WAT data generation method and device, computer equipment and storage medium

Abstract

The invention relates to a WAT data generation method and device, computer equipment and a storage medium. The method comprises the following steps: in response to a WAT data generation request, pre-processing a target process condition to obtain a condition embedding vector; based on standard normal distribution sampling, determining an initial noise vector; inputting the initial noise vector and the condition embedding vector into a preset diffusion model to obtain target WAT data corresponding to the target process condition; the preset diffusion model is a pre-trained diffusion model which is not guided by a classifier. By adopting the method, the diversity and stability of generated results can be improved.

Description

Technical Field

[0001] This application relates to the field of wafer technology, and in particular to a WAT ​​data generation method, apparatus, computer device, and storage medium. Background Technology

[0002] With the rapid development of integrated circuit manufacturing processes, wafer acceptance test (WAT) data plays a crucial role in process modeling, defect analysis, and yield optimization. However, acquiring real WAT data is costly and time-consuming, especially during new process development or small-batch pilot production, where data scarcity limits the training of downstream models. Traditional techniques, such as generative adversarial networks (GANs), interpolation, and physical simulation, can generate virtual data. However, these methods still suffer from insufficient consistency in the distribution of generated data, and the training and generation processes also exhibit low stability due to network complexity.

[0003] Therefore, it is evident that existing technologies for generating WAT data still suffer from issues of low diversity and stability in the generated results. Summary of the Invention

[0004] Therefore, it is necessary to provide a WAT ​​data generation method, apparatus, computer equipment, and storage medium that can improve the diversity and stability of the generated results, addressing the aforementioned technical problems.

[0005] Firstly, this application provides a WAT ​​data generation method, the WAT data generation method comprising:

[0006] In response to the WAT data generation request, the target process conditions are preprocessed to obtain the condition embedding vector;

[0007] The initial noise vector is determined based on sampling from a standard normal distribution.

[0008] The initial noise vector and the conditional embedding vector are input into a preset diffusion model to obtain target WAT data corresponding to the target process conditions; the preset diffusion model is a pre-trained diffusion model without classifier guidance.

[0009] In one embodiment, the preprocessing of the target process conditions to obtain the condition embedding vector includes:

[0010] Obtain classification features that match the preset diffusion model;

[0011] Based on the classification features, the target process conditions are one-hot encoded to obtain a conditional embedding vector.

[0012] In one embodiment, the preset diffusion model includes a downsampling network, an upsampling network, and a conditional embedding network, wherein:

[0013] The downsampling network is used to extract features from the input data to obtain multi-scale features;

[0014] The conditional embedding network is used to determine the mapping control parameters based on the conditional embedding vector;

[0015] The upsampling network is used to generate target WAT data based on the multi-scale features and the mapping control parameters.

[0016] In one embodiment, the upsampling network includes a feature amplification module and a feature fusion module, wherein:

[0017] The feature amplification module is used to amplify the multi-scale features to obtain stable distribution features;

[0018] The feature fusion module is used to obtain the multi-scale features of the downsampling network based on skip connections; and to perform at least one splicing and / or fusion based on one or more of the stable distribution features, the multi-scale features and the mapping control parameters to generate target WAT data.

[0019] In one embodiment, the step of inputting the initial noise vector and the conditional embedding vector into a preset diffusion model includes:

[0020] Obtain raw WAT sample data; the raw WAT sample data includes process data information and process condition information;

[0021] The process data information is cleaned and normalized, and the process condition information is uniquely encoded to obtain the target WAT sample data.

[0022] The target WAT sample data is used to train the diffusion model to obtain the preset diffusion model.

[0023] In one embodiment, training the diffusion model to be trained using the target WAT sample data includes:

[0024] The diffusion model to be trained is initialized based on preset hyperparameters;

[0025] Noise is added to the process data information to obtain disturbed process data information;

[0026] Based on the perturbation process data and the process condition information, the parameters of the diffusion model to be trained are iteratively updated through a reverse denoising process.

[0027] In one embodiment, obtaining the target WAT data corresponding to the target process conditions includes:

[0028] The target WAT data is evaluated based on at least one preset evaluation algorithm;

[0029] If the evaluation result of the target WAT data meets the preset conditions, then the target WAT data is output.

[0030] Secondly, this application provides a WAT ​​data generation apparatus, the apparatus comprising:

[0031] The condition processing module is used to preprocess the target process conditions in response to the WAT data generation request to obtain the condition embedding vector.

[0032] The noise determination module is used to determine the initial noise vector based on standard normal distribution sampling.

[0033] The data generation module is used to input the initial noise vector and the conditional embedding vector into a preset diffusion model to obtain target WAT data corresponding to the target process conditions; the preset diffusion model is a pre-trained diffusion model without classifier guidance.

[0034] Thirdly, this application provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0035] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0036] The aforementioned WAT data generation method, apparatus, computer equipment, and storage medium, in response to a WAT ​​data generation request, preprocess the target process conditions to obtain a conditional embedding vector, determine an initial noise vector based on standard normal distribution sampling, and input the initial noise vector and conditional embedding vector into a preset diffusion model to obtain target WAT data corresponding to the target process conditions. The preprocessing operation unifies the input format of process parameters, introduces generation diversity through standard normal distribution sampling while ensuring denoising stability, and combines a pre-trained classifier-free guided diffusion model to achieve condition-driven progressive data generation. This allows the generated results to maintain reasonable statistical variation while satisfying specific process constraints, avoiding pattern collapse and training instability issues. It can effectively simulate the WAT data distribution under small sample conditions, achieving the technical effect of improving the diversity and stability of the generated results. Attached Figure Description

[0037] Figure 1This is an application environment diagram of the WAT data generation method in one embodiment;

[0038] Figure 2 This is a flowchart illustrating a WAT ​​data generation method in one embodiment;

[0039] Figure 3 This is a flowchart illustrating the WAT data generation method in another embodiment;

[0040] Figure 4 This is a schematic diagram of the diffusion model in one embodiment;

[0041] Figure 5 This is a schematic diagram of the noise addition process in one embodiment;

