Intelligent cutting method and system for semiconductor processing part based on path reconstruction

Abstract

The invention discloses an intelligent cutting method and system for a semiconductor machined part based on path reconstruction, and relates to the technical field of semiconductor manufacturing processes, and the method comprises the steps: monitoring a machined part image, and extracting the edge contour and internal feature point information; fitting and constructing a three-dimensional contour model; planning a cutting path based on the model and outputting initial parameters; and cutting according to the initial parameters and iteratively optimizing the path according to the feedback information until the cutting task is completed. The technical problems that in the semiconductor device manufacturing process, a traditional cutting mode is difficult to cope with dynamic interference of semiconductor materials, fine modeling and protection on the edges and the interiors of the machined parts are lacked, cutting precision is insufficient, and the machined parts are prone to being damaged and scrapped are solved, fine modeling on the edges and the interiors of the machined parts is achieved, and the cutting precision is improved. The cutting path is optimized by combining dynamic interference real-time feedback, the cutting precision is improved, the damage rejection rate of machined parts is reduced, and the technical effects of high precision and high reliability of semiconductor device manufacturing are met.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor manufacturing process technology, and in particular to a method and system for intelligent cutting of semiconductor workpieces based on path reconstruction. Background Art

[0002] In semiconductor device manufacturing, the quality of workpiece cutting is closely related to device performance, yield rate, and production efficiency. Existing technologies mostly use fixed path planning and open-loop control modes, which can complete basic cutting in a relatively stable environment. However, as semiconductor devices develop towards high precision and miniaturization, traditional technologies are gradually showing their limitations in complex scenarios. Traditional methods have difficulty avoiding the protection areas of key components because they do not finely model the edge contours and internal features of the workpieces. In addition, due to dynamic interference such as material thermal sensitivity and equipment vibration, the fixed path cannot be adaptively adjusted in real time, resulting in insufficient cutting accuracy and a high scrap rate of workpieces. In addition, due to the lack of real-time feedback in open-loop control, it is impossible to optimize the path based on thermal deformation, vibration, etc., making it difficult to meet the needs of semiconductor manufacturing for high-precision and high-reliability cutting processes. Summary of the Invention

[0003] The present application provides a method and system for intelligent cutting of semiconductor workpieces based on path reconstruction, which is used to solve the technical problems in the semiconductor device manufacturing process, that is, traditional cutting methods are difficult to cope with the dynamic interference of semiconductor materials, and lack refined modeling and protection of the edges and interior of the workpieces, resulting in insufficient cutting accuracy and easy damage and scrapping of the workpieces.

[0004] The first aspect of the present application provides a method for intelligent cutting of semiconductor workpieces based on path reconstruction, the method comprising: extracting edge features and internal features based on the workpiece image acquired through monitoring, and obtaining the edge contour of the workpiece and the internal feature point information of the workpiece; constructing a three-dimensional contour model based on the edge contour of the workpiece and the internal feature point information of the workpiece; planning the cutting path based on the three-dimensional contour model, and outputting the initial cutting path and initial cutting parameters; controlling the cutting of the workpiece according to the initial cutting path and initial cutting parameters, iteratively optimizing the path based on the cutting feedback information, and performing iterative cutting control until the cutting task is completed.

[0005] The second aspect of the present application provides an intelligent cutting system for semiconductor workpieces based on path reconstruction, the system comprising: a workpiece information acquisition module, for extracting edge features and internal features based on the workpiece image acquired through monitoring, and obtaining the edge contour of the workpiece and the internal feature point information of the workpiece; a three-dimensional model construction module, for fitting and constructing a three-dimensional contour model based on the edge contour of the workpiece and the internal feature point information of the workpiece; an initial path output module, for planning the cutting path based on the three-dimensional contour model, and outputting the initial cutting path and initial cutting parameters; a path iterative optimization module, for controlling the cutting of the workpiece according to the initial cutting path and initial cutting parameters, performing iterative path optimization according to the cutting feedback information, and performing iterative cutting control until the cutting task is completed.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application collects images of workpieces to extract edge features and internal feature point information, constructs a three-dimensional contour model, and plans the initial cutting path. Using a dynamic reconstruction module, the cutting path is optimized based on real-time feedback from thermal deformation, equipment vibration, and other factors. Through multi-objective fitness evaluation and feedforward neural network correction, intelligent iterative optimization of the cutting path is achieved. This solves the problem of traditional fixed-path cutting being unable to avoid key features and insufficient response to dynamic interference, precisely controls the cutting path and parameters, improves the cutting accuracy and reliability of semiconductor workpieces, and meets the requirements of high-precision manufacturing. This achieves the technical effect of finely modeling the edges and internal features of workpieces, optimizing the cutting path with real-time feedback from dynamic interference, improving cutting accuracy, and reducing the damage and scrap rate of workpieces, meeting the high-precision and high-reliability requirements of semiconductor device manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0008] Figure 1 This is a flow chart of a method for intelligent cutting of semiconductor workpieces based on path reconstruction provided in an embodiment of the present application.

[0009] Figure 2 This is a structural diagram of a semiconductor processing part intelligent cutting system based on path reconstruction provided in an embodiment of the present application.

[0010] Explanation of the accompanying symbols: workpiece information acquisition module 1, three-dimensional model construction module 2, initial path output module 3, path iterative optimization module 4. DETAILED DESCRIPTION

[0011] The present application provides a method and system for intelligent cutting of semiconductor workpieces based on path reconstruction, which is used to solve the technical problems in the semiconductor device manufacturing process, that is, traditional cutting methods are difficult to cope with the dynamic interference of semiconductor materials, and lack refined modeling and protection of the edges and interior of the workpieces, resulting in insufficient cutting accuracy and easy damage and scrapping of the workpieces.

[0012] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only some of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0013] It should be noted that the terms "first", "second", etc. in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or inherent to these processes, methods, products or devices.

