A wafer defect detection method and device based on multi-modal data and a graph neural network
By constructing a multimodal dataset and graph neural network, the problem of cross-interface defect modeling in heterogeneous dual wafers was solved, enabling accurate prediction of latent interface voids and improving detection accuracy and yield.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TIANJIN GUORUI MICROELECTRONICS TECHNOLOGY CO LTD
- Filing Date
- 2026-02-28
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies struggle to effectively model cross-interface defects between heterogeneous dual wafers, especially the formation mechanism of latent interface voids in hybrid bonding processes. They also cannot accurately predict the spatial correlation effects under dynamic process constraints, resulting in low detection accuracy.
A multimodal dataset is constructed, including micromorphology, transient physical field and layout logic features. A dual-graph structure is established using graph neural networks. The morphology mismatch is quantified through topological convolution and cross attention mechanism. Combined with physical field projection and spatiotemporal evolution memory network, the micromorphology and physical interaction features in the bonding process are captured to achieve the prediction of latent defects.
It improves the accuracy of identifying interface voids in hybrid bonding processes, enhances the physical interpretability of the model, adapts to the engineering scalability of advanced processes, and improves the yield.
Smart Images
Figure CN122134682A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the application of artificial intelligence in the field of semiconductor manufacturing technology, specifically to a wafer defect detection method and apparatus based on multimodal data and graph neural networks. Background Technology
[0002] As semiconductor manufacturing technology evolves towards advanced process nodes, 3D integrated circuits (3D ICs) have become a key technology for continuing performance improvement. Among these technologies, hybrid bonding processes employing ultra-fine pitch (e.g., less than 10 micrometers) are a core technique for achieving high-density vertical interconnects. This process simultaneously achieves metallic bonding between copper pads and molecular bonding between dielectric layers on an atomically flat wafer surface, resulting in extremely high interconnect density and excellent electrical performance. However, hybrid bonding processes face challenges, with a key defect being latent interface voids. These voids are difficult to detect directly using conventional optical or acoustic detection methods during the bonding compression process. They typically form and grow during subsequent thermal annealing processes due to the combined effects of residual stress release at the interface, metal atom diffusion, and initial interface imperfections. These interface voids can cause increased interconnect resistance, decreased electromigration reliability, and in severe cases, physical delamination of the chip, affecting product yield and reliability.
[0003] In related technologies, most wafer defect detection methods focus on identifying defects on the surface of a single wafer, or on static non-destructive testing after bonding. For example, convolutional neural networks are used to extract and classify features from scanning electron microscope or optical images of the wafer surface to identify surface defects such as particles and scratches. However, these methods are difficult to apply to predicting cross-interface defects between two heterogeneous wafers (referred to as the first wafer and the second wafer) in hybrid bonding processes. The formation mechanism of latent interface voids is complex, and its causes are highly dependent on the asymmetric microstructure features between the surfaces of the two wafers to be bonded (e.g., the morphological differences between them after chemical mechanical planarization), the transient physical field distribution of the bonding equipment during the pressing process (e.g., pressure gradient, temperature compensation timing), and the local logic features of the pads determined by the chip layout. Existing technologies lack the ability to effectively characterize and model the cross-interface spatial correlation effects caused by microstructure mismatch between heterogeneous wafers under dynamic process constraints. These technologies cannot establish a nonlinear physical interaction model among the "topological features of the first wafer, the topological features of the second wafer, and process parameters." They cannot accurately predict how the local protrusions of the first wafer and the local depressions of the second wafer co-evolve under specific pressure curves, ultimately leading to stress concentration in local areas and inducing microscopic voids. Therefore, how to utilize deep learning techniques that process multimodal data and complex spatial topologies to predict and detect latent interface voids in hybrid bonding processes is a key technical problem that needs to be solved in the field of advanced semiconductor manufacturing processes. Summary of the Invention
[0004] This invention provides a wafer defect detection method based on multimodal data and graph neural networks, comprising the following steps: Step 1: Construct a multimodal heterogeneous dataset for heterogeneous dual wafers, including microstructure data of the first and second wafers, transient physical field timing data of the bonding machine, and layout logic features. Step 2: Establish a wafer space topology model based on a dual-graph structure, abstract the pads as graph nodes, and define the edges of the graph according to physical proximity and exhaust paths; perform topological convolution operation on the feature graph to aggregate spatial correlation features and generate node embedding vectors that represent local background information. Step 3: Design a cross-interface bipartite interaction and spatial alignment mechanism, including: establishing a spatial topological alignment model and calculating the spatial projection relationship between the two wafer node sets; constructing a cross-interface interaction bipartite graph and establishing interaction edges based on the projection overlap; using a cross-attention mechanism to calculate similarity in order to obtain the interface shape sensing weight that characterizes the degree of interface shape mismatch, and generating a feature vector that characterizes the interface bonding state. Step 4: Implement energy projection from the global physics field to local grid points. Use the physics field mapping operator to project the transient physics field time series data onto the interaction edges of the cross-interface interactive bipartite graph to generate local physical features for characterizing the pad-level micro-physical state. Step 5: Predict latent defects based on spatiotemporal evolution memory. Input the feature vector and local physical features into the spatiotemporal evolution memory network for feature enhancement, generate a composite feature vector, and use a classifier to calculate the probability score of latent interface voids based on the composite feature vector. Step 6: Output the detection results and generate a full-image defect risk heat map of the wafer based on the occurrence probability score.
