Graph-free wafer defect detection method using ultrasonic multi-frequency signals

Abstract

The application discloses a kind of ultrasonic multi-frequency signal's no-pattern wafer defect detection method, belong to semiconductor manufacturing detection technical field, including the synchronous acquisition of each detection point's multi-frequency ultrasonic signal and at least one auxiliary signal, pre-processing constructs multi-channel feature vector.The feature vector is processed by probability fusion, and the probability fusion feature distribution parameterized by mean vector and variance vector is generated.Next, sampling is carried out from the distribution and multiple random inferences are carried out, and the defect category, prediction confidence and fusion uncertainty measure of each point are calculated.For high uncertainty points, the contribution difference of multi-frequency ultrasonic and auxiliary signals to uncertainty is analyzed inversely.Based on the prediction confidence statistics of all points, a dynamic threshold is determined, the uncertainty area is identified by combining the contribution difference, and a final quality judgment report is generated by applying differentiated rules.The application realizes intelligent information fusion, internal quantification and adaptive decision of reliability, and improves the robustness of detection.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing and testing technology, and in particular to a method for detecting patternless wafer defects using ultrasonic multi-frequency signals. Background Technology

[0002] In the semiconductor manufacturing process, unpatterned wafers (bare silicon wafers) are the cornerstone of integrated circuits, and the uniformity and integrity of their internal materials directly determine the performance and yield of the final chip. Therefore, before incorporating patterning processes, efficient and accurate non-destructive testing of defects such as micropores, cracks, and inclusions within the wafer is a key step in controlling front-end process quality and reducing production costs.

[0003] Currently, the industry primarily relies on technologies based on single physical principles for internal defect detection in patternless wafers. Among these, ultrasonic testing technology has become an important detection method due to its sensitivity to internal material discontinuities and its certain penetration depth. Traditional ultrasonic testing often uses a single center frequency, and its detection effectiveness is limited by frequency characteristics: high-frequency ultrasound has high resolution for surface and near-surface micro-defects but weak penetration and rapid signal attenuation; low-frequency ultrasound has strong penetration but insufficient sensitivity to micro-defects. This contradiction between frequency and defect characteristics often makes it difficult to balance detection rate and signal-to-noise ratio when dealing with diverse defects of different depths and scales using single-frequency ultrasonic testing.

[0004] To improve detection capabilities, the fusion detection approach using multi-frequency ultrasonic excitation or multi-modal sensing (such as combining ultrasound and thermal waves) has begun to attract attention. Most existing fusion strategies are at the stage of simple data layer concatenation or fixed-weight feature layer fusion. These methods have significant limitations: First, they fail to establish a deep correlation model between different frequencies or modal signals and the physical characteristics of defects. When multiple source signals represent the same area differently or seem contradictory (e.g., abnormal echoes at one frequency while normal echoes at others, or inconsistent indications between ultrasonic and thermal signals), simple fusion methods struggle to make reliable judgments, easily leading to misjudgments or missed detections. Second, existing methods are essentially deterministic systems, outputting a "hard" decision result, unable to self-assess the confidence level or reliability of the current detection judgment. When faced with "unfamiliar" defect patterns with extremely low signal-to-noise ratios, indistinct defect features, or insufficient coverage by training data, the system may still provide a definite but potentially erroneous result, lacking a risk warning mechanism and posing potential risks to quality control. Furthermore, defect determination usually relies on a pre-set fixed threshold, which is difficult to adapt to the normal fluctuations in material properties, surface conditions and background noise of wafers from different manufacturers and batches. The generalization ability and stability of the detection standard need to be improved.

[0005] Therefore, it is necessary to provide a method for detecting patternless wafer defects using ultrasonic multi-frequency signals to solve the above-mentioned technical problems. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for detecting patternless wafer defects using ultrasonic multi-frequency signals. It introduces a probabilistic fusion feature distribution and uncertainty decoupling mechanism, thereby achieving highly robust detection and accurate determination of patternless wafer defects.

[0007] This invention provides a method for detecting patternless wafer defects using ultrasonic multi-frequency signals, the method comprising the following steps: For each detection point on the wafer, its multimodal raw signal is acquired and preprocessed to construct a multi-channel feature vector for each detection point. The multimodal raw signal includes a multi-frequency ultrasonic signal and at least one auxiliary signal. The multi-channel feature vectors are subjected to probabilistic fusion processing to obtain the probabilistic fusion feature distribution corresponding to each detection point, wherein the probabilistic fusion feature distribution is parameterized by the mean vector and the variance vector. Multiple feature samples are sampled from the probabilistic fusion feature distribution of each detection point, and multiple random forward propagation inferences are performed on the multiple feature samples to calculate the defect category prediction result, prediction confidence and fusion uncertainty measure of each detection point. For detection points where the fusion uncertainty measure exceeds a preset warning value, the corresponding variance vector and mean vector are inverted to decouple and quantify the difference in contribution of the multi-frequency ultrasonic signal and the auxiliary signal to the fusion feature uncertainty of the detection point. Based on the comparison results of the fusion uncertainty metric and dynamic threshold at each detection point, and combined with the contribution difference, uncertainty regions are identified and a final quality assessment report for the wafer is generated, wherein the dynamic threshold is determined by the statistical distribution of the predicted confidence scores of all detection points.

[0008] Preferably, the step of acquiring and preprocessing the multimodal raw signals of each detection point on the wafer to construct a multi-channel feature vector for each detection point includes: For each detection point, multi-frequency ultrasonic signals and auxiliary signals are collected simultaneously, and the echoes of each independent frequency channel of the multi-frequency ultrasonic signals are preliminarily conditioned. For the echoes of each frequency channel after preliminary conditioning, feature transformation is performed to extract time-domain correlation features that characterize the difference between the signal of each channel and the reference state, forming an ultrasound feature group; The ultrasound feature set is fused and spliced ​​with the auxiliary features extracted from the auxiliary signal to generate a multi-channel feature vector for the detection point.