[0014] Example 1, as Figure 1 As shown, a method for intelligent cutting of semiconductor workpieces based on path reconstruction is provided, wherein the method comprises: Step A100: extracting edge features and internal features of the workpiece based on the image acquired through monitoring, and obtaining information on the edge contour of the workpiece and internal feature points of the workpiece.

[0015] Specifically, a model containing edge and internal feature extraction branches is pre-trained, and the model is used to extract the edges and internal features of the workpiece image, and output edge contours and internal feature point information. The specific steps are described in detail in A110-A120.

[0016] Step A200: constructing a three-dimensional contour model based on the edge contour of the workpiece and the internal feature point information of the workpiece.

[0017] Optionally, first generate an initial model based on the triangular mesh and the edge contour and internal feature point information of the workpiece through three-dimensional reconstruction, and then output the three-dimensional contour model through mesh simplification, hole repair and smoothing optimization. The specific steps are described in detail in A210-A220.

[0018] Step A300: performing cutting path planning based on the three-dimensional contour model, and outputting an initial cutting path and initial cutting parameters.

[0019] In one embodiment of the present application, key feature points and protection areas are identified and marked based on a three-dimensional contour model, and a first path is obtained by randomly planning the cutting path based on this constraint. The first cutting parameters are randomly generated and combined with the path generation scheme and the fitness is evaluated. The initial cutting path and parameters are output according to the fitness planning. The specific steps are described in detail in A310-A340.

[0020] Step A400: Controlling the cutting of the workpiece according to the initial cutting path and initial cutting parameters, iteratively optimizing the path according to cutting feedback information, and performing iterative cutting control until the cutting task is completed.

[0021] Specifically, the workpiece is first cut according to the planned initial cutting path and parameters using high-precision cutting equipment (such as laser cutters and precision dicing machines). During the cutting process, multiple sensors deployed in the processing area (such as infrared thermal imagers to monitor thermal deformation and accelerometers to collect equipment vibration data) are used to monitor and obtain cutting feedback information at fixed points with a sampling interval of 10ms-50ms. This information covers the workpiece's thermal deformation characteristics (such as the temperature gradient and thermal stress distribution in the cutting area) and the equipment's vibration characteristics (such as the vibration frequency and amplitude of the cutting head).

[0022] Next, fixed-point monitoring captures feedback about thermal deformation and equipment vibration during the workpiece cutting process. This feedback is used to iteratively correct the initial cutting plan, and then iteratively control the cutting process according to the corrected plan. The specific steps are detailed in A410-A420. This process of cutting-monitoring-correction-cutting is repeated, with each iteration improving the alignment of the cutting path with the actual processing requirements. After multiple iterations (the specific number depends on the workpiece's precision requirements and initial deviations), the workpiece's cutting task is completed.

[0023] By accurately collecting feedback data, intelligently correcting cutting plans and iteratively executing them, it effectively compensates for interference caused by material properties and equipment operation in semiconductor processing, allowing the cutting path to always adapt to the real-time status of the workpiece, thereby improving the cutting accuracy of semiconductor workpieces, reducing the cutting defect rate, and ensuring the quality and performance of the workpieces.

[0024] Furthermore, step A100 in the method provided in the embodiment of the present application includes: A110: Pre-training feature extraction model, wherein the feature extraction model includes an edge feature extraction branch and an internal feature extraction branch.

[0025] A120: Utilize the edge feature extraction branch and the internal feature extraction branch to extract edge features and internal features of the workpiece image, and output the edge contour of the workpiece and the internal feature point information of the workpiece.

[0026] In an embodiment of the present application, the feature extraction model is a pre-trained model including an edge feature extraction branch and an internal feature extraction branch.

[0027] Specifically, first, an image of the workpiece surface is acquired using industrial image acquisition equipment (such as a CCD camera with a resolution of ≥5 million pixels). The image covers the complete outline of the workpiece and the details of the internal area. The acquired image is input into a pre-trained dual-branch feature extraction model. The edge feature extraction branch of this model uses a lightweight convolutional neural network architecture to identify pixel-level edge information in the image. Through layer-by-layer convolution and pooling operations on the feature map, the coordinate data of the workpiece edge contour is output (edge ​​positioning error ≤ 0.01mm). The internal feature extraction branch, which runs simultaneously, targets internal structures such as key device areas and protection areas in the image, and realizes feature point positioning and region segmentation through a multi-task learning mechanism. The output is the coordinates of internal feature points and the range of the protection area (feature point positioning error ≤ 0.005mm, region segmentation accuracy ≥ 95%).

[0028] The model processes a single image in ≤50ms, meeting the real-time requirements of online inspection for semiconductor processing. Compared to traditional single-branch feature extraction techniques, such as the Canny algorithm, which only extracts edges, or the Halcon region detection tool, which only identifies internal structures, the dual-branch model can simultaneously extract both contours and internal features, avoiding the feature correlation errors caused by traditional step-by-step processing, improving feature extraction completeness, and reducing the rate of missed detection of key features.

[0029] Through the parallel processing of the workpiece image by the pre-trained dual-branch feature extraction model, efficient and accurate extraction of edge contours and internal feature point information is achieved, providing complete geometric data support for the subsequent three-dimensional contour model construction, solving the problems of incomplete information and low efficiency of traditional single-modal feature extraction methods, and laying a data foundation for protection area avoidance and precision optimization in cutting path planning.

[0030] Furthermore, step A110 in the method provided in the embodiment of the present application includes: A111: Based on the processing logs of similar workpieces, a sample workpiece image set, a sample edge contour set, and a sample internal feature distribution set are collected.

[0031] A112: Using a sample workpiece image set as input and a sample edge contour set as supervision, the convolutional neural network is trained until convergence to obtain an edge feature extraction branch.