[0005] Further, in step one, the micro-morphology data includes the exhaust path width; in step two, the initial feature vector of the graph node includes the center coordinates of the pad, the recess depth, the local morphology slope, and the material properties, and the weight features of the graph edge include the Euclidean distance between the centers of the pads and the effective cross-sectional area of the exhaust channel; in step three, calculating the spatial projection relationship includes calculating the three-dimensional spatial projection distance between the node sets of the first wafer and the second wafer in the pressed state, establishing interactive edges based on the projection overlap includes establishing interactive edges when the physical overlap area between nodes exceeds a preset threshold, and calculating the similarity between nodes includes using the node embedding vector of the first wafer as the query and the node embedding vector of the second wafer as the key and value, and obtaining the interface morphology sensing weight by calculating the dot product similarity.
[0006] Furthermore, the steps for constructing a multimodal heterogeneous dataset for the first and second wafers also include: preprocessing the microstructure data, transient physical field time-series data, and layout logic features. The preprocessing includes scaling each feature data to a uniform numerical range using a normalization method, and using time-series resampling technology to ensure that the sampling rate of the microstructure data and the sampling rate of the transient physical field time-series data are consistent in the time dimension. For missing data points in the microstructure data, data completion is performed using a bicubic interpolation algorithm or a deep learning-based super-resolution reconstruction technique.
[0007] Furthermore, the steps for implementing energy projection from the global physical field to local grid points specifically include: constructing an energy projection model using a physical information neural network, the loss function of which includes a data fitting term and a physical constraint residual term, wherein the physical constraints include the interface mechanical equilibrium equation and the heat conduction equation; training with minimizing the physical constraint residual term as the optimization objective to learn the mapping relationship that projects the global total pressure onto each interaction edge of the cross-interface interaction bipartite graph, and calculating the local equivalent contact pressure of each pad node pair under the pressure time sequence as a local physical feature.
[0008] Furthermore, the steps for predicting latent defects based on spatiotemporal evolution memory specifically include: constructing a spatiotemporal evolution memory network, which integrates a temporal convolutional network and a convolutional long short-term memory network to learn and store the evolutionary trajectory of the physical interaction features at the bonding moment, and to capture the cumulative effects of the pressure ramp-up stage, constant pressure stage, and depressurization stage on the interface microstructure during the pressing process; introducing a reverse causal tracing module based on contrastive learning, which learns a discriminative subspace in the feature space that can distinguish between defective samples and normal samples by comparing and analyzing known defective samples with normal samples, so as to identify the core combination of causes leading to the formation of latent defects.
[0009] Furthermore, the steps for designing the cross-interface bipartite interaction and spatial alignment mechanism also include: when updating the edge features of the cross-interface interaction bipartite graph, a transient trapped gas pressure evolution operator is introduced. The operator is coupled with the exhaust path width in the micro-morphology data and the permeability constant simplified based on Darcy's law to calculate the transient trapped gas pressure value generated under the dynamic compression sequence. This transient trapped gas pressure value forms a nonlinear cancellation relationship with the feature vector characterizing the interface bonding state.
[0010] Furthermore, the steps for outputting the detection results also include: using a trained deep classifier as a surrogate model, and using gradient ascent or genetic algorithm to search in the parameter space for a combination of pressure timing curves and temperature compensation strategies that minimize the total defect risk of the entire wafer, thereby generating a collaborative optimization scheme for the machine process parameters.
[0011] Another aspect of the present invention provides a wafer defect detection device based on multimodal data and graph neural networks, comprising: a dataset construction module for acquiring microscopic morphology data of a first wafer and a second wafer, transient physical field time-series data of a bonding machine, and layout logic features, and preprocessing the acquired data to generate a heterogeneous dataset; a dual-graph topology modeling module for abstracting pads as graph nodes and defining graph edges, constructing a first feature graph and a second feature graph, and generating node embedding vectors through topological convolution; a cross-interface interaction module for calculating the spatial projection relationship between the node sets of the two wafers and constructing a cross-interface interaction bipartite graph, and generating feature vectors representing the interface bonding state using a cross-attention mechanism; a physical field projection module for constructing an energy projection model using a physical information neural network, projecting transient physical field time-series data onto the interaction edges of the cross-interface interaction bipartite graph, and generating local physical features; a defect prediction module for feature enhancement using a spatiotemporal evolution memory network and calculating the probability score of latent interface voids; and a result output module for generating a wafer full-map defect risk heat map based on the probability score and generating a collaborative optimization scheme for machine process parameters.