[0042] Figure 6 This is a schematic diagram illustrating the quality assessment results of the data generated by each model in one embodiment;

[0043] Figure 7 This is a structural block diagram of a WAT ​​data generation device in one embodiment;

[0044] Figure 8 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0046] The WAT data generation method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or located in the cloud or on other network servers. Terminal 102, in response to a WAT ​​data generation request from server 104, can preprocess the target process conditions to obtain a conditional embedding vector; determine an initial noise vector based on standard normal distribution sampling; and input the initial noise vector and the conditional embedding vector into a preset diffusion model to obtain target WAT data corresponding to the target process conditions. The preset diffusion model is a pre-trained diffusion model without classifier guidance. Further, the preset diffusion model can be obtained by terminal 102 through communication with server 104. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, and tablets. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0047] In one embodiment, such as Figure 2 As shown, a WAT ​​data generation method is provided, which can be applied to... Figure 1 Taking terminal 102 as an example, the WAT data generation method includes:

[0048] Step S100: In response to the WAT data generation request, the target process conditions are preprocessed to obtain the condition embedding vector.

[0049] The target process conditions can be a set of process parameters involved in the integrated circuit manufacturing process. For example, the target process conditions may include one or more of continuous and discrete parameters. In an exemplary embodiment, the target process conditions may include, but are not limited to, one or more physical or chemical characteristic parameters such as doping concentration, etching time, and photoresist thickness, which can be obtained through production line measurements, process design documents, or simulation systems.

[0050] Conditional embedding vectors can be vectors used to characterize target process conditions. Their dimensions are consistent with the vector dimensions of the input parameters of the preset diffusion model, and they can be generated through preprocessing techniques such as data standardization, normalization, or feature engineering. In a specific embodiment, continuous parameters can be standardized to eliminate dimensional differences, while discrete parameters can be converted into dense vector representations through encoding or embedding.

[0051] Preprocessing the target process conditions can involve standardizing and vectorizing the original process parameters, thereby unifying heterogeneous parameters into a numerical format that the model can handle. This allows the generation process to effectively correlate with specific process conditions while reducing noise interference in the input data.

[0052] Step S200: Determine the initial noise vector based on standard normal distribution sampling.

[0053] The standard normal distribution can be a Gaussian distribution with a mean of 0 and a variance of 1.

[0054] The initial noise vector can be a random vector used to initiate the inverse generation process of the diffusion model, with the same dimension as the spatial dimension of the target WAT data. For example, when the WAT data contains 100 measurement features, the initial noise vector also has a dimension of 100. The initial noise vector can be generated by independent random sampling from a standard normal distribution.

[0055] Sampling based on a standard normal distribution can be exemplified by calling a random number generator to independently sample each dimension to construct a complete vector, thus ensuring the randomness required in the generation process. By utilizing the uniform distribution characteristic of the standard normal distribution, the bias in generating the initial state can be avoided, which helps to improve the stability of the subsequent denoising process.

[0056] Step S300: Input the initial noise vector and the conditional embedding vector into the preset diffusion model to obtain the target WAT data corresponding to the target process conditions.

[0057] The preset diffusion model can be a pre-trained diffusion model without classifier guidance.

[0058] The initial noise vector and conditional embedding vector are input into a preset diffusion model. For example, at each denoising step of the model, conditional prediction guided by the conditional embedding vector and / or unconditional prediction based solely on the noise input can be performed, and then linearly weighted and merged according to a preset ratio to obtain the corresponding WAT data. In an exemplary embodiment, the model uses 70% weight to reference the conditional guiding signal and 30% weight to retain the noise randomness, thus gradually completing the reverse diffusion process from noise to structured WAT data. By restoring the data structure through a gradual denoising mechanism, the output data both conforms to the target process constraints and retains reasonable natural variation characteristics. The classifier-free guidance mechanism avoids the training complexity caused by relying on an additional discriminator network, and by directly controlling the generation path through gradient direction optimization, the controllability and stability of the generation process can be improved.

[0059] This embodiment provides a WAT ​​data generation method that, in response to a WAT ​​data generation request, preprocesses the target process conditions to obtain a conditional embedding vector. Based on standard normal distribution sampling, an initial noise vector is determined. The initial noise vector and the conditional embedding vector are input into a preset diffusion model to obtain target WAT data corresponding to the target process conditions. The preprocessing operation unifies the input format of process parameters, introduces generation diversity and ensures denoising stability through standard normal distribution sampling, and combines a pre-trained classifier-free guided diffusion model to achieve condition-driven progressive data generation. This allows the generated results to maintain reasonable statistical variation while satisfying specific process constraints, avoiding pattern collapse and training instability. It can effectively simulate the WAT data distribution under small sample conditions, achieving the technical effect of improving the diversity and stability of the generated results.

[0060] In one embodiment, the target process conditions are preprocessed to obtain a conditional embedding vector, including:

[0061] Obtain classification features that match the preset diffusion model;

[0062] Based on classification features, the target process conditions are one-hot encoded to obtain the conditional embedding vector.

[0063] The classification features can be discrete process parameters that need to be treated as classification variables to characterize category information under different process conditions. For example, classification features can include, but are not limited to, one or more of the following: process step selection (e.g., etching, deposition), material type (e.g., metal, semiconductor), etching gas type, and photolithography process step selection.

[0064] Obtaining classification features that match the preset diffusion model can be achieved by obtaining the classification features used by the preset diffusion model during training to perform one-hot encoding of process conditions.

[0065] Conditional embedding vectors can be standardized vector forms used to represent target process conditions, with the same or compatible input structure as the diffusion model. They can be generated by one-hot encoding based on categorical features and integrating them into a fixed-dimensional vector representation. In an exemplary embodiment, the conditional embedding vector can be in binary vector form, where each dimension corresponds to a possible class value, set to 1 only at the corresponding class position and 0 otherwise. For example, when a parameter takes the value "process type B" and its value set has three classes, the corresponding encoding result is [0, 1, 0]. This vector form ensures that the discrete features of different process parameters are input into the model in a uniform format, avoiding the introduction of spurious order relationships due to numerical encoding.