[0032] A113: Using a sample workpiece image set as input and a sample internal feature distribution set as supervision, a convolutional neural network is trained until convergence to obtain an internal feature extraction branch, which is combined with the edge feature extraction branch to construct a feature extraction model.

[0033] Optionally, first, skilled practitioners can rely on historical processing logs of similar workpieces and use industrial-grade image acquisition equipment (e.g., a CCD camera with a resolution of 5 megapixels or higher) to acquire a sample set of workpiece images. Simultaneously, they can generate a sample edge profile set (containing the coordinate data of the workpiece's outer contour, with an accuracy of ±0.001mm) and a sample internal feature distribution set (recording the coordinates of key feature points and the extent of protected areas, with a positioning error of 0.005mm or less) based on manual annotation or high-precision inspection equipment (e.g., a coordinate measuring machine). The sample size of these three datasets must meet the requirements for machine learning training. Typically, the number of samples of similar workpieces should be 1,000 or more to cover the features of workpieces of different sizes and materials.

[0034] The training phase of the edge feature extraction branch then begins. Using a sample workpiece image set as input, a convolutional neural network (CNN) architecture (such as a lightweight variant of ResNet-18) is employed, with a sample edge contour set as the supervisory signal, to construct a pixel-level edge detection task. During training, a batch size of 32, an initial learning rate of 0.001, and the Adam optimizer are used, with a mean squared error (MSE) as the loss function. Training is repeated for 50-100 rounds until the loss converges to ≤0.01. At this point, the model achieves pixel-level accuracy in detecting workpiece edge contours, with an edge positioning error of ≤1 pixel, corresponding to an actual size error of ≤0.01mm. This output is then output to the edge feature extraction branch.

[0035] Next, the internal feature extraction branch is trained. Using the same CNN architecture, the input remains the sample workpiece image set, but the supervisory signal is switched to the sample internal feature distribution set. The task is defined as multi-task learning of keypoint localization and region segmentation. By adjusting the loss function weights, such as 60% for keypoint localization and 40% for region segmentation, and training for 80-120 rounds until the combined loss value is ≤0.02, the model's positioning error for internal feature points is ≤0.5 pixels, corresponding to an actual error of ≤0.005mm, and a protected area segmentation accuracy of ≥95%, the internal feature extraction branch is generated.

[0036] Finally, the edge feature extraction branch and the internal feature extraction branch are integrated in parallel within the network structure to form a dual-branch feature extraction model. This model can simultaneously process edge and internal features of workpiece images and meet the real-time requirements of semiconductor processing.

[0037] Through the above steps, the dual-branch feature extraction model constructed based on historical processing data realizes the refined extraction of the edge contour and internal key features of semiconductor processing parts, providing a high-precision data foundation for subsequent 3D modeling and cutting path planning, and effectively solving the cutting deviation problem caused by incomplete feature extraction in traditional methods.

[0038] Furthermore, step A200 in the method provided in the embodiment of the present application includes: A210: Based on the triangular mesh, three-dimensional reconstruction is performed according to the edge contour of the workpiece and the internal feature point information of the workpiece to generate an initial three-dimensional contour model.

[0039] A220: Perform mesh simplification, hole repair, and smooth optimization on the initial three-dimensional contour model, and output a three-dimensional contour model.

[0040] In the embodiments of this application, a triangular mesh refers to the geometric structure used for 3D reconstruction based on the edge contours and internal feature point information of the workpiece. This is achieved by topologically connecting 2D feature data (edge ​​contour vertices and internal feature points) in 3D space to generate an initial 3D structural framework composed of a large number of triangular facets. A 3D contour model is constructed by first performing 3D reconstruction using a triangular mesh algorithm based on the workpiece edge contours and internal feature point information to generate an initial model. This model is then output as a high-precision model after mesh simplification, hole repair, and smoothing optimization.

[0041] Specifically, in the construction of the three-dimensional contour model of the semiconductor workpiece, first, the workpiece edge contour coordinate data (accuracy ±0.01mm) and internal feature point coordinate information (positioning error ≤0.005mm) output from the aforementioned feature extraction step are used as input, and three-dimensional reconstruction is performed based on the triangular mesh algorithm in computer graphics (such as Delaunay triangulation). The vertices of the two-dimensional edge contour and the internal feature points are topologically connected in three-dimensional space, and by constructing non-overlapping triangular facets between the point sets, the circumcircle of each triangle does not contain other points, thereby generating an initial three-dimensional contour model composed of a large number of triangular facets. This model preliminarily restores the three-dimensional geometric structure of the workpiece, but may have problems such as mesh redundancy (the number of facets can reach thousands), local holes (such as unclosed areas caused by missing feature points), or surface roughness.

[0042] Next, to improve the computational efficiency and geometric accuracy of the model, the initial model needs to be mesh simplified. For mesh simplification, the error metric-based QEM (Quadric Error Metrics) algorithm is used. This algorithm calculates error metrics for each vertex and edge, identifies vertices and edges that contribute little to the model's contour features, and gradually deletes them. During the deletion process, the retention of key model contour features (such as inflection points on the edges of the workpiece and the neighborhood structure of internal feature points) is continuously evaluated to ensure that the contour error of the simplified model is controlled within a range of ≤±0.01mm. Taking an initial model containing 2,000 triangular facets as an example, after processing with the QEM algorithm, the number of facets can be reduced by 40%-60%, or to 800-1,200. This significantly reduces the computational complexity of the model while maintaining the integrity of key geometric features.

[0043] Hole repair is then performed using a curvature-guided Poisson reconstruction algorithm. First, the algorithm automatically scans the surface of the initial 3D contour model and detects hole areas with an area ≥ 0.01mm². For each hole, the algorithm generates an interpolated surface that matches the surrounding geometric structure based on the curvature distribution characteristics of the mesh surrounding the hole, and generates new triangular patches to fill the gap through triangulation. During the repair process, by calculating the normal vector continuity and vertex position deviation between the repaired area and the surrounding mesh, it ensures that the geometric continuity error between the repaired area and the original model is ≤ 0.005mm, ultimately achieving 100% surface closure of the model, eliminating geometric defects caused by missing feature points or reconstruction errors.