[0012] This invention discloses a wafer defect detection method based on multimodal data and graph neural networks. The core of this method lies in establishing a deep learning modeling framework that couples physical constraints with spatial topological relationships to address latent defects at heterogeneous dual-wafer interfaces in hybrid bonding processes. This scheme constructs a multimodal heterogeneous dataset by integrating the microscopic topography of the first and second wafers, the transient physical field of the bonding machine, and the temporal layout logic features. It abstracts pad distribution features using the dual-graph topology and designs a cross-interface bipartite interaction and spatial alignment mechanism, quantifying the degree of topographic mismatch between the two wafer surfaces through a cross-attention operator. Based on this, a physical information neural network is introduced to perform energy projection from the global physical field to local grid points, mapping macroscopic process parameters to pad-level microscopic physical features. Finally, a spatiotemporal evolution memory network integrating temporal convolution and convolutional long short-term memory is used to capture the feature evolution trajectory throughout the entire bonding cycle. Combined with a reverse causal tracing mechanism, it achieves accurate prediction of the probability of latent interface voids, thus constructing a technical path from defect identification to closed-loop optimization of process parameters.
[0013] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects.
[0014] First, by constructing a dual-graph interactive neural network architecture, the modeling of the micro-morphological mismatch between the first wafer and the second wafer in the hybrid bonding process can be realized, which can capture the physical correlation between heterogeneous surfaces and improve the recognition accuracy of interface voids.
[0015] Second, a projection mechanism from the global physical field to local grid points was introduced, and a correlation model between macroscopic machine tool process parameters and microscopic pad-level defects was established. This enabled the deep learning model to perceive key physical constraints such as pressure gradients and temperature compensation, thereby enhancing the physical interpretability of the model.
[0016] Third, the spatiotemporal evolution memory network and the reverse causal tracing module are used to perform long-range attribution of latent defects, identify the combination of bonding transient physical states that lead to void formation after annealing, and provide decision-making basis for prevention and optimization at the process end.
[0017] Fourth, the technical solution can handle large-scale heterogeneous graph calculations, is compatible with advanced processes with micro-pitch less than 10 micrometers, has engineering scalability, and helps improve the yield of hybrid bonding processes. Attached Figure Description
[0018] The embodiments of the present invention will now be described with reference to the accompanying drawings.
[0019] Figure 1 This is a flowchart of a wafer defect detection method based on multimodal data and graph neural networks according to an embodiment of the present invention.
[0020] Figure 2 This is a schematic diagram of heterogeneous dual-wafer dual-graph topology modeling and cross-interface two-part interaction mechanism according to an embodiment of the present invention.
[0021] Figure 3 This is a schematic diagram of the energy projection mechanism from the global physical field to local grid points according to an embodiment of the present invention.
[0022] Figure 4 This is a structural block diagram of a wafer defect detection device based on multimodal data and graph neural networks according to an embodiment of the present invention.
[0023] Figure 5 This is a schematic diagram of the spatiotemporal evolution memory network structure according to an embodiment of the present invention.
[0024] Figure 6 This is a performance comparison chart between the wafer defect detection method according to an embodiment of the present invention and existing technology methods.
[0025] Figure 7 This is a three-dimensional schematic diagram of the microstructure of a wafer surface according to an embodiment of the present invention.
[0026] Figure 8 This is a schematic cross-sectional view of the physical process of void formation at the dual-wafer bonding interface according to an embodiment of the present invention. Detailed Implementation
[0027] The embodiments of the present invention will now be described with reference to the accompanying drawings.
[0028] It should be noted that the terms used in the embodiments of the present invention, such as "first" and "second," are only used for distinguishing descriptions and do not have a specific order or significance. Where there is no conflict, the embodiments of the present invention and the features within them can be combined with each other.
[0029] Example 1 This embodiment provides a wafer defect detection method based on multimodal data and graph neural networks. (Refer to...) Figure 1 This method addresses the challenge of effectively modeling cross-interface spatial correlation effects caused by microstructure mismatch between the surfaces of two heterogeneous wafers in hybrid bonding processes, under dynamic process constraints. This makes it difficult to accurately predict latent interface voids formed during semiconductor manufacturing. The method includes the following steps: Step S100: Construct a multimodal heterogeneous dataset for heterogeneous dual wafers.
[0030] This step is used to collect multi-dimensional data related to latent interface voids, providing input for deep learning models. Specifically, this step acquires full feature data of the first and second wafers to be bonded by integrating a semiconductor manufacturing execution system (MES) and various metrology devices.
[0031] In its implementation, this step includes several sub-steps. First, the microstructure data of the first and second wafers are acquired. Preferably, atomic force microscopy (AFM) is used to perform high-density scanning of the dual-wafer surfaces after chemical mechanical planarization (CMP) to acquire three-dimensional topographic data at the nanoscale. This data accurately reflects the dishing depth of each copper pad region and the erosion depth of the dielectric region, and can be organized into a high-resolution spatial feature lattice characterizing the wafer surface height information. Figure 7The figure illustrates the nanoscale topographic undulations on the wafer surface. The "Dishing" region represents the dish-shaped depressions formed on the copper pads after CMP processing, typically tens of nanometers deep; the "Erosion" region represents the surface corrosion of the dielectric region after CMP, exhibiting fluctuating characteristics. The X and Y axes in the figure represent the micrometer-level planar coordinates of the wafer surface, while the Z axis represents the nanometer-level height variation. Next, transient physical field timing data of the bonding machine are acquired. Throughout the entire wafer bonding cycle, a series of dynamic physical parameters are collected in real-time with millisecond-level time resolution using high-frequency sensors built into the machine. These parameters include pressure gradient signals from each independent heating zone of the bonding head, temperature compensation timing curves of the wafer stage, and displacement sensor data for precise closed-loop control. These timing data collectively constitute a description of the global process energy field environment acting on the wafer at the moment of bonding. Finally, the layout logic characteristics of the wafer are analyzed. By parsing Graphical Data System II (GDSII) or Open Artwork System Exchange Standard (OASIS) layout design files, geometric and logical information related to stress distribution and interface gas venting paths is extracted. This includes, for example, the geometry, spatial coordinates, and pad density of each copper pad in local areas. Furthermore, the width and length of venting paths for residual interface gases, as well as the electrical connections between adjacent pads, are extracted. Further, planarization models can be calculated using finite element analysis to simulate the layout, predict the global stress distribution that may form after CMP processing, and use this stress map as a background feature for subsequent models.