[0066] Obtaining classification features that match the preset diffusion model can be achieved by analyzing the input condition dimension configuration of the diffusion model, clarifying which process parameters need to be treated as classification variables, and selecting the set of parameters to participate in the encoding, thereby ensuring that the dimension of the encoded vector is consistent with the dimension of the model input.

[0067] Based on classification features, one-hot encoding is performed on the target process conditions. For example, the current value of each classification parameter can be mapped to its predefined category space, with 1 set in the corresponding position and 0 set in the rest, thereby generating a fixed-dimensional binary encoded vector. By uniformly adopting one-hot encoding, not only can the preprocessing process be simplified, but a clear and unambiguous discrete feature representation can also be provided for the diffusion model.

[0068] This embodiment provides a WAT ​​data generation method that obtains classification features that match a preset diffusion model and performs one-hot encoding on target process conditions based on these features to obtain conditional embedding vectors. One-hot encoding converts discrete process parameters into fixed-dimensional binary vectors, ensuring the consistency between the mathematical representation of the classification features and their physical meaning, and avoiding spurious ordering relationships caused by numerical encoding. By pre-screening classification features that match the model input requirements, the compatibility between feature dimensions and model structure is ensured, reducing the risk of input errors. The generated conditional embedding vectors, as conditional inputs to the diffusion model, can accurately represent specific process combinations, strengthening the mapping relationship between generated data and process parameters. This achieves the technical effect of improving the accuracy and constraint consistency of generated data under multiple process conditions.

[0069] In one embodiment, the preset diffusion model includes a downsampling network, an upsampling network, and a conditional embedding network, wherein:

[0070] Downsampling networks are used to extract features from input data to obtain multi-scale features;

[0071] Conditional embedding networks are used to determine mapping control parameters based on conditional embedding vectors;

[0072] Upsampling networks are used to generate target WAT data based on multi-scale features and mapping control parameters.

[0073] The downsampling network can be a structured component in the diffusion model that implements feature encoding, used to perform layer-by-layer spatial dimension compression and semantic information abstraction on the input data. For example, the downsampling network can include one or more downsampling operation units such as convolutional layers, pooling layers, and attention mechanisms to perform hierarchical processing of the input data. For example, the downsampling network can include, but is not limited to, one or more of the following: residual downsampling modules, strided convolutional structures, and local self-attention blocks.

[0074] Multi-scale features can be feature representations with differentiated analytical granularity output from different network layers, with spatial resolution decreasing layer by layer while semantic abstraction increases accordingly. In an exemplary embodiment, multi-scale features can encompass a joint expression of high-frequency detail information and low-frequency structural information, such as in WAT data generation tasks, corresponding to the coupling relationship between different process parameters or the statistical dependence between multiple measurement indicators. Multi-scale features can be collected at the output of each layer through a downsampling network, and further, can be passed to subsequent modules using a skip connection mechanism.

[0075] Mapping control parameters can be dynamic parameters used to adjust the behavior of layers within the diffusion model, imposing conditional adjustments on neural network weights, activation responses, or attention distributions. Mapping control parameters can include, but are not limited to, one or more of the following: affine coefficients in layer normalization, query or key scaling factors in the attention mechanism, and modulation weights of deconvolution kernels. In an exemplary embodiment, mapping control parameters can be injected into the upsampling network at each time step of the diffusion process to achieve fine-grained control of the denoising path.

[0076] For example, the mapped control parameters can be obtained by performing a nonlinear mapping function through a conditional embedding network. For instance, a lightweight fully connected network can map the conditional embedding vector into a set of finely tuned control signals and decompose them into subsets of parameters applicable to different network layers. In one specific embodiment, this mapping process may include performing multiple linear transformations and nonlinear activations on the conditional embedding vector, ultimately outputting a set of control parameters that match each layer of the upsampled network, enabling process condition information to be transformed into kinetic adjustment factors that directly affect the generation process.

[0077] Target WAT data can be virtual measurement data that conforms to the electrical performance, geometric parameters, or defect distribution characteristics under specific process conditions. Its dimensions are consistent with real WAT data, and it can be used for process modeling or defect prediction tasks. Target WAT data can be gradually recovered from a noisy state through an upsampling network during the back diffusion process.

[0078] For example, the target WAT data can be generated by performing inverse feature reconstruction through an upsampling network. For instance, the spatial resolution of the feature map can be progressively improved through the synergistic effect of deconvolutional layers, interpolation upsampling, and skip connections, with control parameters from the conditional embedding network fused at each layer. In one specific embodiment, the upsampling network receives the current noise estimation features in each denoising step, supplements the information with multi-scale features from the corresponding layer, and modulates the activation response using mapping control parameters, ultimately outputting target WAT data that approximates the true distribution, thereby achieving a gradual transformation of noisy signals into structured data.

[0079] This embodiment provides a WAT ​​data generation method, which utilizes a pre-defined diffusion model including a downsampling network, an upsampling network, and a conditional embedding network. The downsampling network extracts features from the input data to obtain multi-scale features; the conditional embedding network determines mapping control parameters based on conditional embedding vectors; and the upsampling network generates target WAT data based on the multi-scale features and mapping control parameters. The downsampling network performs hierarchical compression on the input data to extract multi-scale features covering both high-frequency details and low-frequency structures, fully expressing the coupling relationships between process parameters and the statistical dependencies of measurement indicators. The conditional embedding network transforms the semantic information of process conditions into... The mapping control parameters involved in the diffusion process control enable dynamic adjustment of the generation path, ensuring that the statistical characteristics of the generated data remain consistent with the input process conditions. By using an upsampling network to fuse multi-scale features and mapping control parameters during the back diffusion process, and employing a dual-channel information processing mechanism to maintain the stability of the macroscopic structure of the data and the condition specificity of the microscopic parameters, the system enhances the responsiveness to process changes while maintaining high fidelity in the generated results. This allows for accurate simulation of the distribution characteristics of real WAT data under small sample conditions, improving the usability and reliability of the generated data in process parameter optimization and defect prediction tasks, and reducing computational overhead and data annotation requirements.

