A semiconductor device cleaning path planning and early warning method and system

Abstract

The present invention discloses a semiconductor device cleaning path planning and early warning method and system, belonging to the field of computer data information processing technology; the method comprises acquiring surface parameter data of a semiconductor device to be cleaned, generating a comprehensive feature vector after normalization processing, and constructing a three-dimensional surface model; segmenting the surface area and generating n cleaning units, where n is the number of cleaning units and n is a natural number; using a path planning algorithm to generate a cleaning path; starting cleaning and monitoring the surface parameters of the semiconductor device in real time, dynamically adjusting the cleaning parameters and the cleaning path; using the matching results of the surface information after cleaning and the target feature value, identifying the uncleaned area, and optimizing the cleaning path until the cleaning is completed; the present invention can realize real-time early warning of the laser cleaning process of the semiconductor device, ensure consistency with the target feature value during the cleaning process, and can flexibly process complex surfaces, avoid repeated cleaning and energy waste, and improve overall efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of computer data information processing, and in particular to a semiconductor device cleaning path planning and early warning method and system. Background Art

[0002] With the rapid development of the semiconductor industry, the precision and complexity of semiconductor devices are constantly improving, and the requirements for their cleaning processes are becoming increasingly stringent. Among the many cleaning technologies, laser cleaning has become an important means of cleaning semiconductor devices because of its advantages such as non-contact, no chemical pollution, and strong adaptability. However, traditional laser cleaning technologies mostly rely on manual operation, that is, the operator completes the cleaning task by manually adjusting the equipment parameters and editing the cleaning path. This method not only requires a high level of skill from the operator, but also takes a lot of time to set up the program, resulting in low cleaning efficiency. In addition, since the cleaning path cannot be dynamically adjusted during the cleaning process, the cleaning accuracy may be reduced or even missed due to different device shapes. The existing cleaning mode that combines manual and machine operations is prone to human errors, affecting the overall cleaning effect, and it is difficult to meet the high-precision cleaning needs of modern semiconductor devices.

[0003] At the same time, during the actual cleaning process, the surface temperature of semiconductor devices changes in real time with the intensity and duration of laser irradiation. Traditional technologies generally lack real-time temperature monitoring and early warning capabilities. When the surface temperature is too high, it can cause thermal damage to the device and even pose a safety hazard. Moreover, because temperature changes can cause tiny displacements of the device, traditional technologies have difficulty manually identifying and promptly correcting the cleaning path, which in turn affects the comprehensiveness and uniformity of the cleaning process. To address these shortcomings, existing technologies urgently need an intelligent laser cleaning method that can monitor surface state changes in real time during the cleaning process, dynamically adjust the cleaning path, and provide feedback on the results. Summary of the Invention

[0004] In order to more efficiently and accurately plan and warn the cleaning path of semiconductor devices and dynamically optimize the cleaning process, the present invention provides a semiconductor device cleaning path planning and warning method and system.

[0005] In order to achieve the above-mentioned purpose of the invention, the present invention provides a semiconductor device cleaning path planning and early warning method, which includes the following steps: Step 1: Obtain the surface temperature, reflectivity and spatial depth information of the semiconductor device to be cleaned, generate a comprehensive feature vector after normalization processing, and construct a three-dimensional surface model based on the spatial depth information; at the same time, perform feature extraction on the surface image of the semiconductor device before cleaning to generate an initial image feature vector.

[0006] Step 2: Using the comprehensive feature vector and the three-dimensional surface model, the surface area of ​​the semiconductor device is segmented to generate n cleaning units.

[0007] Step 3: Based on the surface area segmentation results of the semiconductor device, a path planning algorithm is used to generate a cleaning path in combination with the optimization goals of path length and laser energy usage.

[0008] Step 4: Start cleaning and monitor the temperature, reflectivity, and spatial depth information of the semiconductor device surface in real time, and dynamically adjust the cleaning parameters and cleaning path.

[0009] Step 5: Use the matching results of the surface information after cleaning and the target feature value to identify the uncleaned area and optimize the cleaning path until cleaning is completed.

[0010] Among them, step 1 specifically includes: collecting the surface temperature, reflectivity and spatial depth information of the semiconductor device, and performing time-space synchronization correction to establish a multimodal feature matrix; normalizing the data in the multimodal feature matrix, and generating a comprehensive feature vector using a weighted fusion algorithm; constructing a three-dimensional surface model based on the spatial depth information; at the same time, performing feature extraction on the surface image before cleaning to generate an initial image feature vector.

[0011] The surface temperature is obtained by collecting the surface temperature of the semiconductor device at a high frame rate using an infrared thermal imager, and smoothing the dynamic temperature changes using a time-series interpolation algorithm; the reflectivity is obtained by obtaining reflectivity data of multiple bands using a multispectral imaging device, and extracting key spectral features through principal component analysis (PCA); the spatial depth information includes single-point depth measurement values, point cloud data, and depth maps; the spatial depth information is obtained by combining a laser rangefinder and a structured light sensor.

[0012] The generation formula of the comprehensive feature vector is as follows.

[0013]

[0014] Where: T(x) represents the surface temperature value of a specific cleaning area; R(x) represents the principal component eigenvalue of the reflectivity; D(x) represents the eigenvalue of the spatial depth information; μ T ,μ R ,μ D are the mean values ​​of temperature, reflectivity and spatial depth information features respectively; σ T ,σ R ,σ D are the standard deviations of temperature, reflectivity and depth characteristics respectively; W = [w T ,w R ,w D ],w T ,w R ,w D is the weight vector of temperature, reflectivity and spatial depth information features, and satisfies wT +w R +w D =1.

[0015] The step 2 is specifically as follows: based on the comprehensive feature vector and the three-dimensional surface model, the surface of the semiconductor device is divided into n cleaning units by a dynamic segmentation algorithm, and the segmentation granularity is dynamically adjusted according to the complexity of the cleaning unit. The dynamic segmentation algorithm adopts a dynamic clustering segmentation method based on the comprehensive feature vector and spatial depth information, and the segmentation granularity of the dynamic segmentation algorithm is f seg (s i ) is as follows.