[0009] Preferably, the step of performing probabilistic fusion processing on the multi-channel feature vectors to obtain the probabilistic fusion feature distribution corresponding to each detection point includes: Ultrasonic feature components and auxiliary feature components are extracted from the multi-channel feature vectors respectively, and their respective modal built-in confidence is evaluated; Based on the modality built-in confidence, dynamic fusion weights are assigned to the ultrasound feature components and auxiliary feature components; The ultrasonic feature components and auxiliary feature components are weighted and fused using the dynamic fusion weights to form intermediate fused features; Based on the intermediate fusion features, a probabilistic fusion feature distribution parameterized by the mean vector and the variance vector is generated through a pre-set probability mapping network.

[0010] Preferably, the step of sampling multiple feature samples from the probabilistic fusion feature distribution of each detection point, and performing multiple random forward propagation inferences on the multiple feature samples to calculate the defect category prediction result, prediction confidence, and fusion uncertainty measure for each detection point includes: Based on the variance vector of the probabilistic fusion feature distribution, importance sampling is performed on each detection point to obtain multiple feature samples. Each feature sample is then input into a pre-set Bayesian neural network classifier for multiple random forward propagations to obtain multiple intermediate prediction probability distributions for each feature sample. The multiple intermediate prediction probability distributions of each feature sample are integrated to obtain the preliminary prediction results and cognitive uncertainty estimates of each sample. The preliminary prediction results of all feature samples are then integrated again to generate the defect category prediction results and prediction confidence of the detection point. The fusion uncertainty measure of the detection point is calculated by combining the cognitive uncertainty estimate of each feature sample with the variance vector of the probabilistic fusion feature distribution.

[0011] Preferably, for detection points where the fusion uncertainty metric exceeds a preset warning value, the corresponding variance vector and mean vector are inverted to decouple and quantify the difference in contribution of the multi-frequency ultrasonic signal and the auxiliary signal to the fusion feature uncertainty of the detection point, including: Based on the variance vector, calculate the relative gradient contribution of the feature coding channels of the multi-frequency ultrasound signal and the feature coding channels of the auxiliary signal in each dimension of the fusion feature distribution; Based on the relative contribution of the gradient, source weights of uncertainty are assigned to the ultrasound feature components and the auxiliary feature components; The variance vector is weighted and decomposed using the source weights to obtain the contribution of the uncertainty components corresponding to the multi-frequency ultrasound signal and the auxiliary signal, respectively, and the contribution difference is obtained.

[0012] Preferably, the contribution difference is the ratio or difference between the contribution of the uncertainty component of the multi-frequency ultrasound signal and the contribution of the uncertainty component of the auxiliary signal.

[0013] Preferably, the step of identifying uncertainty regions and generating a final quality assessment report for the wafer based on the comparison results of the fusion uncertainty metric and dynamic threshold at each detection point, combined with the contribution difference, includes: The fusion uncertainty measure of each detection point is compared with the dynamic threshold, and the detection points with the fusion uncertainty measure higher than the dynamic threshold are initially screened to form a candidate uncertainty point set; For each detection point in the candidate uncertainty point set, determine whether the corresponding contribution difference exceeds a preset difference threshold; From the set of candidate uncertainty points, detection points whose contribution differences exceed the difference threshold are selected, and clustering is performed based on the spatial distribution of the detection points. The region formed by points that meet the preset conditions in space is identified as the final uncertainty region. For the detection points within the final uncertainty region, quality judgment rules are applied to generate a final quality judgment report containing regional judgment conclusions.

[0014] Preferably, the dynamic threshold is determined by the statistical distribution of the predicted confidence scores of all detection points, specifically: Calculate the mean and standard deviation of the predicted confidence scores for all detection points; The dynamic threshold is calculated based on a linear combination of the mean and the standard deviation, wherein the dynamic threshold is positively correlated with the mean and negatively correlated with the standard deviation.

[0015] Compared with related technologies, the ultrasonic multi-frequency signal-based method for detecting patternless wafer defects provided by this invention has the following advantages: This invention provides a method for detecting patternless wafer defects using ultrasonic multi-frequency signals. By introducing a probabilistic fusion feature distribution, the mean vector is used to represent the central trend of the features, and the variance vector is used to represent the uncertainty of the features. This enables probabilistic modeling of the fusion process of multimodal signals, effectively capturing the uncertainty information brought about by signal noise and feature variation. By sampling from the probabilistic fusion feature distribution and performing multiple random forward propagation inferences, a fusion uncertainty metric is calculated, achieving a quantitative assessment of the reliability of the prediction results for detection points, avoiding the blind interpretation of traditional deterministic models in ambiguous regions. Furthermore, for high-uncertainty detection points, by inverting the variance vector and mean vector, the contribution difference between the multi-frequency ultrasonic signal and the auxiliary signal to the uncertainty is decoupled and quantified, enabling the tracing of the uncertainty source. This helps identify the specific modal source causing the uncertainty and provides accurate data support for subsequent process analysis. Finally, by combining the fusion uncertainty metric and dynamic threshold, and introducing contribution difference for dual screening and clustering, the uncertainty region can be accurately identified, significantly reducing the false positive and false negative rates, and improving the robustness and accuracy of patternless wafer defect detection. Attached Figure Description

[0016] Figure 1 The flowchart illustrates a method for detecting patternless wafer defects using ultrasonic multi-frequency signals, as provided by this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the drawings, not all structures. Moreover, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0018] It should also be noted that, for ease of description, the accompanying drawings show only the parts relevant to the invention and not all of them. Before discussing exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the drawings. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0019] This invention provides a method for detecting patternless wafer defects using ultrasonic multi-frequency signals, with reference to... Figure 1 As shown, the detection method includes the following steps: S1: For each detection point on the wafer, acquire its multimodal raw signal and preprocess it to construct a multi-channel feature vector for each detection point. The multimodal raw signal includes multi-frequency ultrasonic signal and at least one auxiliary signal.