[0080] In one embodiment, the upsampling network includes a feature amplification module and a feature fusion module, wherein:

[0081] The feature magnification module is used to magnify multi-scale features to obtain stable distribution features;

[0082] The feature fusion module is used to obtain multi-scale features of the downsampled network based on skip connections; and to perform at least one concatenation and / or fusion based on one or more of stable distribution features, multi-scale features and mapping control parameters to generate target WAT data.

[0083] The stable distribution features can be feature representations that have undergone spatial resolution enhancement and possess controlled statistical properties. They can be obtained by performing inverse feature reconstruction operations on multi-scale features combined with normalization. For example, the spatial dimension of stable distribution features can be expanded through deconvolution layers, upsampling layers combined with convolutional layers, or pixel shuffling. Their mean and variance can be constrained through mechanisms such as batch normalization (BN) and layer normalization to stabilize their distribution. In an exemplary embodiment, if the input multi-scale features are a 16×16 feature map, after deconvolution processing, it can be expanded into a 32×32 feature map. Simultaneously, the introduction of the ReLU activation function and the normalization layer can work together to suppress the propagation of high-frequency noise. In a specific embodiment, during the time-step reversal process of the diffusion model, when the current feature map contains high noise residue, the stable distribution features can be forced to approximate a Gaussian distribution through layer normalization, thereby avoiding feature distortion or abnormal gradient fluctuations during the generation process.

[0084] Furthermore, the feature amplification module may include inverse reconstruction units such as transposed convolution, bilinear interpolation upsampling combined with convolutional layers, and pixel shuffle, thereby improving the feature space resolution while using normalization methods to regulate the statistical characteristics of feature distribution, reducing the noise amplification effect caused by the amplification process, and achieving stable restoration of the macroscopic structure.

[0085] Skip connections can be cross-layer pathways connecting corresponding layers of downsampling and upsampling networks, used to transfer multi-scale features extracted during downsampling. For example, skip connections can include the connection method in the U-Net architecture that transfers feature maps from the nth layer of the encoder to the nth layer of the decoder.

[0086] The multi-scale features of a downsampling network can be a combination of spatial local details and global semantic information captured during the encoding stage of the training phase, and can be extracted through the layer-by-layer downsampling process of a convolutional neural network. For example, these features may include one or more of the following: the edge contours of wafer surface defects, local doping uneven regions, spatial correlation patterns of process parameters, etc.

[0087] For example, the feature fusion module can integrate feature information from different levels by concatenating stable distribution features with downsampled features passed through skip connections in the channel dimension or by using weighted summation.

[0088] The concatenation and / or fusion operation can be a process of combining multiple feature tensors, which can be achieved through one or more of the following: channel concatenation, element-wise addition, element-wise multiplication, attention mechanisms, etc. For example, the concatenation and / or fusion operation can first concatenate stable distribution features with multi-scale features introduced by skip connections to form a high-dimensional fusion tensor, and then adjust the number of channels and remove redundant information through 1×1 convolution.

[0089] Mapping control parameters can be applied to the fused features in the form of channel attention, dynamically adjusting the response intensity of each feature channel. In one specific embodiment, if the mapping control parameters indicate that the current generation task focuses on the impact of doping concentration on electrical parameters, the fusion module can enhance the weights of doping-related feature channels through an attention mechanism while suppressing the responses of irrelevant variables. This process can be iterated multiple times, gradually refining the feature representation to ensure that the generated result statistically approximates the distribution of the real WAT data.

[0090] To generate target WAT data, for example, feature concatenation, normalization, convolution transformation, and conditional modulation can be performed sequentially at each upsampling level to output WAT data that conforms to physical laws and process constraints.

[0091] This embodiment provides a WAT ​​data generation method, which utilizes an upsampling network including a feature amplification module and a feature fusion module. The feature amplification module amplifies multi-scale features to obtain stable distribution features. The feature fusion module obtains multi-scale features from the downsampling network based on skip connections. Based on one or more of the stable distribution features, multi-scale features, and mapping control parameters, at least one concatenation and / or fusion is performed to generate target WAT data. The feature amplification module uses deconvolution or upsampling techniques to improve feature resolution and combines a normalization mechanism to stabilize feature distribution, suppress noise diffusion, and ensure the macroscopic consistency of the generated structure. The feature fusion module reuses the multi-scale features of the downsampling network through skip connections, retains local detail information, and performs channel concatenation or weighted fusion with the stable distribution features to avoid loss of details. Furthermore, by incorporating process conditions guided by mapping control parameters during the fusion process, the generated results can meet specific process constraints, making the final output target WAT data closer to the real measurement results in terms of spatial structure, statistical characteristics, and process relevance. Especially under small sample conditions, it can still generate high-quality data with reasonable parameter fluctuations and boundary behaviors, which can be used to enhance the training robustness and generalization ability of yield prediction models.

[0092] In one embodiment, the process includes the following steps before inputting the initial noise vector and conditional embedding vector into a preset diffusion model:

[0093] Obtain raw WAT sample data; raw WAT sample data includes process data information and process condition information;

[0094] The process data information is cleaned and normalized, and the process condition information is uniquely encoded to obtain the target WAT sample data.

[0095] The target WAT sample data is used to train the diffusion model to obtain the preset diffusion model.

[0096] The raw WAT sample data can be a set of wafer test data actually collected during the integrated circuit manufacturing process, used to reflect the physical characteristics and measurement results of the wafer under specific process conditions. The raw WAT sample data can be extracted from the test equipment or database in the production line. For example, the raw WAT sample data may include, but is not limited to, measured values ​​of electrical or structural parameters such as resistivity, thin film thickness, and critical dimension (CD), as well as the corresponding process parameter settings.

[0097] Process data information can be quantitative measurement results reflecting the physical state of the wafer, which can be obtained through online or offline testing equipment. For example, process data information may include one or more of resistivity, dielectric layer thickness, doping concentration, etc. In a specific embodiment, process data information includes electrical performance parameters such as current, resistance, capacitance, threshold voltage, and leakage current.