[0032] Finally, the collected multi-source heterogeneous data were preprocessed. To eliminate the influence of different physical dimensions, a min-max normalization method was used to uniformly scale each feature data to a numerical range of [0, 1]. To ensure effective alignment between static terrain spatial features and dynamic transient physical field time-series data in the model, time-series resampling techniques, such as linear interpolation, were employed to maintain consistent sampling rates across the time dimension. For missing terrain data points that might arise during AFM scanning due to probe movement or surface contamination, a bicubic interpolation algorithm was used for local smoothing data completion, thereby ensuring the integrity and accuracy of the input dataset.
[0033] Step S200: Establish a wafer space topology model based on a dual-graph structure.
[0034] This step utilizes graph neural networks to model spatial relationships, mathematically abstracting and modeling the pad distribution characteristics and their interactions on the surfaces of the first and second wafers. (Refer to...) Figure 2 This step, through a heterogeneous dual-wafer dual-graph topology modeling mechanism, lays the foundation for subsequent cross-interface interactive analysis.
[0035] Specifically, this step first constructs a first feature map (defined as Graph A) 201 for the first wafer and a second feature map (defined as Graph B) 202 for the second wafer. In each feature map, each copper pad is abstracted as a graph node 203. The initial feature vector of a node contains the multidimensional attributes of the pad. For example, the feature vector h_i of a node can be represented as [x, y, d, g, m, a], where x and y are the center coordinates of the pad, d is the depth of the depression formed after the CMP process, g is the gradient or slope of the local topography, m is the material property of the region where the pad is located (e.g., the Young's modulus of copper and dielectric materials), and a is the effective dominance area of the pad calculated based on the Voronoi diagram. This area is used to characterize the neighborhood influence range of the pad.
[0036] Secondly, the edges 204 of the graph are defined based on the physical proximity between the pads and the venting paths in the process design. Preferably, the K-Nearest Neighbors (KNN) algorithm is used to connect the spatially nearest nodes. Simultaneously, in conjunction with the pre-defined physical venting logic in the layout design, pad nodes located on the same venting channel are also connected via edges. The edge weight features include physical quantities describing the relationships between nodes, such as the Euclidean distance between the centers of two pads and the effective cross-sectional area of the venting channel connecting the two pads. To physically model the venting process of interfacial gases during bonding, a permeability constant simplified based on Darcy's law is also introduced into the edge features to characterize the flow resistance of gas in the microchannels.
[0037] Finally, a local topological convolution operation is performed on each constructed feature map. By stacking multiple layers of Graph Convolutional Networks (GCNs), the model learns and aggregates the spatial association features between each pad node and its neighboring pads. In each layer of the GCN, the node feature update process uses adjacency matrix, degree matrix, and weight matrix for calculation. To prevent the excessive smoothing problem of node feature convergence after multi-layer propagation in deep graph neural networks, techniques such as layer normalization and residual connection are introduced during graph convolution calculation to generate deep node embedding vectors that can effectively represent local background information.
[0038] Step S300: Design a cross-interface two-part interaction and spatial alignment mechanism.
[0039] This step is used to accurately model the spatial matching relationship between two independent wafers during the physical lamination process and the nonlinear interactions between their surfaces. (Refer to...) Figure 2 This mechanism is key to connecting the first feature map 201 and the second feature map 202.
[0040] First, a cross-interface spatial topology alignment model is established. Based on the global alignment coordinates provided by the bonding machine and the actual measurement error data of the alignment marks, the three-dimensional spatial projection distances of each node in the first wafer and each node in the second wafer under ideal bonding conditions are calculated. To improve alignment accuracy, a deformable alignment operator is introduced. This operator can perform non-rigid local compensation on the ideal projection coordinates based on the local warping and distortion field map generated during the vacuum adsorption process of the wafer, thereby obtaining an alignment relationship closer to the physical reality.
[0041] Next, a bipartite interaction graph 205 is constructed. This graph structure describes the interaction between the first wafer and the second wafer. The complete set of nodes of the first wafer and the complete set of nodes of the second wafer are defined as the two vertex sets of this bipartite graph. When the physical overlap area between node i in the first wafer and node j in the second wafer exceeds a preset threshold in the projection direction perpendicular to the wafer surface, an interaction edge 206 is established between node i and node j.