[0020] Specifically, step S1 includes the following steps: S11: For each detection point, simultaneously acquire its multi-frequency ultrasonic signal and auxiliary signal, and perform preliminary conditioning on the echo of each independent frequency channel of the multi-frequency ultrasonic signal.

[0021] In this embodiment, the goal of this step is to synchronously acquire the multimodal raw signals of each detection point on the wafer and to perform preliminary physical layer conditioning on the core multi-frequency ultrasonic signals.

[0022] In practical implementation, an integrated scanning and inspection platform must first be built. The core of this platform is a scanning head equipped with a confocal ultrasonic probe and auxiliary sensors (such as laser thermal wave probes or microwave probes). This scanning head is driven by a high-precision motion platform, enabling precise movement along grating or spiral paths on the wafer plane. The step resolution is typically set between 5 and 20 micrometers to achieve comprehensive coverage. During inspection, the patternless wafer is placed on a temperature-controlled vacuum adsorption stage to ensure stable coupling.

[0023] For each stationary detection point, the system's main controller sends a synchronization trigger signal. The programmable pulse transmitter / receiver in the ultrasonic module sequentially transmits at least three narrow ultrasonic pulses with discrete center frequencies (e.g., 50MHz, 100MHz, 200MHz) to excite the probe to generate ultrasonic waves. In a preferred embodiment of the invention, the ultrasonic probe and the corresponding scanning and acquisition system can be integrated or implemented using an ultrasonic scanning microscope. An ultrasonic scanning microscope is a mature non-destructive testing device that performs precise scanning on the sample surface using a high-frequency ultrasonic probe and acquires the vertical incident echo signal at each point, generating high-resolution internal acoustic images. This invention utilizes the platform of an ultrasonic scanning microscope and programs its transmission and reception circuits to achieve sequential transmission and synchronous high-fidelity acquisition of the multi-frequency ultrasonic signals.

[0024] The same probe receives the echo signal reflected from inside the wafer. This signal is immediately amplified by a low-noise preamplifier and then synchronously acquired and recorded by a high-speed data acquisition card (sampling rate usually needs to be ≥1GS / s) to obtain and record the complete time-domain waveform (A-scan signal) corresponding to each frequency.

[0025] Simultaneously, during the intervals of ultrasonic emission or according to a preset coding sequence, an auxiliary sensor (such as a pulsed laser) is triggered to excite the same point and its response signal (such as a surface thermal radiation change curve) is acquired synchronously. This step generates a set of raw, unprocessed signal data packets for each detection point.

[0026] The initial conditioning of multi-frequency ultrasound signals is mainly completed in the hardware and acquisition stages, including controlling the excitation waveform through a pulse generator, performing time gain compensation (TGC) on the gain of echoes at different depths through a variable gain amplifier (VGA) to optimize the dynamic range, and performing preliminary filtering using the anti-aliasing filter built into the acquisition system.

[0027] S12: Perform feature transformation on the echoes of each frequency channel after preliminary conditioning, extract the time-domain correlation features that characterize the difference between the signal of each channel and the reference state, and form an ultrasound feature group.

[0028] In this embodiment, this step aims to perform digital signal processing on the acquired raw signals, especially multi-frequency ultrasound signals, to extract more representative features in preparation for subsequent fusion.

[0029] The implementation process is completed in signal processing software or FPGA / DSP hardware. First, the raw ultrasound A-scan signal obtained in S11 is subjected to independent depth preprocessing for each frequency. For each frequency channel, the software applies a digital bandpass filter (such as a Chebyshev filter) centered on that frequency to further suppress noise outside that band.

[0030] Next, a Hilbert transform is performed on the filtered signal to calculate its analytic signal, and the amplitude of this analytic signal is extracted as the time-domain envelope Env_f(t). This step compresses the data and converts phase information into amplitude information, making the defect reflection peaks more prominent.

[0031] Then, the extracted envelope signal is normalized, for example, by linearly scaling its amplitude range to [a specific value]. The interval is defined to eliminate absolute amplitude differences caused by factors such as minor fluctuations in excitation energy. Finally, feature calculations are performed: a set of time-domain correlation features are calculated from the preprocessed envelope signal. For example, a time window containing a bottom echo can be extracted, and the cross-correlation coefficient (or differential and integral features) between the signal within this window and a reference envelope signal obtained from the same location on a known defect-free "gold reference" wafer can be calculated. Each frequency channel will generate one or a set of such correlation feature values. Arranging the feature values ​​calculated from all frequency channels in order constitutes the "ultrasonic feature set" for the detection point, which is in the form of a feature vector.

[0032] S13: The ultrasound feature group is fused and spliced ​​with the auxiliary features extracted from the auxiliary signal to generate a multi-channel feature vector of the detection point.

[0033] In this embodiment, the goal of this step is to integrate the features extracted from signals based on different physical principles into a unified mathematical representation, namely a multi-channel feature vector, for subsequent processing by machine learning models.

[0034] During implementation, auxiliary signals need to be processed in parallel. For synchronously acquired auxiliary signals (such as photothermal attenuation curves), corresponding feature extraction is required based on their physical meaning. For example, for photothermal signals, their attenuation curves can be exponentially fitted to extract features such as the thermal diffusion time constant and the initial temperature rise amplitude; for microwave signals, features of amplitude or phase changes with frequency can be extracted. This process generates an "auxiliary feature vector".

