An artificial intelligence-based semiconductor data detection system and method

Abstract

The application discloses a kind of semiconductor data detection systems and methods based on artificial intelligence, it is related to semiconductor detection technical field, including, through quantum circuit execution parameterization rotating door operation, and construct renormalization group equation generation physical loss term, simultaneously through quantum discriminator and quantum generator construct generation adversarial network, generate enhanced data set;According to enhanced data set construction physical topology adjacency matrix, load source domain convolutional neural network model and carry out feature transformation, through dynamic physical graph attention mechanism and lattice differential homeomorphism mapping, generate physical constraint embedded analysis engine;Deployment physical constraint embedded analysis engine and handle real-time wafer data, generate lattice defect tensor and space-time confidence field, and obtain confidence score, when confidence score meets high confidence condition, process adjustment instruction is triggered, otherwise the data that does not meet condition is stored in feedback queue;The application realizes the deep coupling of quantum generation process and semiconductor lattice physical law.

Description

Technical Field

[0001] This invention relates to the field of semiconductor testing technology, and in particular to a semiconductor data testing system and method based on artificial intelligence. Background Technology

[0002] As semiconductor manufacturing processes advance to the nanometer scale and below, the complexity of device structures increases dramatically, placing higher demands on the precision of defect detection and process control during manufacturing. Traditional inspection techniques mainly rely on optical or electron microscopy combined with thresholding image processing methods, which struggle to address the challenges of modeling high-dimensional, low signal-to-noise ratio, and sparse defect data. In recent years, breakthroughs in artificial intelligence, particularly deep learning, in image recognition and anomaly detection have provided new pathways for online semiconductor inspection. Convolutional neural networks (CNNs) and graph neural networks (GNNs) are widely used for wafer defect classification and localization, but their performance is limited by the scale and diversity of the training data.

[0003] Current mainstream detection architectures mostly adopt a "perception-decision" separation model, lacking a closed-loop feedback path from detection output to process adjustment. This results in delayed anomaly response, making it difficult to meet the real-time requirements of advanced processes. In addition, model updates typically rely on periodic offline retraining, which cannot effectively utilize online substandard samples for incremental learning, easily leading to knowledge forgetting and resource waste. Summary of the Invention

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

[0005] Therefore, this invention provides an artificial intelligence-based semiconductor data detection method to solve the problems of delayed response and untimely model updates caused by the disconnect between detection and control in the prior art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a semiconductor data detection method based on artificial intelligence, comprising: performing a parameterized rotating gate operation through quantum circuits and constructing a renormalization group equation to generate a physical loss term; simultaneously constructing a generative adversarial network through a quantum discriminator and a quantum generator to generate an enhanced dataset; constructing a physical topological adjacency matrix based on the enhanced dataset, loading a source domain convolutional neural network model and performing feature transformation; generating a physically constrained embedded analysis engine through a dynamic physical graph attention mechanism and lattice differential homeomorphism mapping; deploying the physically constrained embedded analysis engine and processing real-time wafer data to generate a lattice defect tensor and a spatiotemporal confidence field, and obtaining a confidence score; triggering a process adjustment command when the confidence score meets a high confidence condition, otherwise storing the data that does not meet the condition in a feedback queue; receiving the process adjustment command, sending a manufacturing equipment control signal to the manufacturing equipment based on the lattice defect tensor to execute dynamic process adjustment, and using the feedback queue to drive incremental learning to retrain and generate the physically constrained embedded analysis engine.

[0007] As a preferred embodiment of the artificial intelligence-based semiconductor data detection method of the present invention, the specific steps of performing a parameterized rotating gate operation through quantum circuits and constructing a renormalization group equation to generate a physical loss term are as follows: Quantum states are generated by performing parameterized rotating gate operations using quantum circuits, and then converted into synthetic defect images using a classical decoder. The renormalization group equation is constructed by predefined action, and the scaling law compliance of the dislocation density observable operator of the synthesized defect image is calculated to generate the physical loss term.

[0008] As a preferred embodiment of the artificial intelligence-based semiconductor data detection method of the present invention, the step of simultaneously constructing a generative adversarial network through a quantum discriminator and a quantum generator to generate an enhanced dataset includes the following specific steps. Generative adversarial networks are constructed by using quantum discriminators and quantum generators, and hybrid quantum operations are performed to generate the trained generative adversarial network. The trained generative adversarial network is optimized based on the entropy maximization constraint, and then combined with the physics loss term to generate an augmented dataset.

