A semiconductor chip product automatic quality detection system and method

Abstract

This invention discloses an automated quality inspection system and method for semiconductor chip products, relating to the field of semiconductor chip quality inspection technology. The method includes: acquiring event stream data from the chip surface; constructing a spatiotemporal distribution feature vector; performing voxelization processing on the event stream data based on the spatiotemporal distribution feature vector to obtain a dynamic spatiotemporal voxel grid; inputting the dynamic spatiotemporal voxel grid into a deformation-aware 3D convolutional kernel for spatiotemporal feature decoupling to generate a decoupled spatiotemporal feature map; simultaneously generating a self-correcting geometric reference matrix through event trajectory analysis; and fusing the geometric reference matrix with the decoupled spatiotemporal feature map to generate a reference fused feature map. This invention dynamically adjusts the spatial resolution and temporal slice depth based on the density gradient magnitude and polarity transition probability, enabling the voxel grid to adaptively conform to the chip surface deformation state, ensuring the physical authenticity of feature extraction.

Description

Technical Field

[0001] This invention relates to the field of semiconductor chip quality inspection technology, and in particular to an automated quality inspection system and method for semiconductor chip products. Background Technology

[0002] In the field of semiconductor chip quality inspection, high-resolution optical imaging and image processing technologies are primarily relied upon. In recent years, with the widespread application of event cameras, detection schemes based on asynchronous event stream data have gradually developed. By capturing pixel-level light intensity changes, they achieve microsecond-level response times, significantly improving detection efficiency. These technologies include frame-based event accumulation and reconstruction, spiking neural network classifiers, and feature extraction methods combined with spatiotemporal filtering. These techniques have, to some extent, solved the data capture problem in high-speed scenarios.

[0003] However, existing technologies have significant limitations: the voxelization method uses a fixed spatiotemporal grid resolution, which cannot adapt to dynamic surface changes caused by chip thermal deformation, resulting in feature distortion or missed detection; convolutional neural networks lack the ability to model the thermal expansion characteristics of materials, and their fixed-structure convolutional kernels are difficult to accurately capture the spatiotemporal feature correlations under deformation, resulting in insufficient generalization ability in high-speed and high-precision detection scenarios. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an automated quality inspection method for semiconductor chip products to solve the problems of distortion in dynamic surface feature extraction under chip thermal deformation environment and weak spatiotemporal feature correlation in high-speed detection scenarios.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an automated quality inspection method for semiconductor chip products, comprising: acquiring event stream data on the chip surface, constructing a spatiotemporal distribution feature vector, and performing voxelization processing on the event stream data according to the spatiotemporal distribution feature vector to obtain a dynamic spatiotemporal voxel grid;

[0008] The dynamic spatiotemporal voxel grid is input into the deformation-aware 3D convolutional kernel to decouple the spatiotemporal features and generate a decoupled spatiotemporal feature map. At the same time, a self-correcting geometric reference matrix is ​​generated through event trajectory analysis. The geometric reference matrix is ​​fused with the decoupled spatiotemporal feature map to generate a reference fusion feature map.

[0009] A pulse firing modulation mechanism is established based on a pulse neural network. The pulse firing rate feature of the reference fusion feature map is extracted. The pulse firing rate feature is then used to perform pulse-frequency domain cross-modal coupling on the reference fusion feature map to generate a pulse frequency domain feature map.

[0010] Defect decisions are made on the pulse frequency domain feature map through neural plasticity rules and spatiotemporal consistency verification mechanisms, and iterative optimization is performed through pulse firing mode association conditions to generate a defect spatial distribution map.

[0011] A neuromorphic feedback model is established and used as a control criterion to correct the defect types in the defect spatial distribution map, and a structured quality inspection report is output.

[0012] As a preferred embodiment of the automated quality inspection method for semiconductor chip products described in this invention, the specific steps for constructing the spatiotemporal distribution feature vector are as follows:

[0013] The event stream data on the chip surface is timestamped and divided into regular spatiotemporal grids. The event density and polarity distribution values ​​within each grid are statistically analyzed.

[0014] Based on the event density value and polarity distribution value, calculate the density gradient change and polarity transition probability between adjacent spatiotemporal grids;

[0015] Based on the density gradient change and polarity transition probability, a spatiotemporal distribution feature vector is constructed.

[0016] As a preferred embodiment of the automated quality inspection method for semiconductor chip products according to the present invention, the specific steps for voxelizing the event stream data are as follows:

[0017] Based on the spatiotemporal distribution feature vector, determine the spatial resolution parameter and the temporal slice depth parameter;

[0018] Configure a 3D mesh coordinate system based on spatial resolution parameters and time slice depth parameters, map event stream data to the 3D mesh coordinate system, and assign voxel weights based on event density values;

[0019] Based on the assigned voxel weights, neighborhood interpolation compensation is performed on low-density voxel meshes in the 3D mesh coordinate system, while the interpolation coefficients are adjusted according to the polarity distribution values.

[0020] Nonlinear compression is performed on the high-density voxel mesh in the three-dimensional mesh coordinate system.

[0021] As a preferred embodiment of the automated quality inspection method for semiconductor chip products described in this invention, the step of decoupling the dynamic spatiotemporal voxel grid into a deformation-aware 3D convolutional kernel for spatiotemporal feature input includes the following steps.

[0022] Based on dynamic spatiotemporal voxel grids, construct deformable convolutional kernel groups;

[0023] Using the deformable convolution kernel group, adaptive receptive field convolution is performed in the spatial dimension, and gated recurrent convolution is performed in the temporal dimension;

[0024] Based on convolutional operations in spatial and temporal dimensions, spatial edge response features and temporal state evolution features are extracted and fused according to spatiotemporal location encoding to generate a decoupled spatiotemporal feature map.