[0016]

[0017] Among them, v c (s i ) is the comprehensive feature vector of the i-th cleaning unit; v c is the current cluster center feature vector; ||v(s i )-v c || 2 is the square of the Euclidean distance, which measures the similarity of features between units; It represents the sum of the absolute values ​​of spatial depth gradients, which is used to reflect the complexity of the surface of the cleaning area; γ is the weight factor dynamically adjusted according to the surface complexity; λ is the granularity adjustment weight; A i is the cleaning unit area; n is the number of cleaning units, n is a natural number; s i represents the i-th cleaning unit; T(s i ) is the segmentation granularity threshold that is dynamically adjusted; T(s i ) is the dynamically adjusted segmentation granularity threshold, as follows.

[0018]

[0019] Among them, T0 is the initial value of the segmentation granularity; β is the complexity weight factor, which is used to adjust the influence of the depth gradient on the granularity threshold; || VD(s i )||1 is the absolute value sum of the gradient depth, which is used to measure the point s i The complexity of nearby surfaces; details are as follows.

[0020]

[0021] e is s i The number of sampled gradient directions near the point, e = 3; |VD k (s)|for s i The depth gradient change value of the point in the kth direction, k≤e; k represents one of the sampled gradient directions in e; ||VD||max is the global maximum depth gradient value, which is used to normalize the gradient value so that ||VD(s i )||1's dynamic adjustment range is between 0 and 1, as follows.

[0022] ||VD|| max =max x∈Y ||VD(s i )||1.

[0023] Y is the collection of surface models of the entire semiconductor device.

[0024] In step 3, based on the segmented cleaning units, a cleaning path planning strategy is generated using a deep reinforcement learning algorithm, and the optimal cleaning path is determined through a path optimization module; the path planning objective function F is defined. path The optimal trade-off between path length and cleaning performance is expressed as follows.

[0025]

[0026] Among them, P j is the three-dimensional coordinate of the j-th path point; ||P j -P j+1 || represents the Euclidean distance between adjacent path points; E laser (x j ) represents the laser power distribution function; v p (x j ) is the comprehensive feature vector of the j-th path point; x j represents the jth path point; v target is the comprehensive feature vector of the target area; α1 and α2 are the weight coefficients of path length and cleaning performance respectively; m is the number of path points; P0 and P1 represent the starting point and end point of the cleaning path respectively.

[0027] The cleaning parameters and cleaning path are dynamically adjusted in step 4, specifically: during the cleaning process, the temperature, reflectivity and spatial depth information of the surface of the semiconductor device are monitored in real time, and the monitoring data are compared with the target characteristic value generated before cleaning; when the difference between the monitoring data and the target characteristic value exceeds a preset threshold, an early warning signal is triggered, and the cleaning parameters, including the power, frequency and pulse width of the laser, are dynamically adjusted, and the cleaning path is dynamically updated according to the cleaning progress.

[0028] The cleaning path is optimized in step 5, specifically: after cleaning, feature extraction is performed on the cleaned surface image, and similarity is calculated between the feature vector of the cleaned image and the feature vector of the initial image before cleaning to evaluate the cleaning effect; quantitative analysis is performed on the uncleaned area, and when the area of ​​the uncleaned area exceeds the tolerance range, an early warning signal is triggered and a local optimization path is started until the cleaning is completed; specifically: the generation process of the local optimization path is as follows.

[0029] 1) Re-assign the order of cleaning path points so that uncleaned areas are cleaned first.

[0030] 2) Dynamically adjust the granularity of the cleaning path and redistribute the distribution of laser energy so that the cleaning path covers the uncleaned area more concentratedly.

[0031] 3) Perform local path fine-tuning in the uncleaned area, and use the objective function to optimize the path length and cleaning effect. The local optimization path objective function is defined as follows.

[0032]

[0033] Among them, P j is the three-dimensional coordinate of the j-th path point, ||P j -P j+1 || is the Euclidean distance between adjacent path points, γ is the weight coefficient in path optimization, δ(x j ) is the uncleaned area point x j The cleaning residual value, z is the total number of path points in the local path.

[0034] In order to better achieve the above-mentioned purpose of the invention, the present invention also provides a semiconductor device cleaning path planning and early warning system, which is used to execute the semiconductor device cleaning path planning and early warning method. The system includes: a data acquisition module for obtaining the surface temperature, reflectivity and spatial depth information of the semiconductor device to be cleaned, and normalizing the information to generate a comprehensive feature vector; at the same time, constructing a three-dimensional surface model based on the spatial depth information.

[0035] The region segmentation module is used to segment the surface of the semiconductor device based on the comprehensive feature vector and the three-dimensional surface model through a region segmentation algorithm, and divide the surface into n cleaning units using a dynamic segmentation strategy.

[0036] The path planning module is used to generate a cleaning path based on the area segmentation result, using a path planning algorithm and combining the optimization goals of path length and laser energy usage.

[0037] The dynamic monitoring and parameter adjustment module is used to monitor the temperature, reflectivity and spatial depth information of the surface of semiconductor devices in real time during the cleaning process, compare the monitoring data with the target characteristic value, and trigger an early warning signal when the difference exceeds the preset threshold, and dynamically adjust the cleaning parameters and update the cleaning path.

[0038] The cleaning effect evaluation and optimization module is used to match and analyze the information on the surface of the semiconductor device after cleaning with the target characteristic value, identify the uncleaned area, and optimize the cleaning path based on the quantitative analysis results of the uncleaned area, and repeat the execution until the cleaning is completed.