[0009] As a preferred embodiment of the artificial intelligence-based semiconductor data detection method of the present invention, the specific steps of constructing a physical topological adjacency matrix based on the enhanced dataset, loading the source domain convolutional neural network model, and performing feature transformation are as follows. Based on the defect image matrix and dislocation density tensor in the enhanced dataset, a physical topological adjacency matrix is ​​constructed using the lattice continuity equation. Load the source domain convolutional neural network model, extract the feature map of the defect image matrix, and segment it into grain node feature vectors; The unitary matrix is ​​obtained by decomposing the dislocation density tensor using singular value decomposition. A lattice covariant convolution kernel is then constructed to perform physical constraint-based feature transformation on the eigenvectors of grain nodes.

[0010] As a preferred embodiment of the artificial intelligence-based semiconductor data detection method of the present invention, the step of generating a physically constrained embedded analysis engine through a dynamic physical graph attention mechanism and lattice differential homeomorphism mapping includes the following specific steps. Based on the physical topology adjacency matrix and the grain node feature vector after feature transformation, a dynamic physical graph attention mechanism is adopted to generate attention weight coefficients between grains. Based on the attention weight coefficients, lattice differential homeomorphism is performed to generate a physically constrained embedded analysis engine.

[0011] As a preferred embodiment of the artificial intelligence-based semiconductor data inspection method of the present invention, the specific steps for deploying a physically constrained embedded analysis engine and processing real-time wafer data to generate lattice defect tensors and spatiotemporal confidence fields are as follows. Deploy a physically constrained embedded analysis engine on edge computing nodes to collect real-time wafer data; Real-time wafer data is processed through a physical constraint embedded analysis engine to generate lattice defect tensors and spatiotemporal confidence fields.

[0012] As a preferred embodiment of the artificial intelligence-based semiconductor data inspection method of the present invention, the following steps are taken: When the confidence score meets the high confidence condition, a process adjustment instruction is triggered; otherwise, the data that does not meet the condition is stored in a feedback queue. The confidence score is calculated based on the spatiotemporal confidence field. When the confidence score exceeds the process stability boundary threshold, a process adjustment command is generated. When the confidence score does not exceed the process stability boundary threshold, the real-time wafer data and lattice defect tensor are encoded into quantum states and stored in the feedback queue.

[0013] In a preferred embodiment of the artificial intelligence-based semiconductor data inspection method of the present invention, the steps of receiving process adjustment instructions and sending manufacturing equipment control signals to the manufacturing equipment according to the lattice defect tensor to perform dynamic process adjustment are as follows: Receive process adjustment instructions, construct instanton gauge fields based on lattice defect tensors, and generate instanton numbers through Chen-Simons integration; Instantaneous numbers are mapped to process parameter vectors to generate control signals for manufacturing equipment, which are then broadcast to the manufacturing equipment via a quantum-classical hybrid channel for dynamic process adjustment.

[0014] As a preferred embodiment of the artificial intelligence-based semiconductor data detection method of the present invention, the specific steps of using a feedback queue to drive incremental learning and retraining to generate a physically constrained embedded analysis engine are as follows: Quantum state data is extracted from the feedback queue and generated into a classical training dataset through geometric flow decoding. Using a classic training dataset, the distribution matching training of the generative adversarial network with renormalized group constraints is performed to generate physically constrained augmented data. Based on the physical constraint augmentation data, the model parameters of the physical constraint embedded analysis engine are retrained using the Lie group symmetric optimization algorithm to generate an updated physical constraint embedded analysis engine.

[0015] Secondly, this invention provides an artificial intelligence-based semiconductor data inspection system, comprising a quantum enhancement module, a feature mapping module, a real-time analysis module, and a closed-loop control module. The quantum enhancement module performs parameterized rotating gate operations via quantum circuitry and constructs a renormalization group equation to generate physical loss terms. Simultaneously, it constructs a generative adversarial network (GAN) with a quantum discriminator and a quantum generator to generate an enhanced dataset. The feature mapping module constructs a physical topological adjacency matrix based on the enhanced dataset, loads a source domain convolutional neural network model, performs feature transformation, and generates a physically constrained embedded analysis engine through a dynamic physical graph attention mechanism and lattice differential homeomorphism mapping. The real-time analysis module deploys the physically constrained embedded analysis engine and processes real-time wafer data, generating a lattice defect tensor and a spatiotemporal confidence field, and obtaining a confidence score. When the confidence score meets a high confidence condition, a process adjustment command is triggered; otherwise, data that does not meet the condition is stored in a feedback queue. The closed-loop control module receives the process adjustment command, sends manufacturing equipment control signals to the manufacturing equipment based on the lattice defect tensor to execute dynamic process adjustment, and uses the feedback queue to drive incremental learning to retrain and generate the physically constrained embedded analysis engine.