[0025] As a preferred embodiment of the automated quality inspection method for semiconductor chip products described in this invention, the specific steps for generating a self-calibrating geometric reference matrix through event trajectory analysis are as follows:

[0026] Extract the set of continuous event trajectory points from the decoupled spatiotemporal feature map, fit it to obtain the motion trajectory plane equation, calculate the angle deviation between the trajectory plane normal vector and the chip reference coordinate system, and construct the rotation compensation sub-matrix.

[0027] Based on the event trajectory point set, the event density change rate is analyzed to generate scaling factor compensation coefficient;

[0028] By fusing the rotation compensation submatrix and the scaling factor compensation coefficient, a self-correcting geometric reference matrix is ​​generated.

[0029] As a preferred embodiment of the automated quality inspection method for semiconductor chip products according to the present invention, the specific steps for extracting the pulse firing rate feature of the reference fusion feature map are as follows:

[0030] The baseline fused feature map is input into the response layer of the spiking neural network, and the threshold parameters of the neurons in the spiking neural network are dynamically adjusted through the firing rate modulation mechanism.

[0031] Within a preset time window, the number of pulse triggers of neurons in the spiking neural network is counted, the pulse trigger frequency per unit time is calculated, a pulse firing rate heatmap is generated, and the firing rate features of high-response regions are extracted from it.

[0032] As a preferred embodiment of the automated quality inspection method for semiconductor chip products according to the present invention, the specific steps for performing pulse-frequency domain cross-modal coupling on the reference fused feature map are as follows:

[0033] Based on the extracted high-response region emission rate features, multi-scale time-frequency transformation is performed on the event stream data to generate a frequency domain energy distribution spectrum.

[0034] Spatial location coding information is extracted from the firing rate features of the high response region, a cross-modal attention weight matrix is ​​constructed, and the frequency domain energy distribution spectrum is combined with the attention weight matrix through matrix multiplication to redistribute the frequency domain energy distribution spectrum.

[0035] The redistributed frequency domain energy distribution spectrum and the firing rate feature map are cascaded to generate a pulse frequency domain feature map.

[0036] As a preferred embodiment of the automated quality inspection method for semiconductor chip products according to the present invention, the specific steps for generating the defect spatial distribution map are as follows:

[0037] High firing rate neuron clusters in the pulse frequency domain feature map are identified, and the Heb learning rule is applied to enhance the connection strength of defect-related neural pathways. At the same time, the pulse temporal-dependent plasticity mechanism is used to adjust the synaptic weight update direction of defect-related neural pathways to generate a defect probability distribution map.

[0038] Extract the probability values ​​of defects at the same spatial location from the defect probability distribution map in a continuous time slice sequence to construct a probability change curve;

[0039] The peak position and duration of the probability change curve are detected to confirm and mark defects, and a defect spatial distribution map is generated through multiple iterations and optimizations.

[0040] As a preferred embodiment of the automated quality inspection method for semiconductor chip products described in this invention, the specific steps for outputting a structured quality inspection report are as follows:

[0041] Based on the aforementioned defect spatial distribution map, a mapping relationship library between defect types and control commands is established, and initial control commands are generated.

[0042] The initial control command is input into the neuromorphic feedback model to simulate the state, and the difference between the simulated state and the actual state represented by the defect spatial distribution map is calculated.

[0043] The initial control command parameter strength is adjusted based on the difference value, and then sent to the actuator for defect correction, outputting a structured quality inspection report.

[0044] Secondly, this invention provides an automated quality inspection system for semiconductor chip products, comprising:

[0045] The event sensing module acquires event stream data from the chip surface, constructs a spatiotemporal distribution feature vector, and performs voxelization processing on the event stream data based on the spatiotemporal distribution feature vector to obtain a dynamic spatiotemporal voxel grid.

[0046] The feature decoupling module inputs a dynamic spatiotemporal voxel grid into a deformation-aware 3D convolutional kernel to decouple spatiotemporal features and generate a decoupled spatiotemporal feature map. At the same time, it generates a self-correcting geometric reference matrix through event trajectory parsing and fuses the geometric reference matrix with the decoupled spatiotemporal feature map to generate a reference fusion feature map.

[0047] The neural decision-making module establishes a pulse firing modulation mechanism based on a spiking neural network, extracts the pulse firing rate features of the benchmark fusion feature map, and performs pulse-frequency domain cross-modal coupling on the benchmark fusion feature map according to the pulse firing rate features to generate a pulse frequency domain feature map.

[0048] The defect optimization module performs defect decision-making on the pulse frequency domain feature map through neural plasticity rules and spatiotemporal consistency verification mechanism, and performs iterative optimization through pulse firing mode association conditions to generate a defect spatial distribution map.

[0049] The closed-loop control module establishes a neuromorphic feedback model, uses the neuromorphic feedback model as a control criterion, corrects the defect types in the defect spatial distribution map, and outputs a structured quality inspection report.

[0050] The beneficial effects of this invention are as follows: The spatial resolution and temporal slice depth are dynamically adjusted based on the density gradient amplitude and polarity transition probability, enabling the voxel grid to adaptively conform to the chip surface deformation state, thus ensuring the physical authenticity of feature extraction; the deformation parameters of the convolution kernel are dynamically adjusted by the material's thermal expansion coefficient, allowing the feature decoupling process to accurately match the actual physical deformation of the chip, significantly improving the spatiotemporal feature correlation; and a pulse-frequency domain cross-modal coupling mechanism is adopted to deeply integrate the pulse firing rate characteristics with the frequency domain energy distribution, enhancing the response intensity of micro-defects. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of an automated quality inspection method for semiconductor chip products.