[0098] Process condition information can be the configuration of process parameters used during data acquisition, which can be obtained through a process recipe system or equipment logs. For example, process condition information may include one or more parameters such as temperature, pressure, exposure dose, etching gas type, and photoresist type. These parameters can be continuous or discrete variables. In a specific embodiment, process condition information includes: process parameters associated with the corresponding wafer testing, such as deposition temperature, gas flow rate, chamber pressure, RF power, etching time, CMP polishing conditions, etc.; metadata such as wafer batch number, test timestamp, test instrument number, etc.

[0099] Cleaning and normalizing process data can include methods such as removing outliers that exceed physically reasonable ranges, filling in missing data using interpolation or mean substitution, and identifying and correcting deviations caused by measurement noise. Normalization can map the cleaned process data to a uniform numerical range, such as [0, 1] or [-1, 1], through standardization or interval scaling, thereby eliminating the impact of differences in parameter magnitudes on model training.

[0100] For example, discrete parameters in process condition information can be transformed using one-hot encoding, which maps each categorical variable to a binary vector. For continuous process condition parameters, an embedding layer can be used to map them into dense low-dimensional vectors to preserve the potential correlations between their values.

[0101] The target WAT sample data can be generated by performing the aforementioned cleaning, normalization, and encoding operations on the original WAT sample data, and its form can be a structured composite vector. In a specific embodiment, the target WAT sample data can be formed by collecting each filtered process data information and its corresponding process condition information.

[0102] The diffusion model to be trained can be a diffusion model with a U-Net architecture as its core network structure, which may include an encoder-decoder structure to capture the spatial dependencies of high-dimensional data. This model can optimize parameters using the backpropagation algorithm and complete the training process using target WAT sample data. Modeling the forward diffusion process can involve gradually adding Gaussian noise to the target WAT sample data during the training phase, generating a series of intermediate states from the original data to pure noise. The inverse denoising process can be learned by predicting the noise component added at each step using a neural network and updating the model weights based on the prediction error (such as mean squared error), enabling the model to gradually master the ability to recover the original data distribution from the noise.

[0103] This embodiment provides a WAT ​​data generation method, which involves acquiring raw WAT sample data, including process data information and process condition information; cleaning and normalizing the process data information, and performing one-hot encoding on the process condition information to obtain target WAT sample data; training a diffusion model with the target WAT sample data to obtain a preset diffusion model; reducing noise interference in the process data and unifying the dimensions through cleaning and normalization; converting discrete and continuous process condition parameters into numerical vectors that the model can process through one-hot encoding and embedding layers; enabling the model to reconstruct the true data distribution from noise through forward diffusion modeling and inverse denoising learning; enhancing the controllability and diversity of the generation process through joint training of conditional and unconditional paths; and improving training stability by combining a classifier-free guided architecture. This results in generated WAT data that reflects the statistical characteristics of specific process conditions while possessing reasonable variation capabilities, supporting closed-loop iteration of data augmentation and process optimization in small sample scenarios. This achieves the technical effects of improving the training efficiency of downstream models, enhancing generalization ability, and accelerating the process decision-making cycle.

[0104] In one embodiment, training the diffusion model to be trained using the target WAT sample data includes:

[0105] The diffusion model to be trained is initialized based on the preset hyperparameters;

[0106] Noise is added to the process data to obtain disturbed process data.

[0107] Based on the perturbation process data and process condition information, the parameters of the diffusion model to be trained are iteratively updated through a reverse denoising process.

[0108] The preset hyperparameters can be configurable parameters used to control the model training process, and may include, but are not limited to, one or more of the following: learning rate, batch size, optimizer type, and diffusion time steps. The learning rate can be a coefficient that determines the step size of the model parameter updates; the batch size can be the number of samples participating in gradient calculation in a single forward propagation; the optimizer type can be an algorithm used to execute the parameter update strategy, such as one of the Adam optimizers; and the diffusion time steps can be the number of time discretization steps defining the noise addition and denoising process.

[0109] Initialization of the diffusion model to be trained can involve assigning initial values ​​to the weights of each layer of the model to establish a reasonable parameter distribution. For example, the initialization of the diffusion model to be trained can be achieved through a specific weight initialization algorithm. In a specific embodiment, one or more of the He initialization or Xavier initialization methods can be used.

[0110] Noise can be added to process data. The noise can be a random signal used to perturb the original data; for example, Gaussian white noise can be used. The noise addition process can be an operation of gradually injecting noise into the process data, achieved through a step-by-step diffusion strategy. Furthermore, noise can be injected according to a preset formula during the diffusion time step.

[0111] The perturbation process data information can be a hybrid representation carrying the original data features and superimposed noise. The noise addition process can be carried out through a progressive injection mechanism, constructing a continuous path from clean data to pure noise, enabling the model to learn the data distribution characteristics at different noise levels, and avoiding the loss of key features due to a one-time high-intensity noise perturbation.

[0112] The parameters of the diffusion model to be trained are iteratively updated through a reverse denoising process. This can be a process in which the model predicts and removes noise based on noisy data and time step information, gradually restoring the original data distribution. In a specific embodiment, noise prediction can be accomplished by inputting perturbation process data information and time step embedding vectors into the diffusion model to be trained. The time step embedding vectors can be generated through sinusoidal position encoding, enabling the model to perceive the current denoising stage and thus output predicted noise.

[0113] Conditional information fusion can involve processing process condition information through a conditional adaptation layer and then integrating it with the noise prediction results at the feature level. In an exemplary embodiment, a multilayer perceptron (MLP) can be used to map the process condition vector into a feature representation consistent with the noise prediction dimension, and then the data can be fused using element-wise multiplication or concatenation. Through loss calculation and parameter updates, the difference between the predicted noise and the actual noise is analyzed, and the gradient is calculated using a backpropagation algorithm. An optimizer (such as Adam) is used to iteratively adjust the model parameters, gradually improving the model's ability to accurately predict noise under different time steps and process conditions.