[0016] The beneficial effects of this invention are as follows: By performing parameterized rotating gate operations through quantum circuits and combining them with renormalization group equations to generate physical loss terms, a deep coupling between the quantum generation process and the physical laws of semiconductor lattice is achieved, thereby overcoming the defect of traditional GAN-generated samples deviating from physical constitutive relations; by constructing a feedback queue to drive incremental learning, a closed-loop mapping of detection error information to process control parameters is achieved, avoiding the knowledge forgetting problem caused by offline retraining. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an artificial intelligence-based semiconductor data inspection method.

[0019] Figure 2 A flowchart for generating a quantum-enhanced dataset.

[0020] Figure 3 A flowchart for building a physical constraint embedded analysis engine.

[0021] Figure 4 This is a flowchart for real-time defect detection and closed-loop control. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a semiconductor data detection method based on artificial intelligence, including the following steps: S1: Perform parameterized rotating gate operations through quantum circuits and construct renormalization group equations to generate physical loss terms. At the same time, construct a generative adversarial network through a quantum discriminator and a quantum generator to generate an enhanced dataset.

[0026] Quantum states are generated by performing parameterized rotating gate operations using quantum circuits, and then converted into synthetic defect images using a classical decoder.

[0027] The specific process includes: First, the quantum bit state is prepared on the quantum computing device by performing a parameterized rotating gate operation. Then, a quantum rotating gate operation controlled by adjustable parameters is applied, causing the quantum bit state to evolve into the target quantum state according to a preset method. A classical decoder measures the density matrix of the quantum state using quantum state tomography, converts the measurement result into a complex probability amplitude distribution, and then generates a two-dimensional pixel matrix through inverse Fourier transform and normalization. Finally, a synthetic defect image reflecting the defect characteristics of the semiconductor material is output.

[0028] The preset method is determined by the rotation angle sequence of the parameterized rotating gate in the quantum circuit, which is optimized through the training process of the quantum generative adversarial network.

[0029] The renormalization group equation is constructed by predefined action, and the scaling law compliance of the dislocation density observable operator of the synthesized defect image is calculated to generate the physical loss term, expressed as: ; in, Represents the physical loss term. Dislocation density observable operator Expected value Indicates the momentum cutoff scale. Indicates anomaly dimension, This represents the weighting coefficient of the variance term. Dislocation density observable operator The variance.

[0030] The specific process includes: constructing renormalization group equations by predefined action quantities; deriving renormalization group equations describing the evolution of physical quantities at different energy scales using variational principles; mapping the synthetic defect image generated by the quantum circuit to the eigenstate space of the dislocation density observable operator; and extracting key feature parameters from the dislocation density distribution function. The renormalization group equations are used to analyze the variation behavior of these key feature parameters at different scales, and the deviation of the logarithmic derivatives of these key feature parameters from the theoretically expected scaling exponent is calculated. In the calculation of the expected value of the dislocation density observable operator, the influence of the variance term on physical consistency is considered simultaneously, establishing a composite evaluation index that includes first-order derivative constraints and second-order wave constraints. The calculation results are quantified into a physical loss term reflecting the degree of matching between the generated image and the physical properties of real lattice defects. This physical loss term is used in subsequent training to guide the optimization direction of the parameterized rotating gate operation, ensuring that the generated quantum state, after decoding, conforms to both data statistical characteristics and solid-state physics laws.

[0031] The expected scaling exponent is predetermined based on theoretical and experimental measurements of dislocation density in semiconductor materials.

[0032] Generative adversarial networks (GANs) are constructed by using quantum discriminators and quantum generators, and hybrid quantum operations are performed to generate the trained GAN.

[0033] The specific process involves a quantum discriminator and a quantum generator forming a generative adversarial network (GAN). The quantum discriminator projects the quantum state output by the quantum generator through quantum measurement operations, converting the measurement result into a classical discrimination score. The quantum generator employs a parameterized quantum circuit structure, generating the target quantum state by adjusting the rotation gate angle parameter. During training, the quantum discriminator continuously optimizes the measurement basis vectors to improve its ability to distinguish generated quantum states, while the quantum generator backpropagates gradient information to adjust the parameterized rotation gate operation, making the generated quantum state closer to the real data distribution. Hybrid quantum operations are manifested in the quantum discriminator performing quantum measurement and the quantum generator performing quantum state preparation; both exchange gradient information through classical communication to achieve collaborative optimization. After multiple iterations of training, the quantum discriminator and quantum generator reach a Nash equilibrium state. At this point, the quantum generator can produce high-quality quantum states, and the quantum discriminator cannot effectively distinguish between generated and real samples, ultimately resulting in a trained GAN. The entire training process strictly maintains the unitarity constraint of the quantum circuit, ensuring that quantum operations conform to the fundamental principles of quantum mechanics.

[0034] Nash equilibrium refers to the state in a generative adversarial network where the quantum discriminator cannot distinguish between generated samples and real samples, and the quantum generator cannot further optimize the quality of generated samples.