[0044] Finally, the model smoothing optimization phase begins, employing the Laplace smoothing algorithm to iteratively adjust the mesh vertex coordinates. The algorithm calculates the average position of each vertex's neighborhood and gradually shifts the vertex toward its neighborhood's center of gravity to reduce the roughness of the mesh surface. Typically, 5-10 iterations are performed, with each iteration recalculating the vertex neighborhood and updating the coordinates. This process improves the model surface smoothness to Ra ≤ 0.2 μm, effectively eliminating jagged edges or localized protrusions generated during triangular mesh reconstruction. This allows the model surface to more closely match the actual physical contours of the semiconductor workpiece, meeting the stringent modeling accuracy requirements of precision machining.

[0045] After the above processing, the output 3D contour model is a triangular mesh data that includes the complete edge contour of the workpiece, the spatial distribution of internal feature points, and smooth surfaces. Its geometric accuracy can reach ±0.005mm, and the number of facets is controlled between 500 and 1500, balancing computational efficiency and accuracy requirements. This model provides precise 3D spatial constraints for subsequent cutting path planning, allowing the cutting path to accurately adapt to the three-dimensional structure of the workpiece.

[0046] Through three-dimensional reconstruction and multi-stage optimization processing based on triangular meshes, two-dimensional feature data is converted into a high-precision three-dimensional model, which solves the problem that traditional two-dimensional modeling cannot reflect the spatial structure of the workpiece. It provides a reliable geometric basis for avoiding internal key feature points and optimizing spatial trajectories in cutting path planning, and ultimately improves the matching degree between the cutting path and the actual structure of the workpiece.

[0047] Furthermore, step A300 in the method provided in the embodiment of the present application includes: A310: Based on the three-dimensional contour model, identify and mark internal key feature points and protection areas.

[0048] A320: Perform random planning of the cutting path with avoiding the internal key feature points and the protection area as a constraint to obtain a first cutting path.

[0049] A330: Randomly generate first cutting parameters in the cutting parameter space of the cutting device, generate a first cutting plan based on the first cutting path, and evaluate and determine a first cutting fitness.

[0050] A340: Perform cutting path planning according to the first cutting fitness, and output an initial cutting path and initial cutting parameters.

[0051] In this embodiment, cutting fitness is determined by constructing a cutting fitness evaluation function and quantitatively evaluating the simulation results of the cutting plan, with the comprehensive optimization objectives of minimizing cutting time, cutting path length, cutting head idle distance, and maximizing cutting accuracy. Initial cutting parameters are the parameters output when planning the cutting path based on the 3D contour model, including cutting depth, cutting width, feed speed, etc.

[0052] Specifically, when analyzing the 3D contour model using computer vision algorithms, the Canny edge detection algorithm is first used to extract edges from the model's 2D projection image to identify the contour boundaries of the workpiece. Subsequently, the ORB feature point matching algorithm (Oriented FAST and Rotated BRIEF) is used to detect key feature points within the edge region, such as chip solder joints and circuit connection points. By calculating the orientation and descriptors of these feature points, critical internal structures are precisely located (with an accuracy of ±0.005mm). For protected areas, such as sensitive device packaging areas, deep learning-based semantic segmentation algorithms (such as U-Net) are used to classify the model surface and automatically identify the boundaries of areas requiring protection. Finally, a coordinate mapping algorithm is used to annotate the detected key feature points and protected areas into the 3D model space, ensuring their spatial positions are fully aligned with the actual workpiece. This process, through the fusion of multiple algorithms, achieves automated analysis of the entire process, from edge detection, feature point location, to region segmentation, providing highly accurate geometric constraints for cutting path planning.

[0053] Next, a random sampling tree algorithm is used to generate the first cutting path, using hard constraints such as avoiding marked key feature points and protected areas. The algorithm first sets a starting point and a target point within the cuttable area of ​​the 3D model. Nodes are generated through random sampling, and the tree structure is expanded from the starting point toward the random node. Before each expansion, the algorithm checks in real time whether the new node falls within the protected area or is close to a key feature point (distance threshold ≤ 0.01 mm). If the constraint is violated, the node is discarded and resampled. If the constraint is satisfied, the node is added to the tree structure, and a rewiring mechanism optimizes the tree edges, connecting the new node to a closer parent node in the tree to shorten the path length. Through repeated iterative sampling and expansion, typically 50-100 times, multiple candidate paths from the starting point to the target point are generated in the tree structure. A path that meets the constraints is randomly selected as the first cutting path. This path generation process ensures that critical internal structures within the workpiece are avoided. The random sampling and optimization mechanism also covers a wide range of possible cuttable areas. The algorithm generates 50-100 random paths per planning cycle, covering different cutting starting points and directions, ensuring diversity in the path space. At the same time, within the parameter space of the cutting equipment, such as cutting depth 0.1-1.5 mm, cutting width 0.05-0.3 mm, and feed speed 50-200 mm / s, a first cutting parameter combination is randomly generated and combined with the first cutting path to form a complete first cutting plan.

[0054] Then, the first cutting scheme is simulated based on the three-dimensional contour model, and a fitness evaluation function is constructed with the goals of minimizing cutting time, path length, idle distance and maximizing accuracy. The first cutting fitness is determined based on this evaluation. The specific steps are described in detail in A331-A333.

[0055] Finally, a new plan is constructed by continuously randomly planning the cutting path and parameters. After simulation and fitness evaluation, the preset number of convergences is reached, and the initial cutting plan with maximum fitness (including the initial cutting path and parameters) is output. The specific steps are described in detail in A341-A344.