[0114] This embodiment provides a WAT ​​data generation method, which initializes the diffusion model to be trained based on preset hyperparameters; adds noise to the process data information to obtain perturbed process data information; based on the perturbed process data information and process condition information, iteratively updates the parameters of the diffusion model to be trained through a reverse denoising process. By reasonably configuring the hyperparameters, the model parameters can have good distribution characteristics in the early stage of training; by combining the perturbed data with time step embedding for noise prediction, and fusing process condition information to achieve feature-level condition-guided parameter iterative updates, the model can achieve stable convergence in complex high-dimensional process data space. This achieves the technical effect of improving the training efficiency and generation quality of the diffusion model in the WAT data generation task, enhancing the correlation between generated samples and real process parameters, and effectively improving data utilization efficiency under small sample conditions.

[0115] In one embodiment, after obtaining the target WAT data corresponding to the target process conditions, the following steps are included:

[0116] The target WAT data is evaluated based on at least one preset evaluation algorithm;

[0117] If the evaluation results of the target WAT data meet the preset conditions, the target WAT data will be output.

[0118] The target WAT data can be a set of test data reflecting the electrical parameters of the device under specific process conditions during semiconductor manufacturing, used to characterize the consistency and stability of the process execution results. In this embodiment, it can be obtained by combining a diffusion model generation method with the target process conditions as input.

[0119] The preset evaluation algorithm can be a mathematical or statistical method used to quantify the quality of generated data. It can measure the degree of fit between the target WAT data and the real WAT dataset or established process standards, and can be performed by calling a pre-configured evaluation module.

[0120] For example, the preset evaluation algorithm may include, but is not limited to, one or more of the following: statistical distribution consistency assessment, process parameter correlation verification, and domain-specific validity indicators. In one exemplary embodiment, the preset evaluation algorithm may include, but is not limited to, maximum mean difference, Wassertein distance, mean squared error, and mean absolute error.

[0121] Preset conditions can be criteria used to determine whether the target WAT data is qualified, and can be constructed by setting thresholds or combining rules. Based on at least one preset evaluation algorithm, the target WAT data is evaluated. For example, this can be done through statistical distribution tests, correlation analysis, and simulation verification to obtain multi-dimensional evaluation results, which are used to comprehensively measure the distribution characteristics, parameter correlation structure, and physical rationality of the target WAT data.

[0122] If the evaluation result of the target WAT data meets the preset conditions, the target WAT data is output. For example, the data output process is triggered when all indicators meet the preset thresholds or rule combinations. Otherwise, the system can send back an evaluation failure signal and start a regeneration mechanism, such as adjusting the noise vector in the diffusion model or optimizing the model parameters and then executing the generation step again.

[0123] This embodiment provides a WAT ​​data generation method that evaluates target WAT data based on at least one preset evaluation algorithm and outputs the target WAT data when the evaluation result meets preset conditions. By introducing one or more quantitative mechanisms such as statistical distribution consistency evaluation, process parameter correlation verification, and domain-specific effectiveness indicators, combined with configurable qualification criteria, the generated data is consistent with the real process data in terms of probability distribution, parameter correlation structure, and physical realizability. This improves the applicability of the synthesized data in downstream tasks, thereby achieving the technical effects of improving the credibility of the generated data, ensuring engineering practicality, and reducing the risk of decision-making bias caused by data distortion.

[0124] To more clearly illustrate the technical solution of this application, a detailed embodiment is also provided.

[0125] In one embodiment, a WAT ​​data generation method is provided, comprising:

[0126] Step 1: WAT data collection and initial preparation.

[0127] like Figure 3As shown, raw WAT data is acquired through a wafer acceptance test system. This data includes, but is not limited to: electrical performance parameters such as current (I), resistance (R), capacitance (C), threshold voltage (Vth), and leakage current (I_leak); process condition labels: process parameters associated with the corresponding wafer test, such as deposition temperature, gas flow rate, chamber pressure, RF power, etching time, and CMP polishing conditions; and metadata such as wafer batch number, test timestamp, and test instrument number (optional for subsequent traceability or online updates). The above data is stored in a database or file system in a record format, ensuring that the data can be retrieved by batch or process condition. A preliminary screening of all raw features is performed: key electrical parameters closely related to WAT performance are pre-selected based on domain expert knowledge; for each selected feature, outliers (such as exceeding the mean ± 3 standard deviations) are identified and removed using a box plot method; a cleaned raw WAT sample set is obtained, containing electrical performance vectors and corresponding process condition labels. The electrical performance characteristics after cleaning were normalized, mainly using max-min normalization, mapping each feature value to the interval [0, 1].

[0128]

[0129] Where, x i，n Let x' represent the i-th data point out of n data samples, where max represents the maximum value and min represents the minimum value. i，n This represents the normalized value.

[0130] Furthermore, the process condition labels are encoded using one-hot encoding, transforming the process conditions into a fixed-dimensional condition vector c. Finally, an initial training set is generated, integrating the normalized electrical performance vector x0 with the condition vector c to form training sample pairs (x0, c). All samples are divided into training and validation sets (e.g., 80% for training and 20% for validation) to monitor overfitting and evaluate the generation effect during training.

[0131] Step 2: Model initialization and hyperparameter setting.

[0132] Set the maximum number of time steps T (400) and set the noise scheduling mode to cosine scheduling.

[0133] Build a denoising network based on U-net, such as Figure 4 As shown, the overall architecture of the model consists of three parts: downsampling, upsampling, and conditional embedding.

[0134] Downsampling employs ResBlocks as the primary feature extraction unit and enhances the model's expressive power through multi-scale information fusion and conditional embedding mechanisms. In the downsampling stage, the input data first undergoes preliminary feature extraction using ResBlocks, followed by max pooling to reduce resolution and extract more abstract representational information. An additional path uses average pooling combined with the GeLU activation function as an auxiliary feature extraction path, thereby enhancing the ability to capture multi-scale information. The outputs of these two branches serve as input for subsequent network processing.