[0035] The trained generative adversarial network is optimized based on the entropy maximization constraint, and then combined with the physics loss term to generate an augmented dataset.

[0036] The specific process includes constructing a generative adversarial network (GAN) using a quantum discriminator and a quantum generator. The trained GAN is then optimized under entropy maximization constraints to make the distribution of the synthesized defect images more uniform and cover more potential defect patterns. At the same time, the physical loss term generated by the renormalization group equation ensures that the dislocation density observable operator of the synthesized defect images meets the scaling law requirements. The physical loss term constrains the consistency between the output of the GAN and the physical laws of semiconductor lattice by identifying the difference between the expected value of the dislocation density and the logarithmic derivative of the momentum cutoff scale, as well as the weighted result of the variance term. The final enhanced dataset contains diverse defect morphologies and strictly follows the basic principles of materials science, providing high-quality input for the subsequent construction of the physical topological adjacency matrix.

[0037] S2: Construct a physical topological adjacency matrix based on the augmented dataset, load the source domain convolutional neural network model and perform feature transformation, and generate a physical constraint embedded analysis engine through dynamic physical graph attention mechanism and lattice differential homeomorphism mapping.

[0038] Based on the defect image matrix and dislocation density tensor in the augmented dataset, a physical topological adjacency matrix is ​​constructed using the lattice continuity equation.

[0039] The specific process includes: enhancing the defect image matrix in the dataset to record the spatial distribution characteristics of defects on the wafer surface; the dislocation density tensor quantifies the degree and direction of lattice distortion; the lattice continuity equation derives the dislocation flux conservation relationship between adjacent grains through the defect image matrix and the dislocation density tensor; the connection strength between grain nodes is obtained by solving the partial differential equations in the lattice continuity equation; the connection strength is normalized and mapped to matrix element values; and the finally constructed physical topological adjacency matrix accurately characterizes the propagation path and interaction strength of lattice defects in space, providing a graph structure input that conforms to the physical nature of materials for the dynamic physical graph attention mechanism.

[0040] Load the source domain convolutional neural network model, extract the feature map of the defect image matrix, and segment it into grain node feature vectors.

[0041] It should be noted that the pre-training process of the source domain convolutional neural network model uses a large-scale semiconductor defect image dataset as input. A deep feature extraction architecture is constructed by alternately stacking multiple convolutional layers and pooling layers. During the training phase, the cross-entropy loss function is used to supervise the model's learning of the defect classification task. The stochastic gradient descent algorithm is used to optimize the convolutional kernel parameters. Batch normalization technology is applied during training to accelerate convergence and prevent overfitting. Finally, the pre-trained source domain convolutional neural network model can effectively extract general feature representations of defect images. These feature representations can be adapted to new semiconductor detection tasks through transfer learning, providing basic feature extraction capabilities for subsequent feature transformation and the construction of a physically constrained embedded analysis engine.

[0042] The specific process includes: a pre-trained source domain convolutional neural network model receives a defect image matrix as input, and extracts local texture features and global structural features of the defect step by step through multi-layer convolution operations to form a feature map with spatial dimensions. The feature map is divided into multiple sub-feature blocks corresponding to different grain regions according to the grain boundary division rules. Each sub-feature block is compressed into a fixed-dimensional numerical vector through spatial pooling operation. These numerical vectors are the grain node feature vectors that characterize the defect characteristics of a single grain region, providing the basic feature representation for subsequent operations to obtain the unitary matrix through singular value decomposition of the dislocation density tensor and construct the lattice covariant convolution kernel.

[0043] Grain boundary partitioning rules refer to the operational guidelines for accurately dividing the feature map into independent grain region sub-blocks based on the crystal structure characteristics of the wafer material and the spatial distribution mapping relationship of the grains determined by electron backscatter diffraction technology or X-ray diffraction patterns, according to the grain boundary coordinates.

[0044] The unitary matrix is ​​obtained by decomposing the dislocation density tensor using singular value decomposition. A lattice covariant convolution kernel is then constructed to perform physical constraint-based feature transformation on the eigenvectors of grain nodes.

[0045] The specific process involves decomposing the dislocation density tensor into a product of three matrices through singular value decomposition. The unitary matrix contains the principal direction information of lattice defects. The orthogonality of the unitary matrix is ​​used to construct a lattice covariant convolution kernel. The weight distribution of the lattice covariant convolution kernel is consistent with the physical propagation characteristics of lattice defects. The eigenvectors of the grain nodes are convolved with the lattice covariant convolution kernel. During the convolution process, the transformation direction of the eigenvectors is constrained by the unitary matrix, so that the transformed eigenvectors retain the original defect characteristics while embedding the physical laws of lattice defects. The final output eigenvectors contain both the morphological characteristics of the grain region and the physical constraints of dislocation density, providing a feature representation that conforms to the principles of materials science for the dynamic physical graph attention mechanism.