[0039] The beneficial effects of the present invention are as follows: the method of the present invention can obtain temperature, reflectivity and depth spatial information in real time during the laser cleaning process of semiconductor devices to achieve real-time early warning, ensure that the cleaning process is consistent with the target characteristic value, and avoid repeated cleaning and energy waste through dynamic segmentation strategy and path planning, can flexibly process complex surfaces, achieve refined cleaning, and improve overall efficiency. Specifically, by obtaining surface multimodal feature information and generating comprehensive feature vectors and three-dimensional surface models, the integrity and spatial accuracy of data expression are significantly improved, providing a reliable basis for subsequent cleaning; through dynamic segmentation strategy, the complex surface area of ​​semiconductor devices is dynamically divided into n cleaning units, which improves the targeting and resolution of the cleaning area; based on path planning, the optimization goals of path length and cleaning performance are integrated to generate an efficient and energy-saving cleaning path, effectively reducing resource consumption; by real-time monitoring of key parameters in the cleaning process and dynamically adjusting the cleaning path and parameters, the stability and efficiency of the cleaning process are guaranteed; finally, by accurately identifying and optimizing the uncleaned area, the comprehensiveness and consistency of the cleaning effect are ensured. Overall, the method of the present invention significantly improves cleaning efficiency and accuracy, while reducing energy consumption, and is widely applicable to high-demand semiconductor device cleaning scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flowchart of Example 1 and Example 2 of the present invention.

[0041] Figure 2 This is a system block diagram of Example 3.

[0042] Figure 3 This is a comparative test data diagram of Example 4.

[0043] Figure 4 This is a comparison diagram of the semiconductor device before and after cleaning using the method of Example 2.

[0044] Figure symbols: A, B, C represent the categories of semiconductor device samples, A1, A2, B1, B2, C1, C2 represent the numbers of semiconductor device samples; n is the number of cleaning units, and n is a natural number. DETAILED DESCRIPTION

[0045] In response to the above-mentioned problems in the prior art, the present invention proposes a semiconductor device cleaning path planning and early warning method and system. By obtaining the surface temperature, reflectivity and spatial depth information of the semiconductor device, combining normalization processing to generate a comprehensive feature vector, and constructing a three-dimensional surface model based on the spatial depth information, the cleaning area can be accurately segmented. Compared with traditional methods, the present invention adopts a dynamic segmentation strategy to dynamically adjust the segmentation granularity of the cleaning unit according to the surface complexity, ensuring that the segmentation result is more adaptable. In terms of path planning, the deep reinforcement learning algorithm is combined with the optimization goals of path length and laser energy usage to dynamically generate the optimal cleaning path, significantly improving the cleaning efficiency and accuracy. In addition, by real-time monitoring of the changes in the surface state during the cleaning process, when the difference between the monitoring data and the target feature value exceeds a preset threshold, the present invention can trigger an early warning signal and dynamically adjust the cleaning parameters and path, significantly reducing the safety hazards and omissions caused by excessive temperature or displacement. After cleaning is completed, the present invention compares the surface feature values ​​before and after cleaning to evaluate the cleaning effect and optimize the path planning of the uncleaned area, ensuring the comprehensiveness and consistency of the cleaning results.

[0046] In order to clearly illustrate the technical features of this solution, this solution is described below through specific implementation methods.

[0047] Example 1

[0048] An embodiment of the present invention provides a semiconductor device cleaning path planning and early warning method, comprising the following steps: Step 1: obtaining surface temperature, reflectivity and spatial depth information of a semiconductor device to be cleaned, generating a comprehensive feature vector after normalization processing, and constructing a three-dimensional surface model based on the spatial depth information; wherein, the surface temperature is acquired by an infrared thermal imager at a high frame rate to collect the surface temperature of the semiconductor device, and a time series interpolation algorithm is used to smooth the dynamic temperature changes; the reflectivity is acquired by a multispectral imaging device to obtain reflectivity data of multiple bands, and key spectral features are extracted by principal component analysis (PCA); the spatial depth information includes single-point depth measurement values, point cloud data, and a depth map; the spatial depth information is acquired by combining a laser rangefinder and a structured light sensor.

[0049] Furthermore, a multimodal feature matrix is ​​formed using the acquired surface temperature, reflectivity, and spatial depth information. This is then normalized using the Dynamic Range Normalization (DRN) algorithm, aligning the data from different modalities. Specifically, the collected surface temperature, reflectivity, and spatial depth information of the semiconductor device is subjected to spatiotemporal synchronization to establish a multimodal feature matrix. The data in the multimodal feature matrix is ​​normalized, and a weighted fusion algorithm is used to generate a comprehensive feature vector. A three-dimensional surface model is constructed based on the spatial depth information to faithfully restore the microscopic geometric properties of the semiconductor device surface, facilitating subsequent region segmentation and path planning. Simultaneously, feature extraction is performed on the pre-cleaned surface image to generate an initial image feature vector.

[0050] The generation formula of the comprehensive feature vector is as follows.

[0051]

[0052] Where: T(x) represents the surface temperature value of a specific cleaning area; R(x) represents the principal component eigenvalue of the reflectivity; D(x) represents the eigenvalue of the spatial depth information; μ T ,μ R ,μ D are the mean values ​​of temperature, reflectivity and spatial depth information features respectively; σ T ,σ R ,σ D are the standard deviations of temperature, reflectivity and depth features respectively. T ,w R ,w D ],w T ,w R ,w D is the weight vector of temperature, reflectivity and spatial depth information features, and satisfies w T +w R +w D =1.

[0053] The comprehensive feature vector of this embodiment is a high-dimensional feature expression generated by a weighted fusion algorithm of multimodal data (including surface temperature, reflectivity, and spatial depth information), which is used to unify the data dimensions of different physical quantities to improve the accuracy and robustness of cleaning path planning.

[0054] Step 2: Use the comprehensive feature vector and three-dimensional surface model to segment the surface area and generate n cleaning units. Specifically: Based on the comprehensive feature vector and three-dimensional surface model in step 1, the surface of the semiconductor device is divided into n cleaning units through a dynamic segmentation algorithm, and the segmentation granularity is dynamically adjusted according to the complexity of the cleaning unit, so that the cleaning units in complex areas are finer and the cleaning units in simple areas are coarser; the cluster center is dynamically updated, and the feature similarity calculation is optimized through neighborhood weighting. Among them, the dynamic segmentation algorithm adopts a dynamic clustering segmentation method based on comprehensive feature vectors and spatial depth information, and the segmentation granularity f of the dynamic segmentation algorithm is 10 ... seg (s i ) is as follows.