[0053] Figure 2 This is a schematic diagram illustrating the decoupling of event perception and feature processing.

[0054] Figure 3 This is a schematic diagram of neural decision-making and defect detection.

[0055] Figure 4 This is a schematic diagram of closed-loop control and quality correction. Detailed Implementation

[0056] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0057] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0058] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0059] Reference Figures 1-4 This is one embodiment of the present invention, which provides an automated quality inspection method for semiconductor chip products, including the following steps:

[0060] S1. Obtain event stream data from the chip surface, construct a spatiotemporal distribution feature vector, and perform voxelization processing on the event stream data based on the spatiotemporal distribution feature vector to obtain a dynamic spatiotemporal voxel grid.

[0061] S1.1. Perform timestamp alignment on the event stream data on the chip surface and divide it into regular spatiotemporal grids, and statistically analyze the event density value and polarity distribution value in each grid;

[0062] Each event in the event stream data on the chip surface includes pixel coordinates, timestamp, and light intensity change polarity.

[0063] The timestamps of the event stream data are aligned using an interpolation algorithm based on hardware clock synchronization to generate a time-synchronized event sequence. The time-synchronized event sequence is divided into regular spatiotemporal grids, and the size and time window length of each grid are set. The event density and polarity distribution values ​​within each spatiotemporal grid are statistically analyzed.

[0064] Among them, the event density value is the number of events per unit area, and the polarity distribution value is the ratio of positive polarity events to negative polarity events.

[0065] S1.2. Based on the event density value and polarity distribution value, calculate the density gradient change and polarity transition probability between adjacent spatiotemporal grids, and construct the spatiotemporal distribution feature vector.

[0066] Specifically, the central difference method is used to calculate the first derivative of the density value in the spatial and temporal dimensions as the density gradient change. At the same time, the polarity transition probability is calculated using the conditional probability formula to describe the statistical law of polarity change between adjacent grids. A spatiotemporal distribution feature vector is constructed based on the density gradient change and the polarity transition probability.

[0067] The spatiotemporal distribution feature vector includes horizontal density gradient, vertical density gradient, temporal density gradient, positive polarity transition probability, negative polarity transition probability, and polarity retention probability.

[0068] Based on density gradient changes and polarity transition probabilities, a spatiotemporal distribution feature vector is constructed. Specifically, the three components of the density gradient vector (horizontal gradient, vertical gradient, and temporal gradient) are extracted as basic spatial features; the positive polarity transition probability, negative polarity transition probability, and polarity preservation probability in the polarity transition probability matrix are used as supplementary features to form polarity features; the basic spatial features and polarity features are combined into a six-dimensional feature vector, and the dimensional differences are eliminated through feature normalization to output a standardized spatiotemporal distribution feature vector.

[0069] It should be noted that the expressions for calculating the first derivative of the density value in the spatial and temporal dimensions are as follows:

[0070] ;

[0071] in, It is the density gradient vector, representing the rate of change of event density in the spatial and temporal dimensions. It is a horizontal gradient, reflecting Directional density change, It is the vertical gradient, reflecting Directional density change, It is a time gradient, reflecting the change of density over time. It is time. Event density represents the number of events per unit area and is used to quantify the event activity in a specific area on the chip surface.

[0072] The expression for calculating the polarity transition probability is:

[0073] ;

[0074] in, From polarity Transferred to The conditional probability, and This represents the event's polarity state, with values ​​of {positive polarity (+1), negative polarity (-1)}. It is the transition event count, i.e., the adjacent grid. arrive Number of transfers, It is the total transfer base, that is, from The sum of all transition events that start from the destination. This is the initial polarity state, representing the initial event polarity of the current spacetime grid. It represents the target polarity state, indicating the polarity of the target event after the transition in adjacent grids. It is a polarity state traversal index, which traverses all possible polarities in the summation (usually taking...). );

[0075] S1.3. Voxelize the event stream data based on the spatiotemporal distribution feature vector to obtain a dynamic spatiotemporal voxel grid.

[0076] Specifically, based on the constructed spatiotemporal distribution feature vector, the spatial resolution parameters and temporal slice depth parameters of the voxel grid are determined;

[0077] Among them, the spatial resolution parameter is dynamically adjusted according to the density gradient magnitude. For example, a high-gradient region uses a high resolution of 4×4 pixels, and a low-gradient region uses a low resolution of 16×16 pixels. The temporal slice depth parameter is determined according to the polarity transition probability. For example, a high transition probability region uses a depth window of 10 milliseconds, and a low transition probability region uses a depth window of 2 milliseconds.

[0078] Configure the 3D mesh coordinate system based on spatial resolution parameters and time slice depth parameters;

[0079] In this system, the origin of the three-dimensional grid coordinate system corresponds to the center of the chip and serves as the chip's reference coordinate system. The X and Y axes correspond to the length and width dimensions of the chip surface, while the Z axis corresponds to the time dimension.

[0080] The event stream data is mapped to a three-dimensional grid coordinate system. The nearest neighbor interpolation method is used to assign each event to the corresponding voxel grid. Weight coefficients are assigned to each voxel grid according to the local event density and the average event density in the event density value.

[0081] Neighborhood interpolation compensation is performed on low-density voxel grids, and bilinear interpolation is used to fill missing data. At the same time, the interpolation coefficients are adjusted according to the polarity distribution value, with positive polarity events assigned high interpolation weights and negative polarity events assigned low interpolation weights.