[0046] Based on the physical topology adjacency matrix and the grain node feature vectors after feature transformation, a dynamic physical graph attention mechanism is used to generate attention weight coefficients between grains.

[0047] The specific process includes: a physical topological adjacency matrix describes the connection relationships and interaction strengths between grain nodes; grain node eigenvectors after feature transformation carry defect feature information under physical constraints; a dynamic physical graph attention mechanism identifies the cosine similarity between grain node eigenvectors and the Hadamard product of the physical topological adjacency matrix, and combines this with the scaling factor derived from the renormalization group equation to obtain a similarity score for each pair of grain node eigenvectors. The similarity score and the connection strength at corresponding positions in the physical topological adjacency matrix jointly determine the initial value of the attention weight coefficients. Then, a learnable nonlinear transformation is used to adjust the distribution of the attention weight coefficients, so that the connections between important grain nodes receive higher weights. The final generated attention weight coefficients accurately reflect the physical correlation strength of defect propagation between grains, providing quantified interaction parameters for subsequent lattice differential homeomorphism mapping.

[0048] Based on the attention weight coefficients, lattice differential homeomorphism is performed to generate a physically constrained embedded analysis engine.

[0049] The specific process includes: attention weighting coefficients quantify the interaction strength between grain nodes; lattice differential homeomorphism mapping uses attention weighting coefficients to construct a differentiable lattice deformation field; the lattice deformation field adjusts the feature space geometry while maintaining lattice continuity, so that the grain nodes corresponding to high-weight connections are closer in the feature space; the mapping process ensures that the deformation meets the material mechanical constraints by solving partial differential equations; and the finally generated physical constraint embedded analysis engine encodes the grain node feature vectors, physical topological relationships and lattice deformation laws into a computable mathematical representation. This mathematical representation retains the original defect characteristics and embeds lattice physical properties, providing a physically meaningful analysis framework for subsequent real-time wafer data processing.

[0050] S3: Deploy a physical constraint embedded analysis engine and process real-time wafer data to generate lattice defect tensors and spatiotemporal confidence fields, and obtain confidence scores. When the confidence score meets the high confidence condition, a process adjustment command is triggered; otherwise, data that does not meet the condition is stored in the feedback queue.

[0051] Deploy a physically constrained embedded analysis engine on edge computing nodes to collect real-time wafer data.

[0052] The specific process involves edge computing nodes continuously acquiring real-time wafer data through integrated high-precision optical sensors and electron probe arrays, including real-time stress distribution maps, thermal field imaging, and carrier concentration distribution. This real-time wafer data is transmitted to the physical constraint embedded analysis engine for processing with millisecond-level latency. The edge deployment architecture ensures low latency characteristics for data acquisition and processing, while avoiding the security risks associated with transmitting sensitive process data to remote servers, providing localized computing support for subsequent real-time defect detection and control command generation.

[0053] Real-time wafer data is processed through a physical constraint embedded analysis engine to generate lattice defect tensors and spatiotemporal confidence fields.

[0054] The specific process includes: after receiving real-time wafer data, the physical constraint embedded analysis engine first converts physical quantities such as stress distribution and temperature gradient into feature vector representations. These feature vectors are then matched with the feature vectors of the grain nodes based on similarity. The dynamic physical graph attention mechanism updates the attention weight coefficients between grain nodes according to the matching results. The updated weight coefficients guide the lattice differential homeomorphism mapping to readjust the feature space structure. The adjusted feature space generates a lattice defect tensor containing defect type, location, and severity through tensor operations. At the same time, the confidence distribution in the feature space is obtained to form a spatiotemporal confidence field. The spatiotemporal confidence field quantifies the reliability of the defect detection results in the time and space dimensions. The lattice defect tensor and the spatiotemporal confidence field together constitute a complete defect analysis result, providing a basis for subsequent process control decisions.

[0055] A confidence score is calculated based on the spatiotemporal confidence field. When the confidence score exceeds the process stability boundary threshold, a process adjustment command is generated, expressed as: ; in, Represents the spatial coordinate vector of the wafer surface. Represents a time variable. Indicates the confidence score. This represents the natural exponential function. Indicates the time window for process characteristics. Indicates time The characteristic value of the defect change rate, Represents the defective tensor field. Represents the time variable of integration. Represents the stress-defect coupling term. Denotes the Frobenius norm. Represents the spatial gradient of the stress field. This represents the characteristic value of the stress gradient.