[0035] Finally, feature concatenation (fusion) is performed. The ultrasound feature vector obtained in S12 and the auxiliary feature vector obtained in this step are concatenated along the feature dimension. This concatenated vector contains all the preprocessed feature information from multi-frequency ultrasound and auxiliary sensing, and serves as the direct input for the probabilistic fusion processing in the subsequent step S2. In software implementation, this concatenation operation can be easily implemented using a general numerical computing library (such as NumPy).

[0036] S2: Perform probabilistic fusion processing on the multi-channel feature vectors to obtain the probabilistic fusion feature distribution corresponding to each detection point, wherein the probabilistic fusion feature distribution is parameterized by the mean vector and the variance vector.

[0037] Specifically, step S2 includes the following steps: S21: Extract the ultrasound feature components and auxiliary feature components from the multi-channel feature vector respectively, and evaluate their respective modal built-in confidence.

[0038] In this embodiment, the goal of this step is to separate feature components from different physical principles from a unified multi-channel feature vector and to preliminarily evaluate the reliability of each component itself, i.e., modal built-in confidence.

[0039] In practice, the first step is to follow the pre-agreed feature arrangement order when constructing the multi-channel feature vector. For example, the first half of the feature vector might come from the ultrasound signal, and the second half from the auxiliary signal. Following this rule, the input feature vector is split into two independent parts: the ultrasound feature component and the auxiliary feature component. Next, each of these split feature components is input into a small neural network specifically designed for confidence assessment. This assessment network analyzes the intrinsic quality of the feature component, such as signal clarity and consistency between features across dimensions, and outputs a confidence score between zero and one. The closer the score is to one, the higher the quality and reliability of the feature component; the closer the score is to zero, the more likely it is to be affected by noise or have incomplete information, resulting in lower reliability.

[0040] S22: Based on the modality built-in confidence, assign dynamic fusion weights to the ultrasound feature components and auxiliary feature components.

[0041] In this embodiment, this step aims to dynamically allocate the weights of the ultrasound feature components and auxiliary feature components in the subsequent fusion process based on their respective confidence scores evaluated in the previous step; that is, dynamic fusion weights. The core of this mechanism is a weight calculation module based on a normalized exponential function. This module receives two confidence scores as input.

[0042] First, each confidence score is processed by an exponential function, amplifying the difference between high and low scores. Then, the two processed values ​​are summed, and the proportion of each value to this sum is calculated. This proportion is the final dynamic fusion weight. Through this calculation, feature components with higher confidence scores automatically receive a larger fusion weight, meaning their information will be given more importance in the final decision; while feature components with lower confidence scores have a correspondingly smaller weight. The sum of the two weights remains constant at one, ensuring a reasonable fusion ratio.

[0043] S23: The ultrasonic feature components and auxiliary feature components are weighted and fused using the dynamic fusion weights to form an intermediate fusion feature.

[0044] In this embodiment, this step utilizes the dynamic fusion weights calculated in the previous step to weight and merge the ultrasound feature components and auxiliary feature components, generating a single, fused intermediate feature representation. This is a direct calculation process.

[0045] The specific operation involves multiplying each value in the ultrasound feature component by its corresponding dynamic fusion weight; simultaneously, each value in the auxiliary feature component is also multiplied by its corresponding dynamic fusion weight. Then, these two weighted and scaled feature vectors are summed along each corresponding dimension. The resulting new vector is the "intermediate fusion feature." It contains information from both modalities and reflects the contribution ratio based on their respective reliability assessments, with higher-quality and more reliable features dominating.

[0046] S24: Based on the intermediate fusion features, a probabilistic fusion feature distribution parameterized by the mean vector and the variance vector is generated through a preset probability mapping network.

[0047] In this embodiment, this step is the core of the probabilistic fusion process. The goal is to map the determined intermediate fusion features into a probability distribution that can simultaneously describe the most likely value of the fused features and its uncertainty.

[0048] This is achieved through a neural network called a "probabilistic mapping network." This network takes intermediate fused features as input and has two parallel output paths. The first output path generates a "mean vector," which represents the most likely central or expected value of the fused feature in each aspect. The second output path generates a "log-variance vector." To ensure the output uncertainty is positive, this path first outputs the log-variance and then converts it into an actual "variance vector" using an exponential function. The magnitude of each element in the variance vector directly quantifies the uncertainty or fluctuation range of the fused feature in the corresponding dimension; a larger value indicates greater uncertainty in the model's judgment of that feature dimension. This uncertainty may stem from conflicts in multi-source information, signal noise, or limitations in the model's own knowledge. Ultimately, the output of step S2 is this probabilistic fused feature distribution parameterized by both the mean and variance vectors.

[0049] S3: Sample multiple feature samples from the probabilistic fusion feature distribution of each detection point, and perform multiple random forward propagation inferences on the multiple feature samples to calculate the defect category prediction result, prediction confidence and fusion uncertainty measure of each detection point.

[0050] Specifically, step S3 includes the following steps: S31: Based on the variance vector of the probabilistic fusion feature distribution, importance sampling is performed on each detection point to obtain multiple feature samples. Each feature sample is then input into a pre-set Bayesian neural network classifier for multiple random forward propagations to obtain multiple intermediate prediction probability distributions for each feature sample.

[0051] In this embodiment, the goal of this step is to efficiently sample from the probabilistic fusion feature distribution generated in step S2, and to use a Bayesian classifier to perform multiple random inferences on each sample to obtain a preliminary classification probability distribution.

