Soft sensor technology data generation method and system based on diffusion model

Abstract

The invention discloses a soft sensor technical data generation method and system based on a diffusion model, and belongs to the field of soft sensor technical data generation. Collecting data in the wafer manufacturing process as a training set, wherein the data comprises process data and measurement data of semiconductor key manufacturing steps; iteratively training a regression enhancement diffusion model by using the training set; in the iterative training process, after each round of training is completed, screening part of the virtual sample subsets according to the non-dominated sorting result and the congestion degree result of the virtual sample subsets, and adding the virtual sample subsets into a training set; and finally, taking random Gaussian noise as input, generating virtual process data by utilizing a return denoising module of the trained regression enhancement diffusion model, and taking a generation result as soft sensor technical data of the semiconductor key manufacturing step. According to the method, the capability of capturing the complex mapping relation by the model is enhanced, the representative high-quality synthetic training sample can be generated for training the soft sensor model, and the performance of the existing soft sensor model is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of soft sensor technology data generation, and in particular relates to a soft sensor technology data generation method and system based on a diffusion model. Background Art

[0002] In the field of semiconductor manufacturing, precise control of the production process is essential to ensure high product quality and operational efficiency. A common industrial practice is to perform a metrology step, which involves sampling and measuring physical features (such as film thickness) after key process steps. These measurements are used to guide subsequent process adjustments aimed at optimizing production results. However, this approach usually relies on limited sample data, usually only sampling 2 to 3 wafers per batch. This data limitation leads to inaccurate production process control and may cause quality fluctuations. In addition, frequent physical measurement steps are not only time-consuming, but also require expensive metrology equipment, which increases production costs and cycle time and reduces overall production efficiency. Soft sensor technology or virtual metrology provides an innovative solution to estimate critical dimensions and material properties without direct physical measurement. This reduces the reliance on expensive metrology tools and reduces the frequency of expensive measurement steps. Nevertheless, reducing the number of physical measurements limits the availability of labeled training samples required to develop effective soft sensor models.

[0003] Recent research has explored data generation methods to address these challenges, but existing solutions still suffer from the following problems:

[0004] 1. Insufficient quality and diversity of generated data: Existing data generation methods usually lack diversity, the quality of generated samples is low, and they cannot fully simulate actual production conditions. This limits the effectiveness of generated data in training soft sensor models, resulting in limited improvement in model performance. At the same time, due to the insufficient quality and diversity of generated data, the trained soft sensor model has poor generalization ability when facing new data and is difficult to adapt to changes and new conditions in the production process.

[0005] 2. Insufficient learning of complex mapping relationships and data distribution: Traditional generative models perform poorly in capturing the complex mapping relationships between process variables and target variables, resulting in the inability of generated virtual samples to accurately reflect changes in actual production. In addition, existing models find it difficult to simultaneously learn the data distribution of features and target variables, which results in the lack of realism and representativeness of the generated data, affecting the model's prediction accuracy.

[0006] 3. Lack of effective optimization and selection mechanism: The existing technology lacks an effective mechanism to optimize the selection of generated data and cannot ensure that the selected data achieves the best balance in distribution and diversity. This deficiency affects the training effect of the model and cannot fully utilize the generated virtual samples to improve the performance of the soft sensor model. Summary of the invention

[0007] In order to solve the defects of the existing soft sensor technology data generation method, the present invention proposes a soft sensor technology data generation method and system based on a diffusion model. The diffusion model is a regression enhanced diffusion model, which is named REDM (Regression-Enhanced Diffusion Model), which is suitable for virtual data generation of soft sensor technology and can widely improve the prediction performance of existing soft sensor models.

[0008] The technical solution adopted by the present invention is as follows:

[0009] In a first aspect, the present invention proposes a method for generating data of soft sensor technology based on a diffusion model, comprising the following steps:

[0010] (1) Collect data from the wafer manufacturing process, including process data and measurement data of key semiconductor manufacturing steps, and use the collected process data and measurement data as training sets;

[0011] (2) Iteratively train the regression enhanced diffusion model using the training set;

[0012] The regression enhancement diffusion model includes a diffusion denoising module, a backhaul denoising module and a regression enhancement module based on a multi-head attention mechanism. The diffusion denoising module gradually adds random noise to the process data of the training set to obtain pure Gaussian noise data. The backhaul denoising module gradually denoises and restores the pure Gaussian noise data to generate virtual process data. The regression enhancement module generates virtual measurement data according to the real process data and the virtual process data, respectively, and updates the regression enhancement diffusion model parameters according to the loss of the virtual measurement data and the real measurement data, the backhaul denoising loss and the Lipschitz constraint loss.