[0056] The specific process includes: a spatiotemporal confidence field records the reliability distribution of defect detection results in spatial location and time series; the spatiotemporal confidence field is converted into a scalar confidence score through integration; the confidence score is compared with a process stability boundary threshold; when the confidence score exceeds the process stability boundary threshold, it indicates that the current detection result meets the credibility requirements, triggering a mechanism to immediately generate a process adjustment command containing process adjustment parameters; the process adjustment command is directly sent to the relevant process equipment through the manufacturing equipment interface for real-time adjustment; the entire process is completed within a millisecond timescale, ensuring that the semiconductor manufacturing process can respond and optimize rapidly based on high-confidence defect detection results.

[0057] The process stability boundary threshold is a dynamic critical value that is predetermined by statistical process control methods based on historical yield data of semiconductor manufacturing processes, material characteristic parameters, and equipment operating specifications.

[0058] When the confidence score does not exceed the process stability boundary threshold, the real-time wafer data and lattice defect tensor are encoded into quantum states and stored in the feedback queue.

[0059] The specific process includes the following steps: when the confidence score is lower than the process stability boundary threshold, the stress distribution and temperature gradient information contained in the real-time wafer data and the defect features recorded by the lattice defect tensor are converted into a superposition state qubit sequence by a quantum encoder. The qubit sequence is written into the storage array of the feedback queue in timestamp order. The feedback queue uses a quantum random access memory structure to store the encoded quantum state data, ensuring that the physical characteristics and defect attributes of the low-confidence samples can be completely restored in the subsequent incremental learning stage, providing quantized input data for the retraining of the generative adversarial network.

[0060] S4: Receives process adjustment instructions, sends manufacturing equipment control signals to the manufacturing equipment based on the lattice defect tensor to perform dynamic process adjustment, uses feedback queues to drive incremental learning, and retrains to generate a physical constraint embedded analysis engine.

[0061] It receives process adjustment instructions, constructs instanton gauge fields based on lattice defect tensors, and generates instanton numbers through Chen-Simons integration.

[0062] The specific process includes the following steps: after the process adjustment command is triggered, the dislocation density and lattice distortion information recorded in the lattice defect tensor are converted into gauge potential parameters. The gauge potential parameters are used to construct a transient gauge field that describes the topological characteristics of the defect. The transient gauge field forms a nontrivial field configuration in the three-dimensional lattice space. The transient number, which characterizes the topological number of the defect, is obtained by identifying the curvature integral of the Chern-Simons integral on the closed surface. As a topological invariant that quantifies the severity of the defect, the transient number directly reflects the degree of influence of the lattice defect on the electrical performance of the semiconductor device, providing an accurate topological feature description for subsequent process parameter vector mapping.

[0063] Instantaneous numbers are mapped to process parameter vectors to generate control signals for manufacturing equipment, which are then broadcast to the manufacturing equipment via a quantum-classical hybrid channel for dynamic process adjustment.

[0064] The specific process includes converting instantaneous numbers into a multidimensional process parameter vector through nonlinear transformation matrix operations. The process parameter vector contains key process variables such as photolithography exposure dose, etching time, and deposition temperature. The converted process parameter vector is encoded into manufacturing equipment control signals. The manufacturing equipment control signals are transmitted through a quantum-classical hybrid channel. The quantum channel part uses quantum key distribution to ensure instruction security, while the classical channel part achieves high-speed broadcast transmission. After parsing the manufacturing equipment control signals, the receiving end of the manufacturing equipment immediately adjusts the working state of the actuator to complete the real-time dynamic adjustment of the semiconductor manufacturing process. The entire process achieves closed-loop control from defect detection to process optimization within the wafer processing cycle.

[0065] Quantum state data is extracted from the feedback queue and generated into a classical training dataset through geometric flow decoding.

[0066] The specific process includes: the quantum state data stored in the feedback queue is converted into a classical probability distribution through quantum measurement operations; the geometric flow decoder processes these classical probability distributions and reconstructs them into a continuous feature manifold; the sampling points on the feature manifold are orthogonally transformed to form normalized feature vectors; the normalized feature vectors are mapped to the spatiotemporal coordinates of the original real-time wafer data; the reconstructed feature vectors are labeled according to defect category and confidence score; and the final generated classical training dataset contains complete defect feature representations and physical constraint information, providing incremental learning samples that conform to the laws of materials science for the retraining of the generative adversarial network.

[0067] Using a classic training dataset, we perform distribution matching training on the generative adversarial network to renormalize group constraints, thereby generating data that enhances physical constraints.

[0068] The specific process includes: after inputting the classical training dataset into the generative adversarial network (GAN), the quantum generator network generates synthetic defect samples; the discriminator network evaluates the distribution differences between the synthetic samples and the classical training dataset; the renormalization group constraint identifies the invariance of the dislocation density observable operator of the generated samples under momentum space scaling transformation, constructs a physical regularization term and adds it to the loss function of the GAN; during training, the quantum generator network simultaneously optimizes the adversarial loss and the renormalization group constraint loss, so that the generated synthetic defect samples not only match the statistical characteristics of the classical training dataset but also satisfy the physical laws of semiconductor materials. The final output of the physical constraint-enhanced data contains diverse defect patterns that conform to the principles of lattice dynamics, providing training samples with both data diversity and physical rationality for updating the physical constraint embedded analysis engine.