[0042] Furthermore, an interaction layer based on a multi-head cross-attention mechanism is introduced to quantify the degree of shape mismatch between two opposing surfaces. Specifically, the node embedding vector of the first wafer is linearly transformed into a query, and the node embedding vector of the second wafer is linearly transformed into a key and a value. By calculating the dot product similarity between the query and the key, the model learns a set of interface shape sensing weights.
[0043] Finally, the edge features of the bipartite graph are dynamically updated using a gated recurrent unit (GRU) network. Interaction weights are fused with the original terrain features of the nodes to generate feature vectors representing the interface bonding state. To more accurately model the microscopic exhaust dynamics during the interface bonding process, a transient trapped gas pressure evolution operator is introduced when updating the edge features of the bipartite graph. This operator, based on physical formulas coupled with the exhaust path width from the microscopic topography data and a permeability constant simplified based on Darcy's law, is used to calculate the transient trapped gas pressure value generated by the contact closure of local microscopic depressions formed on the surfaces of two opposing pads under a dynamic pressing sequence. This transient trapped gas pressure value forms a nonlinear cancellation relationship with the interface bonding state, enabling more accurate identification of the initial bubble nucleus caused by poor microscopic exhaust. Figure 8The diagram illustrates the physical process of void formation at the bonding interface of two wafers. Both the first and second wafers consist of a silicon substrate, a dielectric layer, and copper pads. Typical regions in the diagram include: a normal bonding region where the upper and lower copper pads are in close contact; a morphology mismatch region where gaps and stress concentration points are created at the interface due to the mismatch between local protrusions on the first wafer and local depressions on the second wafer (indicated by red star marks in the diagram); and a void formation region where the morphology of the voids traps gas within the interface cavities during bonding, forming voids containing trapped gas. The applied mechanical pressure is indicated by arrows pointing in the direction of pressure, and the bonding interface is represented by blue dashed lines. This cross-sectional view demonstrates how the microstructure mismatch and pressure distribution work together to lead to the formation of interface voids.
[0044] Step S400: Perform energy projection from the global physics field to local grid points.
[0045] This step establishes a bridge between macroscopic process parameters and microscopic local physical variables, enabling the model to understand how global process changes affect the bonding state of each pad pair. (Refer to...) Figure 3 This step is achieved through a macro-micro energy projection mechanism based on a physical information neural network (PINN).
[0046] First, a physical field mapping operator is constructed. The global pressure gradient signal P(t) 301 and the temperature compensation signal T(t) 302 of the bonding machine are used as the dynamic input 303 of the operator.
[0047] Secondly, an energy projection model is constructed using a Physical Information Neural Network (PINN) 304. This model learns the mapping relationship to rationally distribute the global total pressure to each interaction edge 206 of the cross-interface interaction bipartite diagram 205 by minimizing the data error and physical constraint residual in the loss function. The model considers the non-uniform distribution of the support stiffness of the wafer stage in space and calculates the local equivalent contact pressure P_local(i,j) 305 for each pad node pair (i, j) under a specific pressure sequence.
[0048] Finally, an interface energy field modulation operator is introduced. Local physical features obtained by projection through the PINN model are injected into the cross-interface interaction bipartite graph as additional features of interaction edges or nodes. For example, P_local(i, j) is used as a new feature of the interaction edge. Furthermore, an edge gating mechanism is used to simulate how local pressure gradients regulate the gas exhaust rate between adjacent pads. This mechanism uses a small neural network to dynamically adjust the weights of edge features related to gas exhaust based on local pressure values, simulating the risk of void formation when local pressure is too low or exhaust paths are blocked.
[0049] Step S500: Prediction of latent defects based on spatiotemporal evolution memory.
[0050] This step addresses the long-term temporal dependency of interface voids, which only becomes apparent after thermal annealing, enabling accurate prediction of latent defects.
[0051] First, construct a spatiotemporal evolution memory network. For example... Figure 5 As shown, the spatiotemporal evolution memory network integrates a temporal convolutional network (TCN) 501 and a convolutional long short-term memory network (ConvLSTM) 502 to learn and store the evolutionary trajectory of physical interaction features at the moment of bonding. The network input is a sequence of full-image feature snapshots (t1, t2, t3, ..., tn) on a series of time slices during the bonding process, which are transmitted to the temporal convolutional network 501 and the convolutional long short-term memory network 502 for processing. The temporal convolutional network 501 adopts an expanded convolutional layer structure, and its receptive field increases with the number of layers to extract multi-scale temporal features. The convolutional long short-term memory network 502 contains multiple LSTM units, corresponding to the pressure ramp-up stage, constant pressure stage, and decompression stage of the bonding process, respectively. It captures the cumulative effect of different stages on the interface microstructure during the bonding process by passing hidden states. The network also includes a contrastive learning reverse causal tracing module 503, which learns a discriminative subspace in the feature space to distinguish between defective and normal samples by comparing and analyzing known defective samples with normal samples. The outputs of the temporal convolutional network 501, the outputs of the convolutional long short-term memory network 502, and the discrimination results of the contrastive learning reverse causal tracing module 503 are fused by the feature fusion module 504 to output a composite feature vector for defect prediction.