[0082] The high-density voxel mesh is subjected to nonlinear compression processing using a logarithmic compression function, and the output is a dynamic spatiotemporal voxel mesh with weighted coefficients. Each voxel contains four attributes: spatial coordinates, timestamp, weight coefficient, and polarity flag.

[0083] S2. Input the dynamic spatiotemporal voxel mesh into the deformation-aware 3D convolutional kernel to decouple the spatiotemporal features and generate a decoupled spatiotemporal feature map. At the same time, generate a self-correcting geometric reference matrix through event trajectory analysis. Fuse the geometric reference matrix with the decoupled spatiotemporal feature map to generate a reference fusion feature map.

[0084] S2.1. Input the dynamic spatiotemporal voxel grid into the deformation-aware 3D convolution kernel to decouple the spatiotemporal features and generate a decoupled spatiotemporal feature map.

[0085] Specifically, based on the thermal expansion characteristics of chip materials, a deformable convolution kernel group is constructed, wherein the deformation parameters of each convolution kernel are dynamically adjusted according to the material expansion coefficient at the current operating temperature; the convolution kernel has independent adaptive capabilities in both spatial and temporal dimensions, and can adjust its structural morphology in real time according to the physical properties of the chip surface.

[0086] Deformable convolution kernels are used to perform adaptive receptive field convolution operations in the spatial dimension. The coverage of the convolution kernels is dynamically adjusted according to the curvature characteristics of local regions in the dynamic spatiotemporal voxel grid. Small receptive fields are used to focus on detailed features in high curvature regions, while large receptive fields are used to capture macroscopic structures in low curvature regions. Spatial edge response features are extracted during the adaptive receptive field convolution operation to accurately identify the physical structural boundaries and contour changes on the chip surface.

[0087] A gated recurrent convolution operation is performed in the time dimension, and the spatial edge response features output in the spatial dimension are used as the current input features. Simultaneously, timestamp sequences of the corresponding regions are extracted from a dynamic spatiotemporal voxel grid, and the interval values ​​between adjacent timestamps are obtained. The current input features and timestamp interval values ​​are input into a forget gate control mechanism, and a historical state retention ratio is generated through feature weighting. This historical state retention ratio is applied to the historical state features to control the fusion weight with the current input features, resulting in a temporal state fusion feature. Based on the temporal state fusion feature, periodic peak events of the event stream waveform in continuous time slices are detected, and the timestamp sequence corresponding to each peak event is recorded. The algorithm calculates the time interval between adjacent peaks based on the timestamp sequence to form a peak interval sequence. Based on the statistical distribution characteristics of the peak interval sequence, it identifies the dominant interval pattern and its fluctuation range. Simultaneously, it extracts the envelope contour of the event stream waveform, detects the decay trend of the waveform amplitude over time, and quantifies the time length required for the waveform amplitude to decay from the peak to a stable state and the slope of the decay curve. Based on the peak interval law and waveform decay characteristics, it establishes a mapping relationship between the peak interval and the chip surface deformation rate. At the same time, based on the correlation characteristics between the decay curve slope and the thermal diffusion efficiency, it extracts the time state evolution characteristics characterizing the deformation rate and thermal diffusion trend.

[0088] Spatial edge response features and temporal state evolution features are fused according to spatiotemporal location encoding. The feature alignment mechanism ensures the correspondence between spatial and temporal dimensions at the same location. A weighted fusion strategy is used to generate a decoupled spatiotemporal feature map.

[0089] The weight ratio of spatial edge response characteristics to temporal state evolution characteristics is dynamically adjusted according to the characteristics of the local region.

[0090] S2.2. Generate a self-correcting geometric reference matrix through event trajectory analysis, and fuse the geometric reference matrix with the decoupled spatiotemporal feature map to generate a reference fusion feature map.

[0091] Specifically, a continuous set of event points describing the chip's motion trajectory is extracted from the decoupled spatiotemporal feature map; a random sampling consensus algorithm is used to fit the optimal motion trajectory plane equation through iterative sampling and outlier removal; the angle between the normal vector and the reference coordinate axis is quantized by vector dot product operation to obtain the spatial angle deviation value, and the spatial angle deviation value is converted into a rotation compensation sub-matrix; at the same time, the rate of change of event density over time is analyzed based on the event point set, and differential operation is performed by the central difference method to obtain the event density sequence of continuous time points, and the density change between adjacent time points is calculated;

[0092] The first derivative of the density change is taken, and the density change rate is output as the scaling factor compensation coefficient.

[0093] By fusing the rotation compensation submatrix and the scaling factor compensation coefficients, a six-DOF self-calibrating geometric reference matrix is ​​generated, which contains spatial compensation parameters in three dimensions: rotation, translation, and scaling. This matrix is ​​used to correct detection deviations caused by changes in chip pose in real time. The spatial compensation parameters in the self-calibrating geometric reference matrix are mapped onto the decoupled spatiotemporal feature map through affine transformation. The pose of the decoupled spatiotemporal feature map is then corrected to generate a reference fusion feature map. This reference fusion feature map contains spatiotemporal feature information and pose correction information of the chip surface.

[0094] S3. Establish a pulse firing modulation mechanism based on a pulse neural network, extract the pulse firing rate features of the benchmark fusion feature map, and perform pulse-frequency domain cross-modal coupling on the benchmark fusion feature map according to the pulse firing rate features to generate a pulse frequency domain feature map.

[0095] S3.1. Establish a pulse firing modulation mechanism based on a spiking neural network and extract the pulse firing rate features from the benchmark fusion feature map.