[0069] Based on the physical constraint augmentation data, the model parameters of the physical constraint embedded analysis engine are retrained using the Lie group symmetric optimization algorithm to generate an updated physical constraint embedded analysis engine.

[0070] The specific process includes: after the physical constraint enhancement data is input into the physical constraint embedded analysis engine, the Lie group symmetry optimization algorithm analyzes the invariant properties of lattice defects under rotation and translation transformations, constructs parameter update rules that maintain crystal symmetry, and adjusts the convolution kernel weights by identifying the gradient direction of the loss function under group action, so that the feature extraction process maintains the transformation law consistent with the lattice symmetry. The parameter space of the trained physical constraint embedded analysis engine forms a symmetry-aware representation of the defect patterns in the physical constraint enhancement data. Finally, the updated physical constraint embedded analysis engine has the ability to identify new defects while maintaining strict adherence to the physical laws of the lattice, achieving a simultaneous improvement in detection accuracy and physical consistency.

[0071] This embodiment also provides an artificial intelligence-based semiconductor data inspection system, including: a quantum enhancement module, a feature mapping module, a real-time analysis module, and a closed-loop control module. The quantum enhancement module is used to perform parameterized rotating gate operations through quantum circuits, construct renormalization group equations to generate physical loss terms, and simultaneously construct a generative adversarial network through a quantum discriminator and a quantum generator to generate an enhanced dataset. The feature mapping module is used to construct a physical topological adjacency matrix based on the enhanced dataset, load a source domain convolutional neural network model and perform feature transformation, and generate a physical constraint embedded analysis engine through a dynamic physical graph attention mechanism and lattice differential homeomorphism mapping. The real-time analysis module is used to deploy the physical constraint embedded analysis engine and process real-time wafer data, generate lattice defect tensors and spatiotemporal confidence fields, and obtain confidence scores. When the confidence score meets the high confidence condition, a process adjustment command is triggered; otherwise, data that does not meet the condition is stored in a feedback queue. The closed-loop control module is used to receive process adjustment commands, send manufacturing equipment control signals to the manufacturing equipment based on the lattice defect tensor to execute dynamic process adjustment, and use the feedback queue to drive incremental learning to retrain and generate the physical constraint embedded analysis engine.

[0072] This embodiment also provides a computer device applicable to the case of an artificial intelligence-based semiconductor data detection method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the artificial intelligence-based semiconductor data detection method proposed in the above embodiment.

[0073] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0074] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the artificial intelligence-based semiconductor data detection method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0075] In summary, this invention achieves deep coupling between the quantum generation process and the physical laws of semiconductor lattice by performing parameterized rotating gate operations using quantum circuits and combining them with renormalization group equations to generate physical loss terms, thereby overcoming the defect of traditional GAN-generated samples deviating from physical constitutive relations; by constructing a feedback queue to drive incremental learning, it achieves closed-loop mapping of detection error information to process control parameters, avoiding the knowledge forgetting problem caused by offline retraining.

[0076] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A semiconductor data detection method based on artificial intelligence, characterized in that: include, The parameterized rotation gate operation is performed using quantum circuits, and the physical loss term is generated by constructing the renormalization group equation. The specific steps are as follows: Quantum states are generated by performing parameterized rotating gate operations using quantum circuits, and then converted into synthetic defect images using a classical decoder. The renormalization group equation is constructed by predefined action, and the scaling law compliance of the dislocation density observable operator of the synthesized defect image is calculated to generate the physical loss term. Simultaneously, a generative adversarial network is constructed using a quantum discriminator and a quantum generator to generate an enhanced dataset; The physical topological adjacency matrix is ​​constructed based on the augmented dataset. The source domain convolutional neural network model is loaded and feature transformation is performed. The specific steps are as follows: Based on the defect image matrix and dislocation density tensor in the enhanced dataset, a physical topological adjacency matrix is ​​constructed using the lattice continuity equation. Load the source domain convolutional neural network model, extract the feature map of the defect image matrix, and segment it into grain node feature vectors; The unitary matrix is ​​obtained by singular value decomposition of dislocation density tensor, and a lattice covariant convolution kernel is constructed to perform physical constraint feature transformation on the eigenvectors of grain nodes. A physical constraint embedded analysis engine is generated by using a dynamic physical graph attention mechanism and lattice differential homeomorphism mapping. Deploy a physical constraint embedded analysis engine to process real-time wafer data, generate lattice defect tensors and spatiotemporal confidence fields, and obtain confidence scores. When the confidence score meets the high confidence condition, a process adjustment command is triggered; otherwise, data that does not meet the condition is stored in the feedback queue. Receive process adjustment instructions and send manufacturing equipment control signals to the manufacturing equipment based on the lattice defect tensor to perform dynamic process adjustment; The physical constraint embedded analysis engine is retrained and generated by using a feedback queue to drive incremental learning. The specific steps are as follows: Quantum state data is extracted from the feedback queue and generated into a classical training dataset through geometric flow decoding. Using a classic training dataset, the distribution matching training of the generative adversarial network with renormalized group constraints is performed to generate physically constrained augmented data. Based on the physical constraint augmentation data, the model parameters of the physical constraint embedded analysis engine are retrained using the Lie group symmetric optimization algorithm to generate an updated physical constraint embedded analysis engine.