[0056] By combining the geometric constraints of the three-dimensional model with the stochastic optimization algorithm, this method achieves global optimization of the cutting path and parameters while avoiding the key internal structures of the workpiece, thereby improving the average cutting accuracy of the initial cutting plan and effectively improving the reliability and production efficiency of semiconductor workpiece cutting.

[0057] Furthermore, step A330 in the method provided in the embodiment of the present application includes: A331: Based on the three-dimensional contour model, perform cutting simulation according to the first cutting plan and output a first cutting simulation result.

[0058] A332: Construct a cutting fitness evaluation function with the comprehensive optimization objectives of minimizing cutting time, cutting path length, cutting head idle distance and maximizing cutting accuracy.

[0059] A333: Using the cutting fitness evaluation function, determine a first cutting fitness based on the first cutting simulation result.

[0060] Specifically, when evaluating and determining the suitability of the first cut, the first cut plan is first physically simulated using finite element simulation software (such as ANSYS) based on a 3D contour model. By inputting cutting path coordinates and cutting parameters (such as cutting depth, cutting width, and feed rate), the cutting head motion trajectory and the stress and deformation of the workpiece are simulated. The first cut simulation results are output, including cutting time (unit: seconds), cutting path length (unit: mm), cutting head travel distance (unit: mm), and cutting accuracy (unit: ±μm). This process accurately simulates the thermal response and mechanical stress distribution of semiconductor materials.

[0061] Then, based on the multi-objective optimization theory, a cutting fitness evaluation function is constructed. The function aims to minimize the cutting time, cutting path length, cutting head idle distance and maximize cutting accuracy, and uses a linear weighting method to normalize the multi-dimensional indicators. For example, the weights of cutting time, cutting path length, idle distance and cutting accuracy are set to 30%, 25%, 20% and 25%, respectively, to construct a function of the form The function (where 、 、 is the preset ideal value, T, L, D are the actual simulation values, A is the actual accuracy value, is the target accuracy value), the function output value range is 0-1, and the larger the value, the higher the fitness.

[0062] Finally, the first cutting simulation result is substituted into the evaluation function to calculate the first cutting fitness value. This numerical value is used to quantitatively evaluate the comprehensive performance of the cutting scheme and provide data support for subsequent path optimization. For example, when the cutting time of a certain scheme is 20s, the path length is 15mm, the idle distance is 5mm, and the accuracy is ±10μm, the corresponding ideal values ​​are 25s, 20mm, 8mm, and ±15μm respectively. , indicating that the overall performance of this scheme is better.

[0063] By combining cutting simulation with multi-objective fitness evaluation, a quantitative evaluation of the initial cutting plan is achieved, avoiding the subjectivity of traditional empirical planning, and achieving a dynamic balance between efficiency (time, path length) and accuracy in the cutting plan. Ultimately, the success rate of the initial cutting plan for semiconductor processing parts is improved, thereby reducing trial and error costs and improving the reliability of the cutting process.

[0064] Furthermore, step A340 in the method provided in the embodiment of the present application includes: A341: Continue to perform random planning of the cutting path to obtain a second cutting path, and randomly generate second cutting parameters in the cutting parameter space of the cutting equipment to construct a second cutting plan.

[0065] A342: Perform cutting simulation according to the second cutting plan and output the second cutting simulation result.

[0066] A343: Using the cutting fitness evaluation function, determine the second cutting fitness according to the second cutting simulation result.

[0067] A344: Continue random planning of cutting schemes and cutting fitness evaluation until a preset number of convergences is reached, and output the cutting scheme with the maximum cutting fitness as the initial cutting scheme, wherein the initial cutting scheme includes an initial cutting path and initial cutting parameters.

[0068] In one embodiment, when planning a cutting path based on the first cutting fitness, an iterative optimization process is first initiated based on the evaluation results of the first cutting solution. A second cutting path is generated within the cuttable region of the three-dimensional contour model using the random sampling tree algorithm described in the aforementioned steps. The algorithmic process is identical to step A320 and will not be further described due to space limitations. Simultaneously, a second set of cutting parameters is randomly generated within the cutting equipment parameter space (e.g., cutting depth 0.1-1.5 mm, feed speed 50-200 mm / s), combining to form a second cutting solution.

[0069] Subsequently, using the same finite element simulation environment as the first cutting solution, the second cutting solution was simulated physically. Data such as cutting time, path length, idle distance, and cutting accuracy were collected, and the second cutting simulation results were output. For example, when the second cutting path length is 18mm, the cutting time is 22s, the idle distance is 6mm, and the accuracy is ±12μm, the simulation results will reflect the actual processing performance of this solution.

[0070] Next, the second-cut simulation results were quantitatively evaluated using a cutting fitness evaluation function. The function automatically calculated and normalized the weighted values ​​of each indicator to obtain the second-cut fitness value. Assuming that the cutting accuracy weight in the fitness function is 30%, if the second solution's accuracy is better than the first, its fitness value may increase from 0.82 to 0.88.

[0071] The above random programming, simulation, and fitness evaluation steps are repeated until a preset number of convergences is reached (typically 50-100 times, determined by those skilled in the art based on actual conditions). During this iteration, the system automatically records the fitness values ​​of all solutions and selects the solution with the highest fitness as the initial cutting solution. For example, after 80 iterations, the fitness value of a solution reaches 0.95, significantly outperforming the initial random solution.

[0072] Through multiple iterations of a stochastic optimization algorithm, combined with a closed-loop feedback mechanism using simulation and fitness evaluation, a global optimization of cutting paths and parameters is achieved. This method improves the overall performance of the initial cutting plan for semiconductor workpieces, narrows the fluctuation range of cutting accuracy, and reduces air movement distance. It effectively solves the imbalance between efficiency and accuracy caused by traditional fixed parameter planning and provides a data-driven optimization strategy for high-precision semiconductor cutting.