[0052] In practice, the first step is "importance sampling." The input is the probabilistic fusion feature distribution corresponding to the current detection point, which is defined by the mean vector and the variance vector. During sampling, it is not simply random sampling, but the sampling process is guided by the magnitude of uncertainty in each dimension represented by the variance vector.

[0053] For dimensions with high variance (i.e., feature dimensions with high uncertainty), the sampling algorithm tends to generate more diverse values ​​in that dimension to fully explore the various possibilities caused by the ambiguity of information; for dimensions with low variance, the sampling is relatively concentrated. In this way, a predetermined number (e.g., K=10) of feature samples are extracted from the distribution.

[0054] Next, each sampled feature is input into a pre-trained Bayesian neural network classifier. The core feature of this classifier is that its internal dropout function is enabled during inference, even when the network is in "training mode" to introduce randomness. For each input feature sample, the classifier performs multiple (e.g., T=20) independent forward propagation calculations. Due to the randomness of Dropout, each forward propagation yields a slightly different instantaneous network state, resulting in a class prediction probability distribution for the same input. Ultimately, T intermediate prediction probability distributions are obtained for each feature sample; these distributions reflect the prediction fluctuations caused by the randomness of the classifier model weights.

[0055] S32: Integrate multiple intermediate prediction probability distributions for each feature sample to obtain preliminary prediction results and cognitive uncertainty estimates for each sample, and perform secondary integration of the preliminary prediction results for all feature samples to generate defect category prediction results and prediction confidence for the detection point.

[0056] In this embodiment, this step aims to perform a two-level integration of the large number of intermediate prediction results generated in S31 to obtain a stable final prediction and estimate the cognitive uncertainty of the model itself.

[0057] The implementation is divided into two levels. First, "intra-sample ensemble" is performed. For each feature sample sampled from the probability distribution, its corresponding T intermediate predicted probability distributions are ensembled. Specifically, the T probability distributions are arithmetically averaged over each class to obtain an "ensemble predicted probability distribution," which serves as the "preliminary prediction result" for that feature sample. Simultaneously, the statistical differences between these T distributions can be calculated (e.g., by calculating their average entropy, or by calculating the variance of the probabilities of the T distributions over each class), which serves as the "cognitive uncertainty estimate" for that feature sample. This estimate quantifies the degree of disagreement within the classifier model itself regarding this specific feature sample.

[0058] Then, a "cross-sample ensemble" is performed. For all K feature samples sampled at the current detection point, their respective "preliminary prediction results" (i.e., the K ensemble prediction probability distributions) are arithmetically averaged again. The probability distribution obtained from this second average is the final "defect category prediction result" for that detection point, where the category with the highest probability is the predicted defect type, and its highest probability value is the "prediction confidence." This two-level ensemble (first ensemble on model stochasticity, then ensemble on input distribution sampling) fully utilizes the idea of ​​Bayesian inference, combining model uncertainty and data uncertainty to obtain robust predictions.

[0059] S33: By combining the cognitive uncertainty estimate of each feature sample with the variance vector of the probabilistic fusion feature distribution, the fusion uncertainty measure of the detection point is calculated.

[0060] In this embodiment, the goal of this step is to calculate a comprehensive "fusion uncertainty metric" that captures both the uncertainty from the input features (represented by the variance vector in step S2) and the uncertainty from the classification model.

[0061] In practice, two aspects of information need to be integrated. The first aspect is the "cognitive uncertainty estimate" calculated in S32, corresponding to each feature sample, which represents the knowledge blind spot of the classification model. The average cognitive uncertainty estimate of all K feature samples can be averaged to obtain the average cognitive uncertainty at the model level. The second aspect is the "variance vector" of the probabilistic fusion feature distribution directly output in step S2, which directly quantifies the inherent ambiguity and conflict degree after multimodal feature fusion. A simple approach is to calculate the average of all elements of this variance vector to obtain the fusion uncertainty at the feature level. Finally, the "fusion uncertainty measure" of the detection point can be calculated by weighted summation of the above two averages, for example: Fusion uncertainty measure = a * (average feature variance) + (1-a) * (average cognitive uncertainty), where a is a preset weight coefficient (e.g., 0.5). Another more complex approach is to couple the variance vector with the cognitive uncertainty of each feature sample during the sampling calculation process. However, the core idea remains the same: the "fusion uncertainty measure" is a composite index that includes both the ambiguity generated during the fusion of the original multimodal signals and the degree of confidence of the classification model in making judgments on the current fused features, thus comprehensively characterizing the overall reliability risk of this detection at this point.

[0062] S4: For detection points where the fusion uncertainty metric exceeds a preset warning value, the corresponding variance vector and mean vector are inverted to decouple and quantify the difference in contribution of the multi-frequency ultrasonic signal and the auxiliary signal to the fusion feature uncertainty of the detection point.

[0063] Specifically, step S4 includes the following steps: S41: Based on the variance vector, calculate the relative gradient contribution of the feature coding channels of the multi-frequency ultrasound signal and the feature coding channels of the auxiliary signal in each dimension of the fusion feature distribution.

[0064] In this embodiment, the objective of this step is to trace and quantify the extent to which the uncertainty (characterized by the variance vector) of the probabilistic fusion feature distribution generated in step S2 can be attributed to the feature coding channels from the multi-frequency ultrasound signal and the feature coding channels from the auxiliary signal.

[0065] The specific implementation employs a contribution analysis method based on gradient backpropagation. First, it is necessary to obtain the probabilistic mapping network that generates the probabilistic fusion feature distribution of the current detection point in step S2. The input to this network is the intermediate fusion feature generated in step S23, which is obtained by weighted fusion of the ultrasonic feature components and auxiliary feature components separated in step S21. During implementation, all network parameters are fixed, and the variance vector corresponding to the current detection point (which is an output of the probabilistic mapping network S24) is used as the target of the analysis.