[0052] Secondly, a reverse causal attribution module based on contrastive learning is introduced. During model training, this module compares and analyzes a large number of known post-annealing defect samples with normal samples, utilizing a self-supervised learning mechanism to discover key feature combinations leading to defect formation. During inference, this module maps the final void location information after hot annealing back to the feature space of the initial bonding stage, identifying the core contributing factors to the formation of latent defects.
[0053] Finally, the probability score of latent interface voids is calculated. The enhanced composite feature vector is then input into a deep classifier. This classifier employs a multilayer perceptron with spatial pyramid pooling (SPP) layers. The SPP layers aggregate contextual information at different scales, enhancing the model's ability to perceive void features of different sizes. The classifier ultimately outputs a void occurrence probability density map corresponding to the wafer coordinate system.
[0054] Step S600: Output the detection results and process control closed loop.
[0055] First, a high-resolution heatmap of the entire wafer defect risk is generated. Different colors are used on this heatmap to represent the probability of void occurrence, marking potential void risk locations. Using feature importance analysis from the model, the dominant physical factors leading to increased risk at each location are automatically correlated and reported.
[0056] Secondly, a collaborative optimization scheme for the machine process parameters is generated. Using a trained prediction model as a surrogate model, the combination of process parameters that minimizes the total defect risk of the entire wafer is searched in the parameter space through gradient ascent or genetic algorithms.
[0057] Finally, real-time early warning and interception logic is triggered. In actual production, the method can continuously perform online prediction of the wafers being bonded. When the predicted local risk value exceeds the safety limit, the system automatically sends an interception signal to the machine or adjusts subsequent process parameters accordingly, realizing closed-loop control from defect detection to defect prevention.
[0058] like Figure 6 The diagram illustrates a comparison of the detection performance between the method provided by this invention and existing technologies. To verify the effectiveness of the method, tests were conducted under a pre-defined experimental environment. This environment included a high-performance computing workstation equipped with a 3.5 GHz processor and a 24 GB graphics processor; the dataset consisted of measured data from an advanced packaging production line, encompassing 300 pairs of heterogeneous wafer samples processed using a hybrid bonding process, including confirmed defect labels identified by scanning acoustic microscopy (CSAM) and cross-sectional scanning electron microscopy (SEM) after thermal annealing. The comparative experiments were conducted on the same hardware platform and test set, with the pad spacing of the test samples all less than 10 micrometers. Experimental results show that traditional CNN methods achieve a detection accuracy of 72% and a recall of 65%, but they struggle to effectively capture cross-interface micro-topographical relationships. Optical detection methods achieve an accuracy of 68% and a recall of 60%, but are limited by optical penetration depth and cannot identify latent voids within the bonding interface. Single-modal methods achieve an accuracy of 78% and a recall of 72%, and although they incorporate graph structures, they lack coupling with multimodal physical fields. In contrast, the method of this invention achieves a detection accuracy of 94% and a recall of 91%. Experimental data demonstrate that by coupling multimodal heterogeneous data with a graph neural network architecture based on physical constraints, this invention improves the accuracy and generalization ability for identifying latent interface voids in hybrid bonding processes, verifying the advanced nature of the technical solution.
[0059] Example 2 This embodiment provides a wafer defect detection device based on multimodal data and graph neural networks. (Refer to...) Figure 4 The device 400 includes: The dataset construction module 401 is configured to acquire the microstructure data of the first and second wafers, the transient physical field time series data of the bonding machine, and the layout logic features; the dataset construction module is also configured to perform normalization processing, time series resampling, and terrain missing data completion on the feature data of each dimension to generate a heterogeneous dataset for subsequent model input.
[0060] The dual-graph topology modeling module 402 is configured to abstract the pads of the first wafer and the second wafer as graph nodes, and define the edges of the graph according to the physical proximity relationship between the pads and the exhaust path, thereby constructing a first feature graph and a second feature graph; the dual-graph topology modeling module is also configured to perform multi-layer topological convolution operation on the feature graph to aggregate the spatial association features between each pad and its neighboring pads, and generate node embedding vectors.
[0061] The cross-interface interaction module 403 is configured to calculate the spatial projection distance between the node sets of the first wafer and the second wafer, and to establish an interaction edge when the projection overlap area exceeds a preset threshold, thereby constructing a cross-interface interaction bipartite graph; the cross-interface interaction module is also configured to use a multi-head cross-attention mechanism to calculate the similarity between nodes to obtain the interface morphology sensing weight, and to combine the transient intercepted gas pressure evolution operator to generate a feature vector representing the interface bonding state.
[0062] The physics field projection module 404 is configured to construct an energy projection model using a physical information neural network, mapping the global transient physics field time series data of the machine to each interaction edge of the cross-interface interaction bipartite graph, and calculating the local equivalent contact pressure of each pad node pair under pressure time series.
[0063] The defect prediction module 405 is configured to capture the evolution trajectory of physical interaction features at the bonding moment using a convolutional long short-term memory network, establish a reverse causal tracing mechanism through contrastive learning, and finally use a deep classifier with a spatial pyramid pooling layer to calculate the probability score of latent interface voids.
[0064] The output module 406 is configured to generate a full-wafer defect risk heatmap based on the occurrence probability score, and use a trained deep classifier as a surrogate model to search in the parameter space for pressure time-series curves and temperature compensation strategies that minimize the total defect risk of the entire wafer, thereby generating a collaborative optimization scheme for the machine process parameters.