[0096] Specifically, the baseline fused feature map is input into the response layer of the spiking neural network, and the activation threshold of neurons in the spiking neural network is dynamically adjusted through the firing rate modulation mechanism.

[0097] Among them, the firing rate modulation mechanism automatically reduces the activation threshold of high-response regions and increases the activation threshold of low-response regions based on the intensity distribution of the baseline fused feature map.

[0098] It should be noted that the activation threshold is generated through dynamic adjustment of the spiking neural network, with a value range of [0.2, 0.8]. It is determined based on the percentile distribution of feature intensity statistically from historical data, ensuring that the neuron response can both fully capture highly active features and effectively suppress background noise.

[0099] The pulse triggering state of neurons is monitored in real time within a preset time window, and the pulse triggering frequency at each spatial location per unit time is statistically analyzed. A pulse firing rate heatmap is generated based on the pulse triggering frequency per unit time. The pulse firing rate heatmap visually represents the neuronal response intensity in different regions of the chip surface in the form of a two-dimensional matrix. The firing rate features of high-response regions are extracted from the pulse firing rate heatmap. The firing rate features of high-response regions include spatial location encoding and response intensity values.

[0100] It should be noted that the preset time window is determined based on the chip production line's transmission cycle time and the minimum periodicity of the event flow, and the window length is dynamically adjusted based on real-time throughput data.

[0101] S3.2. Based on the pulse firing rate characteristics, the baseline fusion feature map is coupled across modes in the pulse-frequency domain to generate a pulse frequency domain feature map.

[0102] Specifically, the event stream data on the chip surface is segmented into time windows, and the Hanning window function is used to suppress spectral leakage. Fourier transform is applied to different time windows to extract the energy distribution in the corresponding frequency bands, generating a three-dimensional frequency domain energy distribution spectrum. The three-dimensional frequency domain energy distribution spectrum corresponds to the frequency band, spatial location, and time slice, respectively. The three-dimensional frequency domain energy distribution spectrum completely records the frequency characteristics and energy propagation path of the chip surface vibration.

[0103] Spatial location coding information is extracted from the emission rate characteristics of high-response regions to construct a cross-modal attention weight matrix. The row dimension of the cross-modal attention weight matrix corresponds to the frequency domain energy band division (low frequency / mid frequency / high frequency), and the column dimension corresponds to the chip surface spatial partition (core area / edge area / pad area). By performing matrix multiplication operation between the cross-modal attention weight matrix and the three-dimensional frequency domain energy distribution spectrum, the spatial redistribution of frequency domain energy is realized, so that high frequency energy is concentrated in high-response regions and low frequency energy diffuses to the structural edge, thereby enhancing the frequency domain characteristic response of defect-sensitive regions.

[0104] The redistributed frequency domain energy distribution spectrum and the pulse firing rate feature map are channel-concatenated, and the redistributed frequency domain energy spectrum is upsampled to improve the spatial resolution and firing rate feature map. Figure 1 To achieve this, coordinate deviations are corrected using a spatial transformation network, and the resulting pulse frequency domain feature map is generated through fusion.

[0105] The pulse frequency domain feature map contains three-dimensional data channels. The first channel retains the original pulse emission characteristics, the second channel carries the redistributed frequency domain energy, and the third channel stores cross-modal correlation weights.

[0106] S4. Defect decisions are made on the pulse frequency domain feature map through neural plasticity rules and spatiotemporal consistency verification mechanisms, and iterative optimization is performed through pulse firing mode association conditions to generate a defect spatial distribution map.

[0107] S4.1. Identify clusters of high firing rate neurons in the pulse frequency domain feature map, and apply the Heb learning rule to enhance the connection strength of defect-related neural pathways. At the same time, use the pulse temporal-dependent plasticity mechanism to adjust the synaptic weight update direction of defect-related neural pathways to generate a defect probability distribution map.

[0108] Specifically, neuronal clusters with significantly higher-than-average pulse firing rates in the pulse frequency domain feature map are identified as high-firing-rate neuronal clusters. The Heb learning rule is applied to enhance the synaptic connection strength of the corresponding neural pathways when the spatial distribution of high-firing-rate neuronal clusters highly overlaps with the spatial distribution of known defect patterns (such as microcracks and solder joint offsets).

[0109] Based on the enhanced neural pathways, a pulse-dependent plasticity mechanism is used to adjust the synaptic weight update direction: the synaptic connections for defect feature propagation are enhanced by positive temporal dependence, while the synaptic connections for noise interference are suppressed by reverse temporal dependence. The adjustment results of the two temporal dependence mechanisms are combined to generate a defect probability distribution map, which quantifies the probability of defect occurrence at each location on the chip surface in the form of a thermal matrix.

[0110] S4.2. Extract the probability values ​​of defects at the same spatial location from the defect probability distribution map in the continuous time slice sequence to construct a probability change curve; detect the peak position and duration of the probability change curve to confirm and mark the defects, and generate a defect spatial distribution map through multiple iterations.

[0111] Specifically, defect probability values ​​at the same spatial location are extracted from the defect probability distribution map in the continuous time slice sequence to construct a probability change curve in the time dimension; the peak position and duration of the probability change curve are detected.