[0066] The gradient of the variance vector with respect to the original ultrasound feature components and auxiliary feature components extracted in step S21 is calculated using an automatic differentiation technique (such as PyTorch). The calculated gradient is a vector, where each element represents the extent to which the final variance vector changes when the corresponding original feature component undergoes a small change.

[0067] Then, the L2 norm (i.e., the square root of the sum of the squares of all gradient values) is calculated for the gradient vectors corresponding to the ultrasound feature components and auxiliary feature components, respectively. These two norm values ​​represent the "relative gradient contribution" of the ultrasound signal feature coding channel and the auxiliary signal feature coding channel to the overall variance (i.e., uncertainty) of the final fused feature distribution. The larger the gradient norm, the more sensitive the feature changes of the signal source are to the magnitude of the final uncertainty, that is, the greater its "contribution".

[0068] S42: Based on the relative contribution of the gradient, assign uncertainty tracing weights to the ultrasound feature components and the auxiliary feature components.

[0069] In this embodiment, this step aims to assign a pair of "source-tracing weights" to the ultrasound feature component and the auxiliary feature component based on the two relative gradient contribution values ​​calculated in S41. These weights will be used to decompose the total uncertainty (variance vector) across the two signal sources. This is a normalization process to ensure that the sum of the two weights is 1, thereby clearly defining the contribution ratio. Specifically, the gradient norm of the ultrasound feature component obtained in S41 (denoted as...) is... The gradient norm of the auxiliary feature components (denoted as ) and the gradient norm of the auxiliary feature components (denoted as ) () as input.

[0070] First, to avoid zero or negative values ​​and to amplify the dominant contribution, these two gradient norms are usually exponentialized and calculated. and Then, the Softmax function is used for normalization to calculate the source weights: Source weights of ultrasound characteristic components

[0071] Source weights of auxiliary feature components

[0072] Obviously, + =1. Through this calculation, if the gradient contribution of the ultrasound signal... Much larger than the auxiliary signal ,but It will be close to 1. A value close to 0 indicates that the uncertainty of the current fusion feature should primarily be traced back to the ultrasound signal source; conversely, if the two are roughly equal, then each weight is close to 0.5. These two source tracing weights... and These are the coefficients obtained through allocation and used for uncertainty decomposition.

[0073] S43: Use the source tracing weight to perform weighted decomposition on the variance vector to obtain the contribution of the uncertainty components corresponding to the multi-frequency ultrasound signal and the auxiliary signal respectively, and obtain the contribution difference.

[0074] In this embodiment, this step utilizes the source tracing weights obtained in S42 to decompose the variance vector output in step S2, ultimately calculating the quantifiable contribution difference. Specifically, this is a linear decomposition and comparison process. First, a weighted decomposition is performed: the total variance vector of the probabilistic fusion feature distribution of the current detection point is obtained. Each element of this variance vector is then multiplied by the ultrasonic source tracing weights. and auxiliary traceability weights This results in two new vectors: The contribution vector of the uncertainty component of the ultrasound signal = variance vector × ; The contribution vector of the uncertainty component of the auxiliary signal = variance vector × ; These two vectors represent the magnitude of the estimated uncertainty "contributed" by their respective signal sources in each feature dimension. To obtain a scalar contribution index, the average (or L2 norm) of these two vectors is typically calculated to obtain the contribution of the total uncertainty component of the ultrasound signal. Contribution of total uncertainty components of auxiliary signals Finally, the contribution difference is calculated. Depending on the specific analytical needs, there are two standard methods for calculating this difference: 1. Ratio: Calculation / This ratio directly reflects the relative multiple relationship between the uncertainty contributions of the two signal sources. When the ratio is significantly greater than 1, it indicates that the ultrasound signal is the main source of uncertainty; when it is significantly less than 1, the auxiliary signal is the main cause.

[0075] 2. Difference: Calculation - This difference reflects the absolute disparity in the uncertainty contributions of the two signal sources. A positive difference indicates that the ultrasound signal contributes more, while a negative difference indicates that the auxiliary signal contributes more.

[0076] The calculated ratio or difference is the final quantified "contribution difference," which clearly reveals which signal mode the uncertainty of the fusion characteristics of the high-uncertainty detection point mainly originates from, thus providing clear directional information for result verification and process diagnosis.

[0077] S5: Based on the comparison results of the fusion uncertainty metric and dynamic threshold at each detection point, and combined with the contribution difference, identify the uncertainty region and generate the final quality judgment report of the wafer, wherein the dynamic threshold is determined by the statistical distribution of the predicted confidence of all detection points.

[0078] Specifically, step S5 includes the following steps: S51: Compare the fusion uncertainty measure of each detection point with the dynamic threshold, and initially screen out the detection points whose fusion uncertainty measure is higher than the dynamic threshold to form a candidate uncertainty point set.

[0079] In this embodiment, the goal of this step is to calculate an adaptive dynamic threshold and use this threshold to perform preliminary screening of all detection points to identify potential high-uncertainty points. The specific implementation is divided into two parts.

[0080] First, calculate the dynamic threshold. Obtain the "predicted confidence" values ​​calculated for all valid detection points on the currently inspected wafer in step S3, forming a set of confidence values. Perform statistical analysis on this set and calculate its arithmetic mean ( ) and standard deviation ( Dynamic threshold () ) is calculated using the following linear combination formula: ,in It is a preset positive coefficient (e.g., =1.5). The logic of this formula is: the threshold is positively correlated with the mean and negatively correlated with the standard deviation. When the overall prediction confidence is high and the distribution is concentrated (small standard deviation), the threshold will be higher, indicating that the judgment result is subject to stricter requirements; when the overall confidence is low or the distribution is dispersed (large standard deviation), the threshold will automatically decrease to adapt to the generally high uncertainty in this detection and avoid generating too many false alarms.