[0055]

[0056] Among them, v c (s i ) is the comprehensive feature vector of the i-th cleaning unit; v c is the current cluster center feature vector; ||v(s i )-v c || 2 is the square of the Euclidean distance, which measures the similarity of features between units; It represents the sum of the absolute values ​​of spatial depth gradients, which is used to reflect the complexity of the surface of the cleaning area; γ is the weight factor dynamically adjusted according to the surface complexity; λ is the granularity adjustment weight; A i is the cleaning unit area; n is the number of cleaning units, n is a natural number; s i represents the i-th cleaning unit; T(s i ) is the dynamically adjusted segmentation granularity threshold.

[0057] Step 3: Based on the surface area segmentation results, a path planning algorithm is used to generate a cleaning path, combined with the optimization goals of path length and laser energy usage, to ensure that the cleaning process achieves the best effect in the shortest time. Specifically: Based on the segmented cleaning units, a deep reinforcement learning algorithm is used to generate a cleaning path planning strategy, and the path optimization module is used to determine the optimal cleaning path; the path planning objective function F is defined. path The optimal trade-off between path length and cleaning performance is expressed as follows.

[0058]

[0059] Among them, P j is the three-dimensional coordinate of the j-th path point; ||P j -P j+1 || represents the Euclidean distance between adjacent path points; E laser (x j) represents the laser power distribution function; v p (x j ) is the comprehensive feature vector of the j-th path point; x j represents the jth path point; v target is the comprehensive feature vector of the target area; α1 and α2 are the weight coefficients of path length and cleaning performance respectively; m is the number of path points; P0 and P1 represent the starting point and end point of the cleaning path respectively.

[0060] Step 4: Start cleaning and monitor the temperature, reflectivity, and spatial depth of the semiconductor device surface in real time, dynamically adjusting the cleaning parameters and cleaning path. During the cleaning process, the temperature, reflectivity, and spatial depth of the semiconductor device surface are monitored in real time, and the monitoring data is compared with the target characteristic value generated before cleaning. When the difference between the monitoring data and the target characteristic value exceeds a preset threshold, an early warning signal is triggered, and the cleaning parameters, including the laser power, frequency, and pulse width, are dynamically adjusted. The cleaning path is also dynamically updated based on the cleaning progress.

[0061] Step 5: Using the matching results of the cleaned surface information and the target feature values, the uncleaned areas are identified and the cleaning path is optimized until cleaning is complete. After cleaning, the cleaned surface image is subjected to feature extraction, and the similarity between the cleaned image feature vector and the initial image feature vector is calculated to evaluate the cleaning effect. The uncleaned areas are quantitatively analyzed. When the area of ​​the uncleaned area exceeds the tolerance range, an early warning signal is triggered and a path optimization solution is recommended. The above steps are repeated until cleaning is complete.

[0062] Example 2

[0063] An embodiment of the present invention provides a semiconductor device cleaning path planning and early warning method, comprising the following steps: Step 1: obtaining surface temperature, reflectivity and spatial depth information of a semiconductor device to be cleaned, generating a comprehensive feature vector after normalization processing, and constructing a three-dimensional surface model based on the spatial depth information; wherein, the surface temperature is acquired by an infrared thermal imager at a high frame rate to collect the surface temperature of the semiconductor device, and a time series interpolation algorithm is used to smooth the dynamic temperature changes; the reflectivity is acquired by a multispectral imaging device to obtain reflectivity data of multiple bands, and key spectral features are extracted by principal component analysis (PCA); the spatial depth information includes single-point depth measurement values, point cloud data, and a depth map; the spatial depth information is acquired by combining a laser rangefinder and a structured light sensor.

[0064] Furthermore, a multimodal feature matrix is ​​formed using the acquired surface temperature, reflectivity, and spatial depth information. This is then normalized using the Dynamic Range Normalization (DRN) algorithm, aligning the data from different modalities. Specifically, the collected surface temperature, reflectivity, and spatial depth information of the semiconductor device is subjected to spatiotemporal synchronization to establish a multimodal feature matrix. The data in the multimodal feature matrix is ​​normalized, and a weighted fusion algorithm is used to generate a comprehensive feature vector. A three-dimensional surface model is constructed based on the spatial depth information to faithfully restore the microscopic geometric properties of the semiconductor device surface, facilitating subsequent region segmentation and path planning. Simultaneously, feature extraction is performed on the pre-cleaned surface image to generate an initial image feature vector.

[0065] The generation formula of the comprehensive feature vector is as follows.

[0066]

[0067] Where: T(x) represents the surface temperature value of a specific cleaning area; R(x) represents the principal component eigenvalue of the reflectivity; D(x) represents the eigenvalue of the spatial depth information; μ T ,μ R ,μ D are the mean values ​​of temperature, reflectivity and spatial depth information features respectively; σ T ,σ R ,σ D are the standard deviations of temperature, reflectivity and depth features respectively. T ,w R ,w D ],w T ,w R ,w D is the weight vector of temperature, reflectivity and spatial depth information features, and satisfies w T +w R +w D =1.

[0068] The comprehensive feature vector of this embodiment is a high-dimensional feature expression generated by a weighted fusion algorithm of multimodal data (including surface temperature, reflectivity, and spatial depth information), which is used to unify the data dimensions of different physical quantities to improve the accuracy and robustness of cleaning path planning.

[0069] Among them, the surface temperature is collected by an infrared thermal imager at a high frame rate on the surface temperature of the semiconductor device, and a time series interpolation algorithm is used to smooth the dynamic temperature changes; the time series interpolation algorithm is a local weighted regression interpolation algorithm, and the surface temperature acquisition method includes the following steps.

[0070] 1) Collect the surface temperature data of semiconductor devices through infrared thermal imager and generate time series data T(t1),T(t2),...,T(t n ), where T(t i ) is at time point t i The surface temperature value of the semiconductor device collected at the time.

[0071] 2) For each time point t i , select a local time window t i-w ,t i+w , where w is the local time window size, which represents the time range used for weighted regression calculation.