[0112] The minimum duration threshold based on the temporal correlation of the Hebbian learning rule is used as the judgment criterion. The minimum duration threshold ranges from 10 milliseconds to 50 milliseconds. The range is determined by statistically analyzing the correlation distribution of the pulse firing patterns of normal chip samples in the time dimension, which is used to distinguish the continuous signal of real defects from transient noise. At the same time, a spatial overlap threshold is used as the judgment condition. The judgment condition is obtained by analyzing the spatial consistency characteristics of the pulse firing patterns. The spatial overlap threshold ranges from 85% to 95%. The range is obtained by analyzing the spatial consistency characteristics of the pulse firing patterns of high-confidence defect samples, which is used to ensure the spatial positioning accuracy of defect marking.

[0113] When the duration of a wave peak exceeds the minimum duration threshold and the spatial overlap meets the spatial overlap threshold requirement, it is marked as a confirmed defect, thereby eliminating misjudgments caused by transient interference.

[0114] For regions that do not meet the defect confirmation criteria, pulse firing mode recalibration is performed. This improves detection sensitivity by lowering the neuron firing rate threshold in regions with a high false alarm tendency, while simultaneously increasing the neuron firing rate threshold in low-response regions to suppress background noise.

[0115] It should be noted that the neuron firing rate threshold ranges from 0.1 to 0.3 (normalized units). This range is set based on the pulse firing statistics of the false alarm region (high firing rate) and the false negative region (low firing rate) in historical detection data, in order to balance detection sensitivity and specificity.

[0116] The defect probability distribution map is updated based on the adjusted parameters, and the spatiotemporal consistency verification is performed again. The optimization process is repeated until the defect label change rate is lower than the convergence threshold based on statistical process control. The output is a defect spatial distribution map containing defect type classification, spatial coordinate matrix, and confidence vector.

[0117] It should be noted that the convergence threshold ranges from 1% to 3%. This range is based on the coefficient of variation theory in statistical process control. By calculating the fluctuation range of the defect marker change rate during the iterative optimization process, the upper limit of the change rate in the continuous stable stage is taken as the judgment boundary to ensure the stability of the optimization process.

[0118] S5. Establish a neuromorphic feedback model, use the neuromorphic feedback model as a control criterion to correct the defect types in the defect spatial distribution map, and output a structured quality inspection report.

[0119] S5.1. Establish a mapping relationship library between defect types and control commands based on the defect spatial distribution map, and generate initial control commands.

[0120] Specifically, a mapping database between defect types and control commands is established based on the defect type classification information (cracks / offsets / contaminants) in the defect spatial distribution map. The mapping rules are as follows:

[0121] Offset defects trigger robotic arm pose correction commands, which include three-dimensional translation compensation parameters obtained based on the defect's spatial coordinates; crack defects directly generate laser repair energy parameters, which include a combination of wavelength, power, and duration that are adaptively adjusted according to the crack size and depth; contaminant defects activate ultrasonic cleaning, and the working parameters of ultrasonic cleaning include frequency, amplitude, and sweep frequency mode that are dynamically set according to the type and distribution density of contaminants.

[0122] Based on the established mapping relationship library, initial control commands are generated according to the spatial coordinate matrix and confidence vector of the defect spatial distribution map, as follows:

[0123] The coordinates of the defect center are transformed into the target position in the robotic arm coordinate system through a coordinate transformation matrix. The intensity of the control parameters is determined based on the specific values ​​in the confidence vector. For example, the laser power is increased proportionally or the ultrasonic treatment time is increased in high-confidence areas.

[0124] S5.2. Input the initial control command into the neuromorphic feedback model to simulate the state and calculate the difference between the simulated state and the actual state represented by the defect spatial distribution map.

[0125] Specifically, the initial control command is input into the neuromorphic feedback model for state simulation. Based on the robotic arm pose correction, laser repair energy parameters, or ultrasonic cleaning settings in the initial control command parameters, the event flow changes on the chip surface after simulation execution are set. The probability of the event occurring after the initial control command is applied is predicted by an event generation algorithm, expressed as:

[0126] ;

[0127] in, To give an initial control command events under conditions The conditional probability of occurrence This is a continuous multiplication operation on subsequent elements. For event indexing, For the first The properties of an event To act on the first The initial control command parameters for each event. The weight matrix is ​​obtained by training with historical data and is used to quantify the causal impact of initial control commands on events.

[0128] The chip surface state is reconstructed by using the predicted probability of event flow occurrence, generating a high-precision surface topography map and event density distribution map. The surface topography map reflects the physical structure changes after instruction execution, and the event density distribution map quantifies the event activity.

[0129] Difference quantification is performed based on the reconstructed chip surface state (simulated state) and the actual defect spatial distribution map (actual state). A difference value matrix is ​​constructed to characterize the deviation between the simulated state and the actual state. The difference value matrix consists of two independently calculated components: the spatial difference matrix measures the positional offset between the simulated defect coordinates and the actual defect coordinates using the Euclidean distance algorithm; the intensity difference matrix calculates the energy deviation between the simulated energy intensity and the actual energy intensity using the relative error percentage algorithm.

[0130] It should be noted that the expression for calculating spatial differences is:

[0131] ;

[0132] in, This represents the spatial difference value. To simulate defect coordinates, The coordinates of the actual defect;

[0133] The expression for calculating the intensity difference is:

[0134] ;

[0135] in, Due to differences in strength, To simulate energy intensity, This represents the actual energy intensity.

[0136] S5.3. Adjust the initial control command parameter strength according to the difference value, send it to the actuator for defect correction, and output a structured quality inspection report.

[0137] Specifically, based on the difference value matrix obtained after the initial control command is simulated by the neuromorphic feedback model, the trajectory of the robotic arm is reversed according to the offset in the spatial difference, and the three-dimensional translation parameters in the pose correction command of the robotic arm are precisely calibrated; at the same time, energy calibration is performed according to the intensity difference value, and when the intensity difference is detected to exceed the difference threshold, the power gradient in the laser repair energy parameter or the amplitude intensity in the ultrasonic cleaning parameter is adjusted proportionally.

