefficient k, and set... It is the first preset threshold, when At that time, the measurement area i is marked as a region of abnormal thermal diffusion. ; The defect prediction model employs a deep neural network or machine learning algorithm, specifically: The piezoelectric ceramic micro-vibration stage was placed in a region of abnormal thermal diffusion. The above is used to collect vibration response data characterized by the original vibration signal; the original vibration signal is subjected to a fourth-order Butterworth bandpass filter to obtain the filtered signal; Based on the original vibration signal, the thermal diffusion anomaly region was calculated respectively. Root mean square characteristics and peak characteristics; Obtaining the thermal diffusion anomaly region Light intensity time series data; and calculate the thermal diffusion anomaly regions respectively. The average optical response and standard deviation of the optical response; Obtaining the thermal diffusion anomaly region The magnitude of temperature change in the thermal response data; The root mean square feature, peak feature, average light response, standard deviation of light response, and temperature change amplitude are used as input features of the defect prediction model; preliminary defect measurement indicators are used as output features of the defect prediction model. Based on the historical training set label distribution, calculate the mean of the training set corresponding to the defect prediction model. with standard deviation Therefore, a second preset threshold is set. ; If the preliminary defect measurement index is greater than the second preset threshold In this case, the areas with abnormal thermal diffusion should be marked in the "Preliminary Defect Table". It is classified as a "suspected defect".

2. The wafer fabrication method based on image segmentation according to claim 1, characterized in that: Set the reference temperature of the wafer surface to be tested, and set the preset temperature change amount; At a reference temperature, thermal response data, characterized by temperature time series data, is acquired for each measurement region i on the surface of the wafer under test during the monitoring period.

3. The wafer fabrication method based on image segmentation according to claim 2, characterized in that: For each region of thermal diffusion anomaly Two adjustable light sources of different wavelengths are applied sequentially to dynamically adjust the reference light intensity of the multi-wavelength measurement light, resulting in the adjusted light intensity. The specific logic includes: Two light sources were selected: a 1.5µm laser and an 850nm LED; and the reference light intensity for both wavelengths was set to be... ; Based on reference light intensity Introducing anomaly regions of thermal diffusion Multiple linear analysis was performed on the temperature variation range and the optimal preset temperature variation to obtain the heat diffusion anomaly region. Adjusted light intensity ; For each region of thermal diffusion anomaly Two light sources were used for irradiation, in a fixed order: 1.5µm laser → 850nm LED; During the irradiation period, the thermal diffusion anomaly region was recorded using the sampling rate of the photoresponse signal. Light intensity time series data.

4. The wafer fabrication method based on image segmentation according to claim 3, characterized in that: The different combinations of polarization structured light include four linear polarization states: 0°, 45°, 90°, and 135°. The four polarization states are switched sequentially, and the corresponding reflected light intensities are collected. , , , ; Based on the collected reflected light intensity , , , Calculate the thermal diffusion anomaly region polarization degree and phase delay; The polarization degree and phase delay were Z-score normalized according to the statistics of the training set, and a multiplicative correction model was constructed based on the normalization results to output the thermal diffusion anomaly regions. The final defect measurement index.

5. The wafer fabrication method based on image segmentation according to claim 4, characterized in that: The second preset threshold is referenced again. If the final defect measurement index is greater than the second preset threshold At that time, the area with abnormal heat diffusion will be... Mark as "Key Re-inspection"; otherwise mark as "Routine Inspection".

6. A wafer fabrication system based on image segmentation, characterized in that: The system is used to execute the image segmentation-based wafer fabrication method according to any one of claims 1-5, comprising: Thermal response data generation module: used to collect thermal image data of the wafer surface to be inspected, and divide the thermal image data into multiple fixed-size measurement areas based on the image segmentation algorithm. During the monitoring period, thermal response data of each measurement area under the action of a preset temperature change is collected to form a multi-region spatiotemporal thermal response matrix. Anomaly Marking Module: Used to analyze the thermal response data of each measurement area during the monitoring period using the thermal diffusion equation to determine the temperature change, mark the abnormal thermal diffusion areas, and determine the optimal preset temperature change amount; Light intensity adjustment module: used to sequentially apply multi-wavelength measurement light to the marked heat diffusion anomaly areas, introduce the thermal response data of the heat diffusion anomaly areas and the optimal preset temperature change amount to perform multivariate linear analysis, so as to dynamically adjust the reference light intensity of the multi-wavelength measurement light and obtain the adjusted light intensity. Initial index generation module: used to apply micro-vibration to each thermal diffusion anomaly region and collect the corresponding vibration response data, while also acquiring the light response data under each thermal diffusion anomaly region; Vibration response data, thermal response data, and optical response data are input into a pre-trained defect prediction model. The defect prediction model uses labeled data from a historical defect database to automatically extract features and perform correlation analysis, generating preliminary defect measurement indicators that reflect the severity of defects in the area. Correction module: Based on the preliminary defect measurement indicators of each thermal diffusion anomaly region, it applies different combinations of polarized structured light to each thermal diffusion anomaly region and corrects the preliminary defect measurement indicators according to the degree of polarization and phase delay.

Images