[0135] After downsampling, the features enter the upsampling stage to gradually restore spatial resolution. Upsampling consists of a two-stage structure: the first stage, Upsampling 1, uses transposed convolution (Conv Transpose) for initial amplification and combines GroupNorm and ReLU for normalization and nonlinear transformation to stabilize the feature distribution. The second stage, Upsampling 2, further refines the features using transposed convolution (Conv Transpose) and ResBlock, and then uses transposed convolution for final upsampling. This part is repeated twice to enhance feature expressiveness. During upsampling, the model concatenates features from the downsampling stage through skip connections, thus fully integrating low-level and high-level information. Furthermore, during feature extraction, the model introduces additional control information through conditional embedding layers to enhance its expressive power. This part, composed of Linear, GeLU, and Linear layers, is responsible for mapping conditional variables to a suitable feature space and fusing them with the main features during upsampling. The conditional information here mainly refers to the diffusion time step t and the data class label c. The model can predict diffusion noise more reasonably and accurately based on time step information, while the category labels will be used for subsequent conditional control to generate WAT data of the corresponding category.

[0136] The experiment used an NVIDIA GeForce RTX4060 graphics card and a Windows 11 operating system. The network model was coded in Python 3.11 using the PyTorch deep learning framework. The network underwent 200 iterations using the AdamW optimizer.

[0137] Step 3: Model training.

[0138] like Figure 5 As shown, in the forward diffusion stage (noise addition), for each real sample x0 and corresponding condition c in the training set, the perturbation data is generated by uniformly sampling time steps t∈{1,2,…,T} according to the formula:

[0139]

[0140] Through multi-step diffusion, the data distribution q(x0) is gradually blurred, finally reaching the standard Gaussian distribution N(0, I) at time T→∞. At each step, the data x... t-1 It will be shrunk (multiplied) ), and add noise N(0, β) t I), which causes the data to gradually lose its features. The reverse generation process aims to start from x. T Starting from ~N(0, I), the original data distribution q(x0) is restored through stepwise denoising. The training objective is to enable the neural network to accurately predict the noise at each time step, i.e., to minimize the loss as much as possible. A condition c is explicitly introduced into the diffusion model, allowing the noise prediction network to directly learn ε. θ (x t Therefore, the optimization objective of the model is as follows:

[0141]

[0142] Step 4: Sampling generation process.

[0143] Sample the initial noise vector x from the standard normal distribution. T ~N(0, I). The vector dimension is consistent with the WAT data dimension (e.g., d-dimensional). Determine the target process condition c, and perform the same normalization / encoding processing on c as during training to generate a conditional embedding vector; then set the current x... T The time step T, the guiding strength s and the condition c are embedded into the input noise prediction network to generate WAT data iteratively.

[0144] Step 5: Sample Generation, Evaluation, and Screening. Based on the same real-world dataset, we generated data using four different generation models and the diffusion model proposed in this study, and evaluated them using MMD, WD, MAE, and MSE. The formulas for the four indicators are as follows:

[0145] (1) Maximum mean discrepancy (MMD):

[0146]

[0147] Among them, y i and Let represent the i-th real data sample and the synthetic data sample; M represents the total amount of data; k(-) represents the Gaussian kernel function.

[0148] (2)Wasserstein distance(WD):

[0149]

[0150] F1(z) and F2(z) are the cumulative distribution functions of the real data and the synthetic data, respectively. It is worth noting that for these two metrics, the smaller the value, the better the data generation performance.

[0151] (3) Mean Squared Error (MSE):

[0152]

[0153] Where n is the number of samples, y i It is the actual value. These are predicted values. Mean squared error (MSE) represents the average of the squared differences between predicted and actual values. MSE squares the error, making larger errors account for a larger proportion of the indicator, and is more sensitive to outliers. It is often used to measure the overall performance of regression models.

[0154] (4) Mean Absolute Error (MAE):

[0155]

[0156] MAE is the average of the absolute values ​​of the differences between predicted and actual values. MAE is relatively less sensitive to outliers and, compared to MSE, focuses more on measuring the average deviation between predicted and actual values.

[0157] like Figure 6 As shown, the quality assessment results of the generated data from different generative models are presented. The experimental results show that VAE performs relatively evenly in terms of MMD (0.057), WD (0.015), MAE (0.022), and MSE (0.009), with the distribution of the generated data closely resembling the real data. infoGAN performs poorly across all metrics, with an MMD as high as 0.306 and a WD of 0.053, indicating a significant difference in the distribution of its generated data from the real data and poor quality. acGAN's MMD (0.855) and WD (0.099) are significantly higher than other models, indicating a large deviation between its generated data and the real data, possibly related to mode collapse. cGAN has the lowest WD (0.009) and a relatively low MMD (0.046), showing that its generated data has a certain degree of good distribution consistency, but its MAE (0.023) and MSE (0.014) are still slightly insufficient compared to the diffusion model.

[0158] This embodiment provides a WAT ​​data generation method that utilizes the forward and inverse processes of a diffusion model to approximate the true WAT data distribution in a high-dimensional space. It also employs a Classifier-Free Guidance mechanism to achieve conditional generation without a discriminator. By combining normalized embedding, residual network structures, and multi-scale feature extraction, the system can generate high-fidelity, highly diverse synthetic data, improving the training performance of downstream prediction models. This method offers advantages such as high-quality data generation, strong controllability, and good training stability, making it suitable for scenarios such as WAT data augmentation, process modeling, and yield analysis.

[0159] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0160] Based on the same inventive concept, this application also provides a WAT ​​data generation apparatus for implementing the WAT data generation method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more WAT data generation apparatus embodiments provided below can be found in the limitations of the WAT data generation method described above, and will not be repeated here.

[0161] In one embodiment, such as Figure 7 As shown, a WAT ​​data generation apparatus is provided, the apparatus comprising:

[0162] The condition processing module 100 is used to preprocess the target process conditions in response to the WAT data generation request to obtain the condition embedding vector.

[0163] The noise determination module 200 is used to determine the initial noise vector based on standard normal distribution sampling.

[0164] The data generation module 300 is used to input the initial noise vector and conditional embedding vector into a preset diffusion model to obtain target WAT data corresponding to the target process conditions. The preset diffusion model is a pre-trained diffusion model without classifier guidance.