[0138] It should be noted that the difference threshold ranges from 5% to 8%. This range is based on the safety margin of the thermal damage threshold of semiconductor materials and is verified through statistical analysis of historical process data to ensure that the energy calibration process effectively corrects deviations while absolutely avoiding damage to the chip's microstructure.

[0139] Based on the severity and distribution density of defects in the defect spatial distribution map, the execution timing of the initial control commands is re-optimized to ensure that high-frequency difference areas are processed first, an optimized control command set is generated and sent to the actuator, and event stream data during the execution process is collected.

[0140] After the actuator completes the defect correction, it collects the actual correction effect through neuromorphic sensors, compares the changes in the event flow pattern before and after the correction, detects the residual defect areas that have not been eliminated, automatically triggers secondary instruction optimization, feeds the residual defect information back to the mapping relationship library update stage for iterative control, until the defect is completely eliminated or the iteration termination condition is met, and finally outputs a structured quality inspection report.

[0141] This embodiment also provides an automated quality inspection system for semiconductor chip products, including:

[0142] The event sensing module acquires event stream data from the chip surface, constructs a spatiotemporal distribution feature vector, and performs voxelization processing on the event stream data based on the spatiotemporal distribution feature vector to obtain a dynamic spatiotemporal voxel grid.

[0143] The feature decoupling module inputs a dynamic spatiotemporal voxel grid into a deformation-aware 3D convolutional kernel to decouple spatiotemporal features and generate a decoupled spatiotemporal feature map. At the same time, it generates a self-correcting geometric reference matrix through event trajectory parsing and fuses the geometric reference matrix with the decoupled spatiotemporal feature map to generate a reference fusion feature map.

[0144] The neural decision-making module establishes a pulse firing modulation mechanism based on a spiking neural network, extracts the pulse firing rate features of the benchmark fusion feature map, and performs pulse-frequency domain cross-modal coupling on the benchmark fusion feature map according to the pulse firing rate features to generate a pulse frequency domain feature map.

[0145] The defect optimization module performs defect decision-making on the pulse frequency domain feature map through neural plasticity rules and spatiotemporal consistency verification mechanism, and performs iterative optimization through pulse firing mode association conditions to generate a defect spatial distribution map.

[0146] The closed-loop control module establishes a neuromorphic feedback model, uses the neuromorphic feedback model as a control criterion, corrects the defect types in the defect spatial distribution map, and outputs a structured quality inspection report.

[0147] This embodiment also provides a computer device applicable to the automated quality inspection method for semiconductor chip products, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the automated quality inspection method for semiconductor chip products as proposed in the above embodiment.

[0148] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0149] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the automated quality inspection method for semiconductor chip products as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0150] In summary, this invention dynamically adjusts the spatial resolution and temporal slice depth based on the density gradient magnitude and polarity transition probability, enabling the voxel grid to adaptively conform to the chip surface deformation state, thus ensuring the physical authenticity of feature extraction. By dynamically adjusting the convolution kernel deformation parameters through the material's thermal expansion coefficient, the feature decoupling process accurately matches the actual physical deformation of the chip, significantly improving the spatiotemporal feature correlation. Furthermore, by employing a pulse-frequency domain cross-modal coupling mechanism, the pulse firing rate characteristics and frequency domain energy distribution are deeply integrated, enhancing the response intensity of micro-defects.

[0151] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An automated quality inspection method for semiconductor chip products, characterized in that: include, Event stream data on the chip surface is acquired, a spatiotemporal distribution feature vector is constructed, and the event stream data is voxelized based on the spatiotemporal distribution feature vector to obtain a dynamic spatiotemporal voxel grid. The specific steps for voxelizing the event stream data are as follows: Based on the spatiotemporal distribution feature vector, determine the spatial resolution parameter and the temporal slice depth parameter; Configure a 3D mesh coordinate system based on spatial resolution parameters and time slice depth parameters, map event stream data to the 3D mesh coordinate system, and assign voxel weights based on event density values; Based on the assigned voxel weights, neighborhood interpolation compensation is performed on low-density voxel meshes in the 3D mesh coordinate system, while the interpolation coefficients are adjusted according to the polarity distribution values. Perform nonlinear compression on the high-density voxel mesh in the three-dimensional mesh coordinate system; The dynamic spatiotemporal voxel grid is input into the deformation-aware 3D convolutional kernel to decouple the spatiotemporal features and generate a decoupled spatiotemporal feature map. At the same time, a self-correcting geometric reference matrix is ​​generated through event trajectory analysis. The geometric reference matrix is ​​fused with the decoupled spatiotemporal feature map to generate a reference fusion feature map. The dynamic spatiotemporal voxel grid is input into a deformation-aware 3D convolutional kernel for spatiotemporal feature decoupling. Includes the following steps, Based on dynamic spatiotemporal voxel grids, construct deformable convolutional kernel groups; Using the deformable convolution kernel group, adaptive receptive field convolution is performed in the spatial dimension, and gated recurrent convolution is performed in the temporal dimension; Based on convolutional operations in spatial and temporal dimensions, spatial edge response features and temporal state evolution features are extracted and fused according to spatiotemporal location encoding to generate a decoupled spatiotemporal feature map. A pulse firing modulation mechanism is established based on a pulse neural network. The pulse firing rate feature of the reference fusion feature map is extracted. The pulse firing rate feature is then used to perform pulse-frequency domain cross-modal coupling on the reference fusion feature map to generate a pulse frequency domain feature map. Defect decisions are made on the pulse frequency domain feature map through neural plasticity rules and spatiotemporal consistency verification mechanisms, and iterative optimization is performed through pulse firing mode association conditions to generate a defect spatial distribution map. A neuromorphic feedback model is established and used as a control criterion to correct the defect types in the defect spatial distribution map, and a structured quality inspection report is output.