2. The semiconductor data detection method based on artificial intelligence as described in claim 1, characterized in that: The process involves simultaneously constructing a generative adversarial network using a quantum discriminator and a quantum generator to generate an enhanced dataset. The specific steps are as follows: Generative adversarial networks are constructed by using quantum discriminators and quantum generators, and hybrid quantum operations are performed to generate the trained generative adversarial network. The trained generative adversarial network is optimized based on the entropy maximization constraint, and then combined with the physics loss term to generate an augmented dataset.

3. The semiconductor data detection method based on artificial intelligence as described in claim 2, characterized in that: The process of generating a physically constrained embedded analysis engine through a dynamic physical graph attention mechanism and lattice differential homeomorphism mapping involves the following specific steps. Based on the physical topology adjacency matrix and the grain node feature vector after feature transformation, a dynamic physical graph attention mechanism is adopted to generate attention weight coefficients between grains. Based on the attention weight coefficients, lattice differential homeomorphism is performed to generate a physically constrained embedded analysis engine.

4. The semiconductor data detection method based on artificial intelligence as described in claim 3, characterized in that: The deployment of a physically constrained embedded analysis engine to process real-time wafer data and generate lattice defect tensors and spatiotemporal confidence fields follows these steps: Deploy a physically constrained embedded analysis engine on edge computing nodes to collect real-time wafer data; Real-time wafer data is processed through a physical constraint embedded analysis engine to generate lattice defect tensors and spatiotemporal confidence fields.

5. The semiconductor data detection method based on artificial intelligence as described in claim 4, characterized in that: The process involves obtaining a confidence score. When the confidence score meets the high confidence condition, a process adjustment command is triggered; otherwise, data that does not meet the condition is stored in a feedback queue. The specific steps are as follows. The confidence score is calculated based on the spatiotemporal confidence field. When the confidence score exceeds the process stability boundary threshold, a process adjustment command is generated. When the confidence score does not exceed the process stability boundary threshold, the real-time wafer data and lattice defect tensor are encoded into quantum states and stored in the feedback queue.

6. The semiconductor data detection method based on artificial intelligence as described in claim 5, characterized in that: The process adjustment command is received, and dynamic process adjustment is executed by sending manufacturing equipment control signals to the manufacturing equipment based on the lattice defect tensor. The specific steps are as follows. Receive process adjustment instructions, construct instanton gauge fields based on lattice defect tensors, and generate instanton numbers through Chern-Simons integration; Instantaneous numbers are mapped to process parameter vectors to generate control signals for manufacturing equipment, which are then broadcast to the manufacturing equipment via a quantum-classical hybrid channel for dynamic process adjustment.

7. An artificial intelligence-based semiconductor data inspection system, based on the artificial intelligence-based semiconductor data inspection method according to any one of claims 1 to 6, characterized in that: It includes a quantum enhancement module, a feature mapping module, a real-time analysis module, and a closed-loop control module. The quantum enhancement module is used to perform parameterized rotating gate operations through quantum circuits and construct renormalization group equations to generate physical loss terms. At the same time, it constructs a generative adversarial network through a quantum discriminator and a quantum generator to generate enhanced datasets. The feature mapping module is used to construct a physical topological adjacency matrix based on the augmented dataset, load the source domain convolutional neural network model and perform feature transformation, and generate a physical constraint embedded analysis engine through dynamic physical graph attention mechanism and lattice differential homeomorphism mapping. The real-time analysis module is used to deploy the physical constraint embedded analysis engine and process real-time wafer data, generate lattice defect tensors and spatiotemporal confidence fields, and obtain confidence scores. When the confidence score meets the high confidence condition, the process adjustment command is triggered; otherwise, the data that does not meet the condition is stored in the feedback queue. The closed-loop control module receives process adjustment instructions, sends manufacturing equipment control signals to the manufacturing equipment based on the lattice defect tensor to perform dynamic process adjustment, and uses a feedback queue to drive incremental learning to retrain and generate a physical constraint embedded analysis engine.

Images