[0073] Furthermore, step A400 in the method provided in the embodiment of the present application includes: A410: During the workpiece cutting process, fixed-point monitoring is performed to obtain cutting feedback information, wherein the cutting feedback information includes at least thermal deformation characteristics and equipment vibration characteristics.

[0074] A420: Iteratively correct the initial cutting plan according to the thermal deformation characteristics and equipment vibration characteristics, and perform iterative cutting control according to the corrected cutting plan.

[0075] In the embodiments of this application, thermal deformation characteristics refer to deformation data related to temperature changes acquired through fixed-point monitoring during the cutting process of a workpiece. These characteristics, such as edge expansion, contraction, or shape shift caused by the thermal effects of cutting, are important components of cutting feedback information. Equipment vibration characteristics refer to equipment vibration-related data acquired through fixed-point monitoring during the cutting process. These include parameters such as vibration frequency, amplitude, and vibration direction. These characteristics reflect the dynamic stability of the cutting equipment and are a key component of cutting feedback information.

[0076] Optionally, during the cutting process, multiple sensors deployed in the processing area provide real-time monitoring at specific points: an infrared thermal imager (accuracy of ±0.5°C) collects the temperature distribution in the cutting area at a 20Hz frequency, extracting thermal deformation characteristics such as edge offset caused by thermal expansion, with a resolution of 0.01mm; a triaxial accelerometer (sampling rate of 1kHz, accuracy of ±0.001g) captures the vibration signals of the cutting equipment and analyzes the vibration frequency (range: 10-500Hz) and amplitude (accuracy of ±0.005mm). Cutting feedback from these two types of sensors is synchronously collected and fused into a multidimensional feedback vector based on timestamps, which is then input into the solution corrector for processing.

[0077] Iterative correction of the initial cutting plan includes building a plan corrector based on a feedforward neural network and collecting sample data for training until convergence, and then using the corrector to perform iterative correction based on thermal deformation and equipment vibration characteristics. The specific steps are described in detail in A421-A422.

[0078] Furthermore, step A420 in the method provided in the embodiment of the present application includes: A421: Build a solution corrector based on a feedforward neural network and collect sample data for training until convergence.

[0079] A422: Utilizing the scheme corrector, the initial cutting scheme is iteratively corrected according to the thermal deformation characteristics and the equipment vibration characteristics.

[0080] In an embodiment of the present application, a feedforward neural network is a neural network architecture based on a multi-layer perceptron, which is used to construct a solution corrector.

[0081] In one embodiment, when constructing a solution corrector, a model framework is first built based on a feedforward neural network (FFNN) architecture. The input layer corresponds to thermal deformation characteristics (such as thermal expansion offset of the workpiece edge, with a resolution of 0.01mm) and equipment vibration characteristics (such as vibration frequency of 10-500Hz and amplitude of ±0.005mm). The hidden layer uses the Reluctant Unit (ReLU) activation function to achieve nonlinear feature mapping. The output layer contains cutting path adjustment parameters (X / Y / Z axis coordinate correction values) and equipment parameter compensation values ​​(such as laser power adjustment and feed speed adjustment). Subsequently, sample data from the historical cutting process (including multiple pairs of thermal deformation-vibration-correction parameter samples) is collected. The network is trained using a stochastic gradient descent algorithm with the mean square error (MSE) as the loss function until the loss converges (training error ≤ 3%), enabling the model to predict correction parameters based on the input characteristics.

[0082] During real-time iterative correction, workpiece thermal deformation characteristics and equipment vibration data are collected in real time using an infrared thermal imager (accuracy of ±0.5°C) and a triaxial accelerometer (sampling rate of 1kHz, accuracy of ±0.001g). After normalization, these data are input into a trained solution corrector. The model outputs path adjustment values ​​(accuracy of ±0.001mm) and equipment parameter compensation values ​​(such as feed speed adjustment of ±2mm / s and laser power adjustment of ±5%), which directly drive the cutting system actuators to dynamically correct the initial cutting path and parameters. After each correction, the system automatically records the new feedback data and updates the model input, forming a closed loop of monitoring, prediction, and correction. Typically, after 3-5 iterations, cutting deviations caused by thermal deformation can be reduced from ±0.05mm to within ±0.01mm, and trajectory fluctuations caused by vibration can be reduced from ±0.03mm to ±0.005mm.

[0083] The solution corrector constructed through a feedforward neural network, combined with real-time multi-source feedback data, realizes dynamic compensation for thermal deformation and equipment vibration during the cutting process of semiconductor workpieces, improves cutting accuracy compared to traditional open-loop control solutions, effectively solves the problem of precision drift caused by dynamic interference during the processing process, and significantly improves the yield and processing consistency of high-precision semiconductor devices.

[0084] In summary, the semiconductor workpiece intelligent cutting method based on path reconstruction provided by the embodiments of the present application has the following technical effects: This application collects images of the edge and internal area of ​​the workpiece, obtains edge contour and internal feature point information through the edge feature extraction branch and internal feature extraction branch of the feature extraction model, performs three-dimensional reconstruction and optimizes the model based on the triangular mesh, identifies key feature points and protection areas in combination with the three-dimensional contour model, and randomly plans the cutting path to avoid constraints. The cutting plan is generated and the fitness is calculated through simulation and fitness evaluation function. The path and parameters are iteratively planned according to the fitness. At the same time, feedback information such as thermal deformation and equipment vibration is monitored at fixed points during the cutting process. The scheme corrector constructed using a feedforward neural network is used to iteratively correct the initial cutting plan, thereby accurately planning the cutting path and parameters of the semiconductor workpiece, significantly improving the cutting accuracy of the semiconductor workpiece and making the cutting process stable and reliable. The application achieves the technical effect of fine modeling of the edge and internal features of the workpiece, optimizing the cutting path in combination with real-time feedback of dynamic interference, improving cutting accuracy, reducing the damage and scrap rate of the workpiece, and meeting the high-precision and high-reliability requirements of semiconductor device manufacturing.