2. The automated quality inspection method for semiconductor chip products as described in claim 1, characterized in that: The specific steps for constructing the spatiotemporal distribution feature vector are as follows: The event stream data on the chip surface is timestamped and divided into regular spatiotemporal grids. The event density and polarity distribution values ​​within each grid are statistically analyzed. Based on the event density value and polarity distribution value, calculate the density gradient change and polarity transition probability between adjacent spatiotemporal grids; Based on the density gradient change and polarity transition probability, a spatiotemporal distribution feature vector is constructed.

3. The automated quality inspection method for semiconductor chip products as described in claim 1, characterized in that: The specific steps for generating a self-calibrating geometric reference matrix through event trajectory analysis are as follows: Extract the set of continuous event trajectory points from the decoupled spatiotemporal feature map, fit it to obtain the motion trajectory plane equation, calculate the angle deviation between the trajectory plane normal vector and the chip reference coordinate system, and construct the rotation compensation sub-matrix. Based on the event trajectory point set, the event density change rate is analyzed to generate scaling factor compensation coefficient; By fusing the rotation compensation submatrix and the scaling factor compensation coefficient, a self-correcting geometric reference matrix is ​​generated.

4. The automated quality inspection method for semiconductor chip products as described in claim 1, characterized in that: The specific steps for extracting the pulse firing rate feature from the baseline fusion feature map are as follows: The baseline fused feature map is input into the response layer of the spiking neural network, and the threshold parameters of the neurons in the spiking neural network are dynamically adjusted through the firing rate modulation mechanism. Within a preset time window, the number of pulse triggers of neurons in the spiking neural network is counted, the pulse trigger frequency per unit time is calculated, a pulse firing rate heatmap is generated, and the firing rate features of high-response regions are extracted from it.

5. The automated quality inspection method for semiconductor chip products as described in claim 1, characterized in that: The specific steps for performing pulse-frequency domain cross-modal coupling on the reference fused feature map are as follows: Based on the extracted high-response region emission rate features, multi-scale time-frequency transformation is performed on the event stream data to generate a frequency domain energy distribution spectrum. Spatial location coding information is extracted from the firing rate features of the high response region, a cross-modal attention weight matrix is ​​constructed, and the frequency domain energy distribution spectrum is combined with the attention weight matrix through matrix multiplication to redistribute the frequency domain energy distribution spectrum. The redistributed frequency domain energy distribution spectrum and the firing rate feature map are cascaded to generate a pulse frequency domain feature map.

6. The automated quality inspection method for semiconductor chip products as described in claim 1, characterized in that: The specific steps for generating the defect spatial distribution map are as follows: High firing rate neuron clusters in the pulse frequency domain feature map are identified, and the Heb learning rule is applied to enhance the connection strength of defect-related neural pathways. At the same time, the pulse temporal-dependent plasticity mechanism is used to adjust the synaptic weight update direction of defect-related neural pathways to generate a defect probability distribution map. Extract the probability values ​​of defects at the same spatial location from the defect probability distribution map in a continuous time slice sequence to construct a probability change curve; The peak position and duration of the probability change curve are detected to confirm and mark defects, and a defect spatial distribution map is generated through multiple iterations and optimizations.

7. The automated quality inspection method for semiconductor chip products as described in claim 1, characterized in that: The specific steps for generating the structured quality inspection report are as follows. Based on the aforementioned defect spatial distribution map, a mapping relationship library between defect types and control commands is established, and initial control commands are generated. The initial control command is input into the neuromorphic feedback model to simulate the state, and the difference between the simulated state and the actual state represented by the defect spatial distribution map is calculated. The initial control command parameter strength is adjusted based on the difference value, and then sent to the actuator for defect correction, outputting a structured quality inspection report.

8. An automated quality inspection system for semiconductor chip products, based on the automated quality inspection method for semiconductor chip products according to any one of claims 1 to 7, characterized in that: include, The event sensing module acquires event stream data from the chip surface, constructs a spatiotemporal distribution feature vector, and performs voxelization processing on the event stream data based on the spatiotemporal distribution feature vector to obtain a dynamic spatiotemporal voxel grid. The feature decoupling module inputs a dynamic spatiotemporal voxel grid into a deformation-aware 3D convolutional kernel to decouple spatiotemporal features and generate a decoupled spatiotemporal feature map. At the same time, it generates a self-correcting geometric reference matrix through event trajectory parsing and fuses the geometric reference matrix with the decoupled spatiotemporal feature map to generate a reference fusion feature map. The neural decision-making module establishes a pulse firing modulation mechanism based on a spiking neural network, extracts the pulse firing rate features of the benchmark fusion feature map, and performs pulse-frequency domain cross-modal coupling on the benchmark fusion feature map according to the pulse firing rate features to generate a pulse frequency domain feature map. The defect optimization module performs defect decision-making on the pulse frequency domain feature map through neural plasticity rules and spatiotemporal consistency verification mechanism, and performs iterative optimization through pulse firing mode association conditions to generate a defect spatial distribution map. The closed-loop control module establishes a neuromorphic feedback model, uses the neuromorphic feedback model as a control criterion, corrects the defect types in the defect spatial distribution map, and outputs a structured quality inspection report.

Images