[0072] 3) Calculate each time point t based on the Gaussian kernel function j t i The weight w ij。

[0073]

[0074] Among them, t i and t j is the current processing time point and the domain time point, σ is the standard deviation of the Gaussian kernel function, and controls the degree of attenuation of the weight with the time point gap.

[0075] 4) Use the weighted regression model to calculate the smoothed temperature value T smooth (t i ).

[0076]

[0077] where w ij is the weight, T(t j ) is the time point t j The corresponding original temperature value.

[0078] 5) Output smoothed temperature time series T smooth (t1), T smooth (t2),…,T smooth (t n ) for subsequent analysis and temperature change trend prediction.

[0079] This method can dynamically adjust the local time window w and the parameter σ of the Gaussian kernel function according to the severity of the temperature change to adapt to different temperature change trends, ensuring that the smoothing process has a higher sensitivity to rapidly changing temperature data and provides a stronger smoothing effect for steadily changing temperature data.

[0080] Step 2: Use the comprehensive feature vector and three-dimensional surface model to segment the surface area and generate n cleaning units. Specifically: Based on the comprehensive feature vector and three-dimensional surface model in step 1, the surface of the semiconductor device is divided into n cleaning units through a dynamic segmentation algorithm, and the segmentation granularity is dynamically adjusted according to the complexity of the cleaning unit, so that the cleaning units in complex areas are finer and the cleaning units in simple areas are coarser; the cluster center is dynamically updated, and the feature similarity calculation is optimized through neighborhood weighting. Among them, the dynamic segmentation algorithm adopts a dynamic clustering segmentation method based on comprehensive feature vectors and spatial depth information, and the segmentation granularity f of the dynamic segmentation algorithm is 10 ... seg (s i ) is as follows.

[0081]

[0082] Among them, v c (s i ) is the comprehensive feature vector of the i-th cleaning unit; v c is the current cluster center feature vector; ||v(s i )-v c || 2 is the square of the Euclidean distance, which measures the similarity of features between units; It represents the sum of the absolute values ​​of spatial depth gradients, which is used to reflect the complexity of the surface of the cleaning area; γ is the weight factor dynamically adjusted according to the surface complexity; λ is the granularity adjustment weight; A i is the cleaning unit area; n is the number of cleaning units, n is a natural number; s i represents the i-th cleaning unit; T(s i ) is the segmentation granularity threshold that is dynamically adjusted; T(s i ) is the dynamically adjusted segmentation granularity threshold, as follows.

[0083]

[0084] Among them, T0 is the initial value of the segmentation granularity; β is the complexity weight factor, which is used to adjust the influence of the depth gradient on the granularity threshold; || VD(s i )||1 is the absolute value sum of the gradient depth, which is used to measure the point s i The complexity of nearby surfaces; details are as follows.

[0085]

[0086] e is s i The number of sampled gradient directions near the point, e = 3; |VD k (s)|for s iThe depth gradient change value of the point in the kth direction, k≤e; k represents one of the sampled gradient directions in e; ||VD|| max is the global maximum depth gradient value, which is used to normalize the gradient value so that ||VD(s i )||1's dynamic adjustment range is between 0 and 1, as follows.

[0087] ||VD|| max =max x∈Y ||VD(s i )||1.

[0088] Y is the collection of surface models of the entire semiconductor device.

[0089] Step 3: Based on the surface area segmentation results, a path planning algorithm is used to generate a cleaning path, combined with the optimization goals of path length and laser energy usage, to ensure that the cleaning process achieves the best effect in the shortest time. Specifically: Based on the segmented cleaning units, a deep reinforcement learning algorithm is used to generate a cleaning path planning strategy, and the path optimization module is used to determine the optimal cleaning path; the path planning objective function F is defined. path The optimal trade-off between path length and cleaning performance is expressed as follows.

[0090]

[0091] Among them, P j is the three-dimensional coordinate of the j-th path point; ||P j -P j+1 || represents the Euclidean distance between adjacent path points; E laser (x j ) represents the laser power distribution function; v p (x j ) is the comprehensive feature vector of the j-th path point; x j represents the jth path point; v target is the comprehensive feature vector of the target area; α1 and α2 are the weight coefficients of path length and cleaning performance respectively; m is the number of path points; P0 and P1 represent the starting point and end point of the cleaning path respectively.

[0092] Step 4 dynamically adjusts the cleaning parameters and cleaning path. Specifically, the cleaning process begins and monitors the temperature, reflectivity, and spatial depth of the semiconductor device surface in real time, dynamically adjusting the cleaning parameters and cleaning path. During the cleaning process, the temperature, reflectivity, and spatial depth of the semiconductor device surface are monitored in real time and compared with the target characteristic values ​​generated before cleaning. When the difference between the monitored data and the target characteristic values ​​exceeds a preset threshold, an early warning signal is triggered, and the cleaning parameters, including the laser power, frequency, and pulse width, are dynamically adjusted. The cleaning path is dynamically updated based on the cleaning progress. The baseline operating parameters for the laser are: laser power: 480W, laser pulse width: 500ns, scanning frequency: 40Hz, buffer length: 4mm, offset length: 5mm. The preset temperature threshold is set between ±1°C and ±5°C. The preset reflectivity threshold is set between ±5% and ±10%. The preset spatial depth threshold is set between ±1% and ±3%.

[0093] In step 5, the cleaning path is optimized by matching the cleaned surface information with the target feature values ​​to identify uncleaned areas and optimizing the cleaning path until cleaning is complete. After cleaning, features are extracted from the cleaned surface image, and the similarity between the cleaned image feature vector and the initial image feature vector is calculated to evaluate the cleaning effect. A quantitative analysis of the uncleaned areas is performed. When the area of ​​the uncleaned area exceeds the tolerance range, an early warning signal is triggered and a local optimization path is initiated until cleaning is complete. The tolerance range is set to 0.5% to 5%. Specifically, the process of generating the local optimization path is as follows.

[0094] 1) Re-assign the order of cleaning path points so that uncleaned areas are cleaned first.