[0085] Example 2, as Figure 2 As shown, based on the same inventive concept as the aforementioned embodiment 1, the embodiment of the present application provides a semiconductor processing part intelligent cutting system based on path reconstruction, the system comprising: The workpiece information acquisition module 1 is used to extract edge features and internal features based on the workpiece image acquired through monitoring, and obtain the edge contour of the workpiece and the internal feature point information of the workpiece.

[0086] The three-dimensional model construction module 2 is used to fit and construct a three-dimensional contour model based on the edge contour of the workpiece and the internal feature point information of the workpiece.

[0087] The initial path output module 3 performs cutting path planning based on the three-dimensional contour model and outputs an initial cutting path and initial cutting parameters.

[0088] The path iteration optimization module 4 is used to control the cutting of the workpiece according to the initial cutting path and initial cutting parameters, perform path iteration optimization according to cutting feedback information, and perform iterative cutting control until the cutting task is completed.

[0089] Furthermore, the workpiece information acquisition module 1 is configured to perform the following steps: A pre-trained feature extraction model, wherein the feature extraction model includes an edge feature extraction branch and an internal feature extraction branch; using the edge feature extraction branch and the internal feature extraction branch, edge features and internal features of the workpiece image are extracted, and the edge contour of the workpiece and the internal feature point information of the workpiece are output.

[0090] Furthermore, the workpiece information acquisition module 1 is configured to perform the following steps: According to the processing logs of similar workpieces, a sample workpiece image set, a sample edge contour set and a sample internal feature distribution set are collected; with the sample workpiece image set as input and the sample edge contour set as supervision, a convolutional neural network is trained until convergence to obtain an edge feature extraction branch; with the sample workpiece image set as input and the sample internal feature distribution set as supervision, a convolutional neural network is trained until convergence to obtain an internal feature extraction branch, and a feature extraction model is constructed in combination with the edge feature extraction branch.

[0091] Furthermore, the three-dimensional model building module 2 is used to perform the following steps: Based on the triangular mesh, three-dimensional reconstruction is performed according to the edge contour of the workpiece and the internal feature point information of the workpiece to generate an initial three-dimensional contour model; the initial three-dimensional contour model is mesh simplified, hole repaired and smoothed to output a three-dimensional contour model.

[0092] Furthermore, the initial path output module 3 is configured to perform the following steps: Based on the three-dimensional contour model, internal key feature points and protection areas are identified and marked; with avoiding the internal key feature points and protection areas as constraints, random cutting path planning is performed to obtain a first cutting path; first cutting parameters are randomly generated in the cutting parameter space of the cutting equipment, a first cutting plan is generated in combination with the first cutting path, and a first cutting fitness is evaluated and determined; cutting path planning is performed according to the first cutting fitness, and an initial cutting path and initial cutting parameters are output.

[0093] Furthermore, the initial path output module 3 is configured to perform the following steps: Based on the three-dimensional contour model, a cutting simulation is performed according to the first cutting scheme, and a first cutting simulation result is output; a cutting fitness evaluation function is constructed with the comprehensive optimization goals of minimizing the cutting time, cutting path length, cutting head idle distance and maximizing cutting accuracy; and the cutting fitness evaluation function is used to evaluate and determine the first cutting fitness based on the first cutting simulation result.

[0094] Furthermore, the initial path output module 3 is configured to perform the following steps: Continue to perform random planning of the cutting path to obtain a second cutting path, and randomly generate second cutting parameters in the cutting parameter space of the cutting equipment to construct a second cutting scheme; perform cutting simulation according to the second cutting scheme, and output a second cutting simulation result; use the cutting fitness evaluation function to determine the second cutting fitness based on the evaluation of the second cutting simulation result; continue to perform random planning of the cutting scheme and cutting fitness evaluation until the preset number of convergence times is reached, and output the cutting scheme with the maximum cutting fitness as the initial cutting scheme, wherein the initial cutting scheme includes an initial cutting path and initial cutting parameters.

[0095] Furthermore, the path iteration optimization module 4 is configured to perform the following steps: During the workpiece cutting process, fixed-point monitoring is performed to obtain cutting feedback information, wherein the cutting feedback information includes at least thermal deformation characteristics and equipment vibration characteristics; the initial cutting plan is iteratively corrected based on the thermal deformation characteristics and equipment vibration characteristics, and iterative cutting control is performed according to the corrected cutting plan.

[0096] Furthermore, the path iteration optimization module 4 is configured to perform the following steps: A scheme corrector is constructed based on a feedforward neural network, and sample data is collected for training until convergence; the scheme corrector is used to iteratively correct the initial cutting scheme according to the thermal deformation characteristics and equipment vibration characteristics.

[0097] An intelligent cutting system for semiconductor processing parts based on path reconstruction provided by an embodiment of the present invention can execute an intelligent cutting method for semiconductor processing parts based on path reconstruction provided by any embodiment of the present invention, and has functional modules and beneficial effects corresponding to the execution method.

[0098] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules may be used and run on the user terminal and / or server, and the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of distinguishing each other and are not used to limit the scope of protection of the present invention.

[0099] The above specific embodiments do not constitute a limitation to the scope of protection of this application. It should be understood by those skilled in the art that various modifications, combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent replacements and improvements made within the spirit and principles of this application should be included in the scope of protection of this application. In some cases, the actions or steps recorded in this application can be performed in an order different from that in the embodiments and can still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A method for intelligent cutting of semiconductor parts based on path reconstruction, characterized in that the method include: Extract edge features and internal features based on the workpiece image acquired through monitoring to obtain the edge contour and internal feature point information of the workpiece; Constructing a three-dimensional contour model based on the edge contour of the workpiece and the internal feature point information of the workpiece; Perform cutting path planning based on the three-dimensional contour model, and output an initial cutting path and initial cutting parameters; The workpiece is cut and controlled according to the initial cutting path and initial cutting parameters, the path is iteratively optimized according to the cutting feedback information, and iterative cutting control is performed until the cutting task is completed.