[0165] In some embodiments, the condition processing module 100 is further configured to:

[0166] Obtain classification features that match the preset diffusion model;

[0167] Based on the classification features, the target process conditions are one-hot encoded to obtain a conditional embedding vector.

[0168] In some embodiments, the preset diffusion model includes a downsampling network, an upsampling network, and a conditional embedding network, wherein:

[0169] The downsampling network is used to extract features from the input data to obtain multi-scale features;

[0170] The conditional embedding network is used to determine the mapping control parameters based on the conditional embedding vector;

[0171] The upsampling network is used to generate target WAT data based on the multi-scale features and the mapping control parameters.

[0172] In some embodiments, the upsampling network includes a feature amplification module and a feature fusion module, wherein:

[0173] The feature amplification module is used to amplify the multi-scale features to obtain stable distribution features;

[0174] The feature fusion module is used to obtain the multi-scale features of the downsampling network based on skip connections; and to perform at least one splicing and / or fusion based on one or more of the stable distribution features, the multi-scale features and the mapping control parameters to generate target WAT data.

[0175] In some embodiments, the apparatus further includes a model training module for:

[0176] Obtain raw WAT sample data; the raw WAT sample data includes process data information and process condition information;

[0177] The process data information is cleaned and normalized, and the process condition information is uniquely encoded to obtain the target WAT sample data.

[0178] The target WAT sample data is used to train the diffusion model to obtain the preset diffusion model.

[0179] In some embodiments, the model training module is also used for:

[0180] The diffusion model to be trained is initialized based on preset hyperparameters;

[0181] Noise is added to the process data information to obtain disturbed process data information;

[0182] Based on the perturbation process data and the process condition information, the parameters of the diffusion model to be trained are iteratively updated through a reverse denoising process.

[0183] In some embodiments, the apparatus further includes a data evaluation module for:

[0184] The target WAT data is evaluated based on at least one preset evaluation algorithm;

[0185] If the evaluation result of the target WAT data meets the preset conditions, then the target WAT data is output.

[0186] Each module in the aforementioned WAT data generation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0187] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 8 As shown, the computer device includes 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, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements a WAT ​​data generation method. The display screen can be an LCD screen or an e-ink display screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0188] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0189] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the WAT data generation method of any of the above embodiments:

[0190] In response to the WAT data generation request, the target process conditions are preprocessed to obtain the condition embedding vector;

[0191] The initial noise vector is determined based on sampling from a standard normal distribution.

[0192] The initial noise vector and the conditional embedding vector are input into a preset diffusion model to obtain target WAT data corresponding to the target process conditions; the preset diffusion model is a pre-trained diffusion model without classifier guidance.

[0193] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the WAT data generation method of any of the above embodiments:

[0194] In response to the WAT data generation request, the target process conditions are preprocessed to obtain the condition embedding vector;

[0195] The initial noise vector is determined based on sampling from a standard normal distribution.

[0196] The initial noise vector and the conditional embedding vector are input into a preset diffusion model to obtain target WAT data corresponding to the target process conditions; the preset diffusion model is a pre-trained diffusion model without classifier guidance.

[0197] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0198] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0199] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0200] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for generating WAT data, characterized in that, The WAT data generation method includes: In response to the WAT data generation request, the target process conditions are preprocessed to obtain the condition embedding vector; The initial noise vector is determined based on sampling from a standard normal distribution. The initial noise vector and the conditional embedding vector are input into a preset diffusion model to obtain target WAT data corresponding to the target process conditions; the preset diffusion model is a pre-trained diffusion model without classifier guidance.

2. The WAT data generation method according to claim 1, characterized in that, The preprocessing of the target process conditions to obtain the conditional embedding vector includes: Obtain classification features that match the preset diffusion model; Based on the classification features, the target process conditions are one-hot encoded to obtain a conditional embedding vector.

3. The WAT data generation method according to claim 1, characterized in that, The preset diffusion model includes a downsampling network, an upsampling network, and a conditional embedding network, wherein: The downsampling network is used to extract features from the input data to obtain multi-scale features; The conditional embedding network is used to determine the mapping control parameters based on the conditional embedding vector; The upsampling network is used to generate target WAT data based on the multi-scale features and the mapping control parameters.

4. The WAT data generation method according to claim 3, characterized in that, The upsampling network includes a feature amplification module and a feature fusion module, wherein: The feature amplification module is used to amplify the multi-scale features to obtain stable distribution features; The feature fusion module is used to obtain the multi-scale features of the downsampling network based on skip connections; and to perform at least one splicing and / or fusion based on one or more of the stable distribution features, the multi-scale features and the mapping control parameters to generate target WAT data.

5. The WAT data generation method according to claim 1, characterized in that, The step of inputting the initial noise vector and the conditional embedding vector into the preset diffusion model includes: Obtain raw WAT sample data; the raw WAT sample data includes process data information and process condition information; The process data information is cleaned and normalized, and the process condition information is uniquely encoded to obtain the target WAT sample data. The target WAT sample data is used to train the diffusion model to obtain the preset diffusion model.

6. The WAT data generation method according to claim 5, characterized in that, The step of training the diffusion model to be trained using the target WAT sample data includes: The diffusion model to be trained is initialized based on preset hyperparameters; Noise is added to the process data information to obtain disturbed process data information; Based on the perturbation process data and the process condition information, the parameters of the diffusion model to be trained are iteratively updated through a reverse denoising process.

7. The WAT data generation method according to claim 1, characterized in that, After obtaining the target WAT data corresponding to the target process conditions, the process includes: The target WAT data is evaluated based on at least one preset evaluation algorithm; If the evaluation result of the target WAT data meets the preset conditions, then the target WAT data is output.

8. A WAT data generation device, characterized in that, The device includes: The condition processing module is used to preprocess the target process conditions in response to the WAT data generation request to obtain the condition embedding vector. The noise determination module is used to determine the initial noise vector based on standard normal distribution sampling. The data generation module is used to input the initial noise vector and the conditional embedding vector into a preset diffusion model to obtain target WAT data corresponding to the target process conditions; the preset diffusion model is a pre-trained diffusion model without classifier guidance.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images