[0095] 2) Dynamically adjust the granularity of the cleaning path and redistribute the distribution of laser energy so that the cleaning path covers the uncleaned area more concentratedly.

[0096] 3) Perform local path fine-tuning in the uncleaned area, and use the objective function to optimize the path length and cleaning effect. The local optimization path objective function is defined as follows.

[0097]

[0098] Among them, P j is the three-dimensional coordinate of the j-th path point, ||P j -P j+1 || is the Euclidean distance between adjacent path points, γ is the weight coefficient in path optimization, δ(x j ) is the uncleaned area point x j The cleaning residual value, z is the total number of path points in the local path.

[0099] When performing local cleaning on uncleaned areas, the cleaning effect needs to be monitored in real time, and the laser power, frequency, and pulse width need to be dynamically adjusted based on the feedback information after cleaning.

[0100] Example 3

[0101] Based on the method of Example 1 or Example 2 of the present invention, an embodiment of the present invention provides a semiconductor device cleaning path planning and early warning system, including the following modules: a data acquisition module, used to obtain the surface temperature, reflectivity and spatial depth information of the semiconductor device to be cleaned, and normalize the information to generate a comprehensive feature vector; at the same time, construct a three-dimensional surface model based on the spatial depth information.

[0102] The region segmentation module is used to segment the surface of the semiconductor device based on the comprehensive feature vector and the three-dimensional surface model through a region segmentation algorithm, and divide the surface into n cleaning units using a dynamic segmentation strategy.

[0103] The path planning module is used to generate a cleaning path based on the area segmentation result, using a path planning algorithm and combining the optimization goals of path length and laser energy usage.

[0104] The dynamic monitoring and parameter adjustment module is used to monitor the temperature, reflectivity and spatial depth information of the surface of semiconductor devices in real time during the cleaning process, compare the monitoring data with the target characteristic value, and trigger an early warning signal when the difference exceeds the preset threshold, and dynamically adjust the cleaning parameters and update the cleaning path.

[0105] The cleaning effect evaluation and optimization module is used to match and analyze the information on the surface of the semiconductor device after cleaning with the target characteristic value, identify the uncleaned area, and optimize the cleaning path based on the quantitative analysis results of the uncleaned area, and repeat the execution until the cleaning is completed.

[0106] The data acquisition module specifically includes: a temperature acquisition unit, which uses an infrared thermal imager to collect surface temperature at a high frame rate, and uses a time series interpolation algorithm to smooth dynamic temperature changes.

[0107] The reflectivity acquisition unit obtains reflectivity data of multiple bands through multispectral imaging equipment and extracts key spectral features through principal component analysis (PCA).

[0108] The spatial depth acquisition unit obtains single-point depth measurement values, point cloud data and depth maps through a combination of a laser rangefinder and a structured light sensor.

[0109] The data normalization processing unit uses the dynamic range normalization (DRN) algorithm to unify the dimension of data of different modalities and generate a multimodal feature matrix.

[0110] The region segmentation module is based on an adaptive threshold method, dynamically adjusts the segmentation granularity according to the complexity of the cleaning unit, and generates n cleaning units.

[0111] The dynamic monitoring and parameter adjustment module can adjust the power, frequency and pulse width of the laser in real time to adapt to the dynamic changes in the surface state during the cleaning process, and dynamically update the cleaning path according to the cleaning progress.

[0112] The cleaning effect evaluation and optimization module includes: an image feature extraction unit for extracting surface image feature vectors before and after cleaning.

[0113] The similarity calculation unit is used to calculate the similarity between the feature vector of the cleaned image and the feature vector of the initial image to evaluate the cleaning effect.

[0114] The path optimization unit is used to perform quantitative analysis on the uncleaned area. When the area of ​​the uncleaned area exceeds the tolerance range, an early warning signal is triggered and an optimized path is recommended to complete the cleaning.

[0115] Example 4

[0116] In order to verify the beneficial effects of the above embodiments, the embodiments of the present invention are scientifically demonstrated through simulation experiments. The experimental preparation and implementation process are as follows.

[0117] 1. Experimental Background and Purpose

[0118] To verify the technical effectiveness of the semiconductor device cleaning path planning and early warning method of the present invention, a comparative experiment was conducted with existing technologies. This experiment simulated the dynamic changes in surface temperature, reflectivity, and spatial depth information during the actual cleaning process. By obtaining comprehensive feature vectors and a three-dimensional surface model, combined with region segmentation and path planning algorithms, cleaning efficiency and accuracy were evaluated.

[0119] 2. Experimental apparatus and materials

[0120] Semiconductor device sample: Standard size (100mm×100mm) graphite sample. The equipment is as follows.

[0121] 1) Infrared thermal imager (resolution 0.01°C, frame rate 100 Hz).

[0122] 2) Multispectral imaging equipment (10 spectral bands, band range 400nm-1000nm).

[0123] 3) Laser rangefinder (accuracy 0.1mm).

[0124] 4) Structured light sensor (resolution 1024×768).

[0125] 5) Laser cleaning system (power range 10W-100W).

[0126] 6) Data processing software: MATLAB is used for comprehensive eigenvector calculation and algorithm verification.

[0127] 3. Experimental conditions

[0128] 1) Sample classification: Samples are divided into three categories based on surface characteristics (temperature gradient range, reflectivity complexity, depth information change, etc.).

[0129] Category A: Good surface uniformity, small temperature gradient (5-10°C), and relatively low changes in reflectivity and depth information.

[0130] Category B: Average surface uniformity, moderate temperature gradient (10-20°C), moderate reflectivity complexity, and large variations in local depth information.

[0131] Category C: The surface complexity is high, the temperature gradient is large (above 20°C), the reflectivity is complex and the depth information gradient fluctuates violently.

[0132] 2) Adjust cleaning parameters.

[0133] Comparison group (existing technology): Cleaning was performed based on a fixed path and power according to the traditional path planning algorithm, without real-time warning and dynamic adjustment.

[0134] Experimental group (present invention): Path planning was generated based on the comprehensive feature vector, and cleaning parameters (laser power, frequency, path) were dynamically adjusted to optimize the cleaning effect.