2. The method for intelligent cutting of semiconductor parts based on path reconstruction according to claim 1, characterized in that: Extract edge features and internal features based on the workpiece image acquired through monitoring to obtain the edge contour and internal feature point information of the workpiece, including: A pre-trained feature extraction model, wherein the feature extraction model includes an edge feature extraction branch and an internal feature extraction branch; The edge feature extraction branch and the internal feature extraction branch are used to extract edge features and internal features of the workpiece image, and output the edge contour of the workpiece and the internal feature point information of the workpiece.

3. The method for intelligent cutting of semiconductor parts based on path reconstruction according to claim 2, characterized in that: Pre-trained feature extraction models, including: According to the processing logs of similar workpieces, a sample workpiece image set, a sample edge contour set, and a sample internal feature distribution set are collected; Using the sample workpiece image set as input and the sample edge contour set as supervision, the convolutional neural network is trained until convergence to obtain the edge feature extraction branch; With the sample workpiece image set as input and the sample internal feature distribution set as supervision, the convolutional neural network is trained until convergence to obtain the internal feature extraction branch, and the feature extraction model is constructed in combination with the edge feature extraction branch.

4. The method for intelligent cutting of semiconductor parts based on path reconstruction according to claim 1, characterized in that: The three-dimensional contour model is constructed by fitting the edge contour of the workpiece and the internal feature point information of the workpiece, including: Based on the triangular mesh, three-dimensional reconstruction is performed according to the edge contour of the workpiece and the internal feature point information of the workpiece to generate an initial three-dimensional contour model; The initial three-dimensional contour model is mesh simplified, hole repaired and smoothed, and a three-dimensional contour model is output.

5. The method for intelligent cutting of semiconductor parts based on path reconstruction according to claim 1, characterized in that: Perform cutting path planning based on the three-dimensional contour model and output an initial cutting path and initial cutting parameters, including: Based on the three-dimensional contour model, identifying and marking internal key feature points and protection areas; Performing random planning of the cutting path with avoiding the internal key feature points and the protection area as a constraint to obtain a first cutting path; Randomly generating a first cutting parameter in a cutting parameter space of a cutting device, generating a first cutting plan based on the first cutting path, and evaluating and determining a first cutting fitness; Cutting path planning is performed according to the first cutting fitness, and an initial cutting path and initial cutting parameters are output.

6. The method for intelligent cutting of semiconductor parts based on path reconstruction according to claim 5, characterized in that: Assessment to determine first-cut fitness, including: Based on the three-dimensional contour model, performing a cutting simulation according to the first cutting scheme, and outputting a first cutting simulation result; With the comprehensive optimization goals of minimizing cutting time, cutting path length, cutting head idle distance and maximizing cutting accuracy, a cutting fitness evaluation function is constructed. The cutting fitness evaluation function is used to evaluate and determine the first cutting fitness according to the first cutting simulation result.

7. The method for intelligent cutting of semiconductor parts based on path reconstruction according to claim 6, characterized in that: Performing cutting path planning according to the first cutting fitness and outputting an initial cutting path and initial cutting parameters includes: Continuing to perform random planning of the cutting path to obtain a second cutting path, and randomly generating second cutting parameters within the cutting parameter space of the cutting device to construct a second cutting plan; Perform cutting simulation according to the second cutting plan and output a second cutting simulation result; Determining a second cutting fitness based on the second cutting simulation result by using the cutting fitness evaluation function; Continue to perform random planning of cutting schemes and cutting fitness evaluation until the preset number of convergences is reached, and output the cutting scheme with the maximum cutting fitness as the initial cutting scheme, wherein the initial cutting scheme includes an initial cutting path and initial cutting parameters.

8. The method for intelligent cutting of semiconductor parts based on path reconstruction according to claim 1, characterized in that: Iterative path optimization and iterative cutting control are performed based on cutting feedback information, including: During the workpiece cutting process, fixed-point monitoring is performed to obtain cutting feedback information, wherein the cutting feedback information includes at least thermal deformation characteristics and equipment vibration characteristics; The initial cutting plan is iteratively corrected according to the thermal deformation characteristics and the equipment vibration characteristics, and iterative cutting control is performed according to the corrected cutting plan.

9. The method for intelligent cutting of semiconductor parts based on path reconstruction according to claim 8, characterized in that: The initial cutting plan is iteratively corrected according to the thermal deformation characteristics and equipment vibration characteristics, including: Build a solution corrector based on a feedforward neural network and collect sample data for training until convergence; The plan corrector is used to iteratively correct the initial cutting plan according to the thermal deformation characteristics and the equipment vibration characteristics.

10. A semiconductor processing part intelligent cutting system based on path reconstruction, characterized in that: A system for implementing a semiconductor workpiece intelligent cutting method based on path reconstruction according to any one of claims 1 to 9, comprising: A workpiece information acquisition module is used to extract edge features and internal features based on the workpiece image acquired through monitoring, and obtain the edge contour of the workpiece and the internal feature point information of the workpiece; A three-dimensional model building module, used for fitting and building a three-dimensional contour model based on the edge contour of the workpiece and the internal feature point information of the workpiece; An initial path output module performs cutting path planning based on the three-dimensional contour model and outputs an initial cutting path and initial cutting parameters; The path iterative optimization module is used to control the cutting of the workpiece according to the initial cutting path and initial cutting parameters, perform path iterative optimization according to cutting feedback information, and perform iterative cutting control until the cutting task is completed.

Images