[0065] Example 3 This embodiment provides the implementation process of the method of the present invention in a specific semiconductor manufacturing scenario, which is used to further illustrate the specific application of the technical solution of the present invention in solving the prediction of latent voids at heterogeneous interfaces.
[0066] In this embodiment, the objects to be processed are a high-speed logic wafer (first wafer) using a 7-nanometer process and a three-dimensional stacked memory wafer (second wafer). The center-to-center spacing between the pads of the first wafer and the second wafer is 8 micrometers, and the diameter of the copper pads is 4 micrometers.
[0067] The specific implementation process is as follows: Step S100 is executed to construct a multimodal heterogeneous dataset. Atomic force microscopy is used to perform a full scan of the surfaces of the first and second wafers to acquire nanoscale topographic relief data. In this embodiment, the average depth of the copper pad recesses on the first wafer is 4.2 nm, and the average depth of the copper pad recesses on the second wafer is 6.5 nm. Simultaneously, pressure timing data during the bonding cycle is acquired in real-time from the bonding machine, with the maximum total pressure set at 45 kN and the bonding time at 12 seconds. Furthermore, the GDSII layout files of the two wafer types are analyzed to extract the coordinate distribution of the pad array and the exhaust path parameters with a width of 1.5 μm. The above data is normalized and time-series aligned to generate a heterogeneous dataset.
[0068] Step S200 is executed to establish a dual-graph topology model. Based on the pad center coordinates, a first feature map and a second feature map are constructed for the first and second wafers respectively. In this embodiment, each feature map contains approximately 1.2 × 10^6 nodes. Adjacent pads are connected using the K-nearest neighbor algorithm, and logical connection edges are established based on the exhaust paths. Through a three-layer graph convolution operation, the local spatial correlation features of each node are aggregated to generate a 512-dimensional node embedding vector.
[0069] Step S300 is executed to perform cross-interface interaction and spatial alignment. Based on the global alignment residual with a precision of 50 nanometers, the vertical projection distance between the first and second wafer node sets is calculated. A cross-interface interaction bipartite graph is constructed. When the overlap area between the first and second wafer nodes in the projection direction is greater than 10 square micrometers, an interaction edge is established. The shape sensing weight is calculated using a cross-attention mechanism, and the transient trapped gas pressure evolution operator is invoked. This operator calculates the local gas pressure value generated during the pressure ramp-up stage based on a 1.5-micrometer-wide exhaust path and a permeability constant of 5.0 × 10^-14 square meters. The generated feature vector characterizes the interface bonding state.
[0070] Step S400 is executed to implement physical field energy projection. The 45 kN total pressure signal set by the machine is input into the physical information neural network. The physical information neural network, based on the elasticity equilibrium equation, projects the macroscopic pressure onto the interaction edge of the cross-interface interaction bipartite graph. The calculation results show that for some pad pairs in the wafer edge region, the local equivalent contact pressure is only 8.2 MPa, which is lower than the preset theoretical bonding threshold.
[0071] Step S500 is executed to predict latent defects. The feature vector representing the interface bonding state and the local equivalent contact pressure are input into the spatiotemporal evolution memory network. The convolutional long short-term memory network captures the microscopic contact evolution stagnation phenomenon caused by insufficient local pressure during the constant pressure stage. The reverse causal tracing module identifies that this anomaly is caused by a severe mismatch between the local height bulge of the first wafer and the concave features of the second wafer. The classifier finally outputs the overall risk probability distribution, identifying three high-risk latent void zones in the edge region of the first wafer, all with an occurrence probability score higher than 0.88.
[0072] Execute step S600 to output the detection results and closed-loop optimization. Generate a high-resolution full-image defect risk heat map of the wafer. Based on the above prediction results, use a surrogate model to search for process parameters and generate an optimization scheme. This scheme recommends increasing the peak pressure of subsequent wafers in this batch to 52 kN and increasing the temperature compensation in the edge region by 15 degrees Celsius. After applying this optimization scheme, acoustic testing after thermal annealing verified that the incidence of interface voids was significantly reduced.
[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A wafer defect detection method based on multimodal data and graph neural networks, characterized in that, The method includes the following steps: Step 1: Construct a multimodal heterogeneous dataset for heterogeneous dual wafers, including microstructure data of the first and second wafers, transient physical field timing data of the bonding machine, and layout logic features. Step 2: Establish a wafer space topology model based on a dual-graph structure to obtain a first feature map and a second feature map, wherein pads are abstracted as graph nodes and the edges of the graph are defined according to physical proximity and exhaust paths; perform topological convolution operation on the first feature map and the second feature map to aggregate spatial correlation features and generate node embedding vectors that represent local background information. Step 3: Design a cross-interface bipartite interaction and spatial alignment mechanism, including: establishing a spatial topological alignment model and calculating the spatial projection relationship between the two wafer node sets; constructing a cross-interface interaction bipartite graph and establishing interaction edges based on the projection overlap; using a cross-attention mechanism to calculate similarity in order to obtain the interface shape sensing weight that characterizes the degree of interface shape mismatch, and generating a feature vector that characterizes the interface bonding state. Step 4: Implement energy projection from the global physical field to local grid points. Use the physical field mapping operator to project the transient physical field time series data onto the interaction edges of the cross-interface interaction bipartite graph to generate local physical features for characterizing the pad-level micro-physical state. Step 5: Predict latent defects based on spatiotemporal evolution memory. Input the feature vector and the local physical features into the spatiotemporal evolution memory network for feature enhancement to generate a composite feature vector. Then, use a classifier to calculate the probability score of latent interface voids based on the composite feature vector. Step 6: Output the detection results and generate a full-view defect risk heat map of the wafer based on the occurrence probability score.