[0135] 3) Test data recording.

[0136] For each sample, the surface cleanliness score before and after cleaning was recorded (0-100 points, the higher the score, the cleaner it was).

[0137] The cleaning path length, cleaning energy consumption (laser power consumption) were recorded, and the cleaning efficiency (cleaning area / time) was calculated.

[0138] Compare the area ratio of uncleaned area after cleaning (uncleaned area / total area).

[0139] The scoring criteria for cleanliness in this embodiment are based on the following quantitative indicators.

[0140] a. Scoring range: 0-100 points (the higher the score, the better the cleanliness).

[0141] b. Basis for scoring.

[0142] ① Clean the surface residue (check with a microscope or electron microscope).

[0143] No residue: full score (100 points).

[0144] ② For every 10% increase in the proportion of residue area, 10 points will be deducted.

[0145] Surface contaminant removal (via mass spectrometry or chemical detection).

[0146] Removal rate ≥95%: full score (100 points).

[0147] For every 5% decrease, 10 points will be deducted.

[0148] ③ Changes in surface roughness (detected by atomic force microscopy).

[0149] Roughness change ≤ 5%: no points deduction.

[0150] For every 5% increase, 5 points will be deducted.

[0151] c. Cleanliness score calculation formula.

[0152] Cleanliness score = (residue score + contaminant removal rate score + roughness change score) / 3.

[0153] IV. Experimental Procedure

[0154] 1. Data collection and processing

[0155] The dynamic temperature changes on the sample surface are captured at a high frame rate using an infrared thermal imager, and the data is smoothed using a time-series interpolation algorithm. Multispectral imaging equipment is used to acquire reflectance data in multiple bands, and principal component analysis is used to extract key features. A laser rangefinder and structured light sensor are used to obtain single-point depth, point cloud, and depth map data to generate a three-dimensional surface model. This data is then subjected to spatiotemporal synchronization correction and normalized using the Dynamic Range Normalization (DRN) algorithm.

[0156] 2. Feature Fusion

[0157] The surface temperature, reflectivity and depth features are combined through a weighted fusion algorithm to generate a comprehensive feature vector, with the weights set as: temperature 0.4, reflectivity 0.3, depth 0.3.

[0158] 3. Region Segmentation

[0159] Using comprehensive feature vectors and a 3D surface model, a dynamic segmentation strategy is used to divide the sample surface into regions and generate multiple cleaning units. The segmentation granularity is dynamically adjusted according to the surface complexity.

[0160] 4. Path Planning

[0161] By combining deep reinforcement learning algorithms with an optimization objective function, the optimal cleaning path is generated. Path planning aims to minimize the path and maximize cleaning efficiency.

[0162] 5. Cleaning and real-time adjustment

[0163] The laser cleaning system cleans using a planned path, monitors surface information in real time, triggers warning signals based on discrepancies, and dynamically adjusts the laser power and path. This process repeats until cleaning is complete.

[0164] 6. Cleaning effect evaluation

[0165] Extract features from the surface after cleaning and compare them with the feature vectors before cleaning to quantify the uncleaned area and optimize the path. See the comparison test data graph. Figure 3 .

[0166] 5. Data Analysis and Results

[0167] Through the data analysis of the comparison group and the experimental group, it can be clearly seen that the technical advantages of the method of the present invention are:

[0168] 1. Cleanliness score: The average cleanliness score of the experimental group samples was significantly higher than that of the control group (the experimental group average was about 90%, and the control group average was about 68%). In particular, under the complex surface conditions of category C, the method of the present invention improved the cleanliness by more than 35%.

[0169] 2. Uncleaned area: The uncleaned area in the experimental group was significantly reduced, and the proportion of uncleaned area on complex surfaces (category C) dropped from 30% in the control group to 8%, a significant reduction.

[0170] 3. Cleaning efficiency and path optimization: By optimizing the path length and laser power distribution, the experimental group reduced the average path length by about 20% and improved the cleaning efficiency by more than 40% (from 0.42cm 2 / s increased to 0.59cm 2 In complex areas (category C), the method of the present invention effectively avoids the path redundancy problem of traditional methods by dynamically adjusting the path and power allocation.

[0171] 4. Reduced energy consumption and economy: The average laser energy consumption of the experimental group was significantly lower than that of the control group (0.12kWh vs. 0.17kWh), indicating that while improving the cleaning effect, it also reduced the waste of laser energy, and had higher economy and environmental protection.

[0172] In summary, the method of the present invention can achieve real-time early warning during the laser cleaning process of semiconductor devices, ensuring consistency with the target characteristic values ​​during the cleaning process. Moreover, through dynamic segmentation strategies and path planning, repeated cleaning and energy waste can be avoided. It can flexibly handle complex surfaces, achieve refined cleaning, and improve overall efficiency.

[0173] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A semiconductor device cleaning path planning and early warning method, characterized in that: The method comprises the following steps: Step 1: obtaining surface temperature, reflectivity and spatial depth information of the semiconductor device to be cleaned, generating a comprehensive feature vector after normalization processing, and constructing a three-dimensional surface model based on the spatial depth information; at the same time, performing feature extraction on the surface image of the semiconductor device before cleaning to generate an initial image feature vector; Step 2: using the comprehensive feature vector and the three-dimensional surface model, segmenting the surface area of ​​the semiconductor device to generate n cleaning units; Step 3: Based on the surface area segmentation results of the semiconductor device, a path planning algorithm is used to generate a cleaning path in combination with the optimization goals of path length and laser energy usage; Step 4: Start cleaning and monitor the temperature, reflectivity, and spatial depth of the semiconductor device surface in real time, and dynamically adjust cleaning parameters and cleaning paths; Step 5: Using the matching results of the cleaned surface information and the target feature values, identify the uncleaned area and optimize the cleaning path until cleaning is completed; The segmentation in step 2 is specifically as follows: based on the comprehensive feature vector and the three-dimensional surface model, the surface of the semiconductor device is divided into n cleaning units by a dynamic segmentation algorithm, and the segmentation granularity is dynamically adjusted according to the complexity of the cleaning unit; The dynamic segmentation algorithm adopts a dynamic clustering segmentation method based on comprehensive feature vectors and spatial depth information. The segmentation granularity of the dynamic segmentation algorithm is f seg (s i ) is as follows: Among them, v c (s i ) is the comprehensive feature vector of the i-th cleaning unit; v c is the current cluster center feature vector; ||v(s i )-v c || 2 is the square of the Euclidean distance, which measures the similarity of features between units; It represents the sum of the absolute values ​​of spatial depth gradients, which is used to reflect the complexity of the surface of the cleaning area; γ is the weight factor dynamically adjusted according to the surface complexity; λ is the granularity adjustment weight; A i is the cleaning unit area; n is the number of cleaning units, n is a natural number; s i represents the i-th cleaning unit; T(s i ) is the dynamically adjusted segmentation granularity threshold.