[0081] Next, preliminary screening is performed. Each detection point on the wafer is traversed, and its "fusion uncertainty metric" value calculated in step S3 is compared with the dynamically calculated threshold. Compare them. All fusion uncertainty metrics greater than [a certain value] will be considered. The detection points are recorded, and their spatial coordinates are collected to form a "candidate uncertainty point set". This point set contains all locations whose overall reliability risk exceeds the adaptive threshold in this detection.

[0082] S52: For each detection point in the candidate uncertainty point set, determine whether the corresponding contribution difference exceeds a preset difference threshold.

[0083] In this embodiment, this step aims to perform secondary filtering on the initially selected candidate points, and based on the "contribution difference" calculated in step S4, to distinguish high uncertainty points caused by severe conflicts in multimodal information.

[0084] Specifically, this is implemented as a threshold judgment process. For each detection point in the "candidate uncertainty point set" obtained in S51, the "contribution difference" value calculated in step S4 is read. As mentioned earlier, this difference value can be a ratio or a difference. A "difference threshold" needs to be preset. If the contribution difference is a ratio, the difference threshold can be set to, for example, 2.0 or 0.5 (depending on which side is dominant); if it is a difference, a scalar threshold is set according to the typical magnitude of the ratio and the difference.

[0085] Next, it is determined whether the absolute value of the contribution difference at this point exceeds a preset difference threshold. The purpose of this determination is as follows: if the contribution difference is very small, it indicates that the ultrasound and auxiliary signals contribute equally to the uncertainty, and the high uncertainty may stem from common problems such as low signal-to-noise ratio; if the contribution difference is large, it indicates that the uncertainty is significantly biased towards a certain signal source, suggesting significant conflicts between multimodal information or an anomaly in a certain signal source, and these points are more valuable for analysis. Detection points with contribution differences exceeding the difference threshold are marked as "high-value uncertainty points".

[0086] S53: From the set of candidate uncertainty points, filter out the detection points whose contribution difference exceeds the difference threshold, and cluster them based on the spatial distribution of the detection points to identify the region formed by the points that meet the preset conditions in space as the final uncertainty region.

[0087] In this embodiment, the goal of this step is to spatially aggregate the discrete, high-value uncertainty points that have passed the above filtering to identify continuous “uncertainty regions”, since process defects or anomalies often have spatial continuity.

[0088] The specific implementation employs a spatial clustering algorithm. The input is a set of two-dimensional planar coordinates of all "high-value uncertainty points" selected in step S52. A density-based spatial clustering algorithm, such as DBSCAN (a density-based spatial clustering application with noise), is used. Two key parameters need to be set: the neighborhood search radius (eps) and the minimum number of points required to form a cluster. The parameter eps can be set according to the scan step resolution (e.g., 2-3 times the step size), and the minimum number of points can be set to 3-5. The DBSCAN algorithm groups these spatially adjacent points into the same cluster and marks isolated points far from any dense area as noise points. All clusters formed by at least the minimum number of points, covering a spatial area, are identified as the "final uncertainty region." These regions appear as connected blocks on the wafer map and may indicate areas with complex defects, material inhomogeneities, or signal interference.

[0089] S54: Apply quality judgment rules to the detection points within the final uncertainty region to generate a final quality judgment report containing regional judgment conclusions.

[0090] In this embodiment, this step is the final decision-making stage, applying differentiated judgment rules to different regions and generating a comprehensive report. The specific implementation is divided into regional judgment and report generation.

[0091] First, apply differentiated quality assessment rules. For all inspection points identified as "final uncertainty areas," a more stringent set of assessment rules is applied. For example, defect assessment in regular areas may only require a prediction confidence level greater than 0.7 to accept the classification result; while for points in these uncertainty areas, the prediction confidence level must be greater than a higher threshold (such as 0.9), and their defect category may need to be marked as "suspected" or forcibly classified as "requiring manual re-inspection." For inspection points outside the uncertainty areas, standard assessment rules are applied.

[0092] Then, generate the final quality assessment report. The report should include the following information: 1) Wafer-level defect statistics (number and area percentage of various types of defects); 2) Defect distribution map, where areas with high uncertainty should be highlighted with different colors or layers; 3) Description of the location, area, average fusion uncertainty measure, and dominant contribution difference (e.g., “ultrasound-dominated conflict” or “auxiliary signal-dominated conflict”) of each identified uncertainty region; 4) Based on preset acceptance criteria (such as not allowing any "confirmed" fatal defects and the total area of ​​"high uncertainty region" not exceeding a certain percentage of the wafer area), give the final judgment conclusion of "acceptable", "unacceptable" or "requires engineering review" for the wafer.

[0093] This report has transformed the test results from "site data" into "actionable quality control conclusions".

[0094] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0095] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0096] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

Claims

1. A method for patternless wafer defect detection of ultrasonic multi-frequency signals, characterized in that, The detection method includes the following steps: For each detection point on the wafer, its multimodal raw signal is acquired and preprocessed to construct a multi-channel feature vector for each detection point. The multimodal raw signal includes a multi-frequency ultrasonic signal and at least one auxiliary signal. The multi-channel feature vectors are subjected to probabilistic fusion processing to obtain the probabilistic fusion feature distribution corresponding to each detection point, wherein the probabilistic fusion feature distribution is parameterized by the mean vector and the variance vector; Multiple feature samples are sampled from the probabilistic fusion feature distribution of each detection point, and multiple random forward propagation inferences are performed on the multiple feature samples to calculate the defect category prediction result, prediction confidence and fusion uncertainty measure of each detection point. For detection points where the fusion uncertainty measure exceeds a preset warning value, the corresponding variance vector and mean vector are inverted to decouple and quantify the difference in contribution of the multi-frequency ultrasonic signal and the auxiliary signal to the fusion feature uncertainty of the detection point. Based on the comparison results of the fusion uncertainty metric and dynamic threshold at each detection point, and combined with the contribution difference, uncertainty regions are identified and a final quality assessment report for the wafer is generated, wherein the dynamic threshold is determined by the statistical distribution of the predicted confidence scores of all detection points.