2. The method according to claim 1, characterized in that, In step one, the micro-morphology data includes the exhaust path width; in step two, the initial feature vector of the graph node includes the center coordinates of the pad, the recess depth, the local morphology slope, and the material properties; the weight features of the graph edge include the Euclidean distance between the pad centers and the effective cross-sectional area of the exhaust channel; in step three, calculating the spatial projection relationship includes calculating the three-dimensional spatial projection distance between the node sets of the first wafer and the second wafer in the pressed state; the step of establishing interactive edges based on the projection overlap includes establishing interactive edges when the physical overlap area between nodes exceeds a preset threshold; calculating the similarity between nodes includes using the node embedding vector of the first wafer as the query and the node embedding vector of the second wafer as the key and value, and obtaining the interface morphology sensing weight by calculating the dot product similarity.
3. The method according to claim 1, characterized in that, The step of constructing a multimodal heterogeneous dataset for the first and second wafers further includes: preprocessing the microstructure data, the transient physical field time series data, and the layout logic features. The preprocessing includes scaling each feature data to a uniform numerical range using a normalization method, and using time series resampling technology to ensure that the sampling rate of the microstructure data and the sampling rate of the transient physical field time series data are consistent in the time dimension.
4. The method according to claim 1, characterized in that, The steps of implementing energy projection from the global physical field to local grid points specifically include: constructing an energy projection model using a physical information neural network, wherein the loss function of the physical information neural network includes a data fitting term and a physical constraint residual term; training the model with minimizing the physical constraint residual term as the optimization objective to learn the mapping relationship that projects the global total pressure onto each interaction edge of the cross-interface interaction bipartite graph, and calculating the local equivalent contact pressure of each pad node under the pressure time series in the transient physical field time series data as the local physical feature.
5. The method according to claim 4, characterized in that, The physical constraint residuals are based on physical constraint conditions including constraint terms constructed based on interface mechanical equilibrium equations and heat conduction equations.
6. The method according to claim 1, characterized in that, The steps for predicting latent defects based on spatiotemporal evolutionary memory specifically include: constructing the spatiotemporal evolutionary memory network, which integrates a temporal convolutional network and a convolutional long short-term memory network to learn and store the evolutionary trajectory of the physical interaction features at the bonding moment; introducing a reverse causal tracing module based on contrastive learning, which learns a discriminative subspace in the feature space that can distinguish between defective samples and normal samples by comparing and analyzing known defective samples and normal samples, so as to identify the core combination of causes that lead to the formation of latent defects.
7. The method according to claim 1, characterized in that, The steps of designing the cross-interface bipartite interaction and spatial alignment mechanism further include: when updating the edge features of the cross-interface interaction bipartite graph, introducing a transient trapped gas pressure evolution operator, which is coupled with the exhaust path width in the micro-morphology data and the permeability constant simplified based on Darcy's law, to calculate the transient trapped gas pressure value generated under the dynamic compression sequence, and the transient trapped gas pressure value forms a nonlinear cancellation relationship with the feature vector representing the interface bonding state.
8. The method according to claim 1, characterized in that, The step of outputting the detection results also includes: using a trained deep classifier as a surrogate model, and using gradient ascent or genetic algorithm to search in the parameter space for a combination of pressure time-series curves and temperature compensation strategies that minimize the total defect risk of the entire wafer, thereby generating a collaborative optimization scheme for the machine process parameters.
9. The method according to claim 3, characterized in that, The preprocessing also includes: for missing data points in the microscopic morphology data, using a bicubic interpolation algorithm or a deep learning-based super-resolution reconstruction technique to complete the data.
10. A wafer defect detection device based on multimodal data and graph neural networks, characterized in that, include: The dataset construction module is configured to acquire microstructure data of the first and second wafers, transient physical field timing data of the bonding machine, and layout logic features, and generate a heterogeneous dataset. The dual-graph topology modeling module is configured to abstract pads as graph nodes and define graph edges based on physical proximity and exhaust paths, generating node embedding vectors through topological convolution operations. The cross-interface interaction module is configured to calculate the spatial projection relationship between the two wafer node sets and construct the cross-interface interaction bipartite graph, and use the cross attention mechanism to generate feature vectors representing the interface bonding state. The physics projection module is configured to use a physical information neural network to construct an energy projection model, projecting transient physical field time series data onto the interaction edges of a cross-interface interactive bipartite graph to generate local physical features. The defect prediction module is configured to use a spatiotemporal evolution memory network for feature enhancement and to calculate a probability score for the occurrence of latent interface holes. The output module is configured to generate a full-map defect risk heat map of the wafer based on the occurrence probability score, and to generate a collaborative optimization scheme for the machine process parameters.