2. A semiconductor device cleaning path planning and early warning method according to claim 1, characterized in that: The comprehensive feature vector in step 1 is obtained by the following method: first, the surface temperature, reflectivity and spatial depth information of the semiconductor device are collected, and time-space synchronization correction is performed, and then a multimodal feature matrix is ​​established; the data in the multimodal feature matrix is ​​normalized, and a weighted fusion algorithm is used to generate a comprehensive feature vector.

3. A semiconductor device cleaning path planning and early warning method according to claim 2, characterized in that: The surface temperature is obtained by collecting the surface temperature of the semiconductor device at a high frame rate using an infrared thermal imager, and smoothing the dynamic temperature changes using a time-series interpolation algorithm; the reflectivity is obtained by obtaining reflectivity data of multiple bands using a multispectral imaging device, and extracting key spectral features through principal component analysis; the spatial depth information includes single-point depth measurement values, point cloud data, and depth maps; the spatial depth information is obtained by combining a laser rangefinder and a structured light sensor.

4. A semiconductor device cleaning path planning and early warning method according to claim 2, characterized in that: The generation formula of the comprehensive feature vector v is: Where: T(x) represents the surface temperature value of a specific cleaning area; R(x) represents the principal component eigenvalue of the reflectivity; D(x) represents the eigenvalue of the spatial depth information; μ T ,μ R ,μ D are the mean values ​​of temperature, reflectivity and spatial depth information features respectively; σ T ,σ R ,σ D are the standard deviations of temperature, reflectivity and depth characteristics respectively; W = [w T ,w R ,w D ],w T ,w R ,w D is the weight vector of temperature, reflectivity and spatial depth information features, and satisfies w T +w R +w D =1.

5. The semiconductor device cleaning path planning and early warning method according to claim 1, characterized in that: The generation of the cleaning path in step 3 is specifically as follows: based on the segmented cleaning units, a cleaning path planning strategy is generated using a deep reinforcement learning algorithm, and the optimal cleaning path is determined through a path optimization module; a path planning objective function F is defined. path The optimal trade-off between path length and cleaning performance is expressed as follows: Among them, P j is the three-dimensional coordinate of the j-th path point; ||P j -P j+1 || represents the Euclidean distance between adjacent path points; E laser (x j ) represents the laser power distribution function; v p (x j ) is the comprehensive feature vector of the j-th path point; x j represents the jth path point; v target is the comprehensive feature vector of the target area; α1 and α2 are the weight coefficients of path length and cleaning performance respectively; m is the number of path points; P0 and P1 represent the starting point and end point of the cleaning path respectively.

6. A semiconductor device cleaning path planning and early warning method according to claim 1, characterized in that: The cleaning parameters and cleaning path are dynamically adjusted in step 4, specifically: during the cleaning process, the temperature, reflectivity and spatial depth information of the surface of the semiconductor device are monitored in real time, and the monitoring data are compared with the target characteristic value generated before cleaning; when the difference between the monitoring data and the target characteristic value exceeds a preset threshold, an early warning signal is triggered, and the cleaning parameters are dynamically adjusted, and the cleaning parameters include the power, frequency and pulse width of the laser, and the cleaning path is dynamically updated according to the cleaning progress.

7. The semiconductor device cleaning path planning and early warning method according to claim 1, characterized in that: The cleaning path is optimized in step 5, specifically: feature extraction is performed on the surface image after cleaning, and similarity is calculated between the feature vector of the image after cleaning and the feature vector of the initial image before cleaning to evaluate the cleaning effect; quantitative analysis is performed on the uncleaned area, and when the area of ​​the uncleaned area exceeds the tolerance range, an early warning signal is triggered and a local optimization path is started until the cleaning is completed.

8. A semiconductor device cleaning path planning and early warning system, characterized in that: The system is used to execute the semiconductor device cleaning path planning and early warning method according to any one of claims 1 to 7, the system comprising: a data acquisition module for acquiring surface temperature, reflectivity, and spatial depth information of the semiconductor device to be cleaned, and normalizing the information to generate a comprehensive feature vector; and simultaneously, constructing a three-dimensional surface model based on the spatial depth information; A region segmentation module, configured to segment the surface of the semiconductor device using a region segmentation algorithm based on the comprehensive feature vector and the three-dimensional surface model, and to divide the surface into n cleaning units using a dynamic segmentation strategy; A path planning module is used to generate a cleaning path based on the region segmentation result, using a path planning algorithm and combining optimization goals of path length and laser energy usage; The dynamic monitoring and parameter adjustment module is used to monitor the temperature, reflectivity, and spatial depth information of the semiconductor device surface in real time during the cleaning process, compare the monitoring data with the target characteristic value, and trigger an early warning signal when the difference exceeds the preset threshold, and dynamically adjust the cleaning parameters and update the cleaning path; The cleaning effect evaluation and optimization module is used to extract features from the surface image after cleaning, calculate the similarity between the feature vector of the cleaned image and the feature vector of the initial image before cleaning, and evaluate the cleaning effect; perform quantitative analysis on the uncleaned area, and when the area of ​​the uncleaned area exceeds the tolerance range, trigger an early warning signal and start the local optimization path until the cleaning is completed.

Images