2. A method for non-graphical wafer defect detection of ultrasonic multi-frequency signals according to claim 1, characterized in that, For each detection point on the wafer, its multimodal raw signal is acquired and preprocessed to construct a multi-channel feature vector for each detection point, including: For each detection point, multi-frequency ultrasonic signals and auxiliary signals are collected simultaneously, and the echoes of each independent frequency channel of the multi-frequency ultrasonic signals are preliminarily conditioned. For the echoes of each frequency channel after preliminary conditioning, feature transformation is performed to extract time-domain correlation features that characterize the difference between the signal of each channel and the reference state, forming an ultrasound feature group; The ultrasound feature set is fused and spliced ​​with the auxiliary features extracted from the auxiliary signal to generate a multi-channel feature vector for the detection point.

3. A method of non-graphical wafer defect detection of ultrasonic multi-frequency signals according to claim 2, characterized in that, The step of performing probabilistic fusion processing on the multi-channel feature vectors to obtain the probabilistic fusion feature distribution corresponding to each detection point includes: Ultrasonic feature components and auxiliary feature components are extracted from the multi-channel feature vectors respectively, and their respective modal built-in confidence is evaluated; Based on the modality built-in confidence, dynamic fusion weights are assigned to the ultrasound feature components and auxiliary feature components; The ultrasonic feature components and auxiliary feature components are weighted and fused using the dynamic fusion weights to form an intermediate fused feature; Based on the intermediate fusion features, a probabilistic fusion feature distribution parameterized by the mean vector and the variance vector is generated through a pre-set probability mapping network.

4. The method of claim 3, wherein the method is a non-graphical wafer defect detection method of ultrasonic multi-frequency signals. The process involves sampling multiple feature samples from the probabilistic fusion feature distribution of each detection point, performing multiple random forward propagation inferences on these feature samples, and calculating the defect category prediction result, prediction confidence, and fusion uncertainty measure for each detection point, including: Based on the variance vector of the probabilistic fusion feature distribution, importance sampling is performed on each detection point to obtain multiple feature samples. Each feature sample is then input into a pre-set Bayesian neural network classifier for multiple random forward propagations to obtain multiple intermediate prediction probability distributions for each feature sample. The multiple intermediate prediction probability distributions of each feature sample are integrated to obtain the preliminary prediction results and cognitive uncertainty estimates of each sample. The preliminary prediction results of all feature samples are then integrated again to generate the defect category prediction results and prediction confidence of the detection point. By combining the cognitive uncertainty estimate of each feature sample with the variance vector of the probabilistic fusion feature distribution, the fusion uncertainty measure of the detection point is calculated.

5. A method of non-graphical wafer defect detection of an ultrasonic multi-frequency signal according to claim 4, characterized in that, For detection points where the fusion uncertainty metric exceeds a preset warning value, the corresponding variance vector and mean vector are inverted to decouple and quantify the difference in contribution of the multi-frequency ultrasonic signal and the auxiliary signal to the fusion feature uncertainty of the detection point, including: Based on the variance vector, calculate the relative gradient contribution of the feature coding channels of the multi-frequency ultrasound signal and the feature coding channels of the auxiliary signal in each dimension of the fused feature distribution; Based on the relative contribution of the gradient, source weights of uncertainty are assigned to the ultrasound feature components and the auxiliary feature components; The variance vector is weighted and decomposed using the source weights to obtain the contribution of the uncertainty components corresponding to the multi-frequency ultrasound signal and the auxiliary signal, respectively, and the contribution difference is obtained.

6. The method for detecting patternless wafer defects using ultrasonic multi-frequency signals according to claim 5, characterized in that, The contribution difference is the ratio or difference between the contribution of the uncertainty component of the multi-frequency ultrasound signal and the contribution of the uncertainty component of the auxiliary signal.

7. A method of non-graphical wafer defect detection of an ultrasonic multi-frequency signal according to claim 6, characterized in that, The step of identifying uncertainty regions and generating a final quality assessment report for the wafer based on the comparison results of the fusion uncertainty metric and dynamic threshold at each detection point, combined with the contribution difference, includes: The fusion uncertainty measure of each detection point is compared with the dynamic threshold, and detection points with fusion uncertainty measures higher than the dynamic threshold are initially screened to form a candidate uncertainty point set; For each detection point in the candidate uncertainty point set, determine whether the corresponding contribution difference exceeds a preset difference threshold; From the set of candidate uncertainty points, detection points whose contribution differences exceed the difference threshold are selected, and clustering is performed based on the spatial distribution of the detection points. The region formed by points that meet the preset conditions in space is identified as the final uncertainty region. For the detection points within the final uncertainty region, quality judgment rules are applied to generate a final quality judgment report containing regional judgment conclusions.

8. A method of non-graphical wafer defect detection of an ultrasonic multi-frequency signal according to claim 7, characterized in that, The dynamic threshold is determined by the statistical distribution of the predicted confidence scores of all detection points, specifically: Calculate the mean and standard deviation of the predicted confidence scores for all detection points; The dynamic threshold is calculated based on a linear combination of the mean and the standard deviation, wherein the dynamic threshold is positively correlated with the mean and negatively correlated with the standard deviation.

Images