[0013] In the process of iterative training of the regression enhanced diffusion model, the training set is updated after each round of training. The updating method is as follows: the virtual process data generated by the feedback denoising module is divided into virtual sample subsets, and some virtual sample subsets are selected according to the non-dominated sorting results and crowding results of the virtual sample subsets and added to the training set;

[0014] (3) Using random Gaussian noise as input, the feedback denoising module of the trained regression enhanced diffusion model is used to generate virtual process data, and the generated results are used as soft sensor technology data for key semiconductor manufacturing steps.

[0015] Furthermore, the key semiconductor manufacturing steps include a chemical vapor deposition step, an etching step, and a chemical mechanical polishing step.

[0016] Furthermore, the process data of the key semiconductor manufacturing steps is a series of sensor values.

[0017] Furthermore, the regression enhancement module includes an encoder, a pre-trained multi-head attention module, and a linear prediction layer; the real process data and the virtual process data are encoded by the encoder respectively, and then the features are extracted by the pre-trained multi-head attention module, and finally the virtual measurement data is generated by the linear prediction layer.

[0018] Furthermore, in the process of iterative training of the regression enhanced diffusion model, the loss functions of the backhaul denoising module and the regression enhancement module are:

[0019]

[0020]

[0021] in, and are the total losses of the backhaul denoising module and the regression enhancement module, respectively. is the loss between the predicted noise and the actual added noise in the backhaul denoising process, It is the loss between the virtual measurement data generated based on the prediction of the real process data and the real measurement data. It is the loss between the virtual measurement data generated by the virtual process data prediction and the real measurement data. is the Lipschitz constraint loss, is the loss coefficient.

[0022] Furthermore, the method of selecting a portion of the virtual sample subsets according to the non-dominated sorting results and the congestion results of the virtual sample subsets includes:

[0023] Calculate the Sliced-Wasserstein distance between each virtual sample subset and the training set;

[0024] Calculate the sample similarity score within each virtual sample subset;

[0025] Based on the Sliced-Wasserstein distance and sample similarity score of each virtual sample subset, the virtual sample subsets are non-dominated sorted to obtain a Pareto front sequence, where each virtual sample subset corresponds to a Pareto front level;

[0026] Calculate the crowding distance of each virtual sample subset in the Pareto front sequence;

[0027] Starting from the first level of the Pareto frontier, virtual sample subsets are selected. If the number of virtual sample subsets at the current level does not meet the screening quantity requirement, the selection continues from the next level. For virtual sample subsets at the same Pareto frontier level, subsets with larger crowding distances are selected first until the screening quantity requirement is met.

[0028] In a second aspect, the present invention proposes a soft sensor technology data generation system based on a diffusion model, which is used to implement the above-mentioned soft sensor technology data generation method based on a diffusion model.

[0029] In a third aspect, the present invention proposes an electronic device, comprising a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement the above-mentioned soft sensor technology data generation method based on the diffusion model.

[0030] In a fourth aspect, the present invention proposes a machine-readable storage medium, which stores machine-executable instructions. When the machine-executable instructions are called and executed by a processor, they are used to implement the above-mentioned soft sensor technology data generation method based on the diffusion model.

[0031] The beneficial effects of the present invention are:

[0032] (1) The present invention proposes a regression enhanced diffusion model, which can be used to generate representative high-quality synthetic training samples. The generated soft sensor technology data of key semiconductor manufacturing steps are used for training soft sensor models, which can significantly improve the performance of existing soft sensor models.

[0033] (2) The present invention introduces a multi-head attention mechanism into the diffusion model to enhance the model's ability to capture complex mapping relationships.

[0034] (3) The present invention divides the virtual process data generated by the feedback denoising module into virtual sample subsets, selects virtual samples with both distribution similarity and diversity according to the non-dominated sorting results and congestion results of the virtual sample subsets, optimizes the synthetic samples and incorporates them into the training set to participate in the training, and ensures that the selected data achieves the best balance in distribution and diversity. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A schematic flow chart of a method for distortion correction and matching of unsupervised scanning electron microscope images according to an embodiment of the present invention;

[0036] Figure 2 A schematic diagram of the training process of the REDM model shown in an embodiment of the present invention;

[0037] Figure 3 A schematic diagram of a sample screening process according to an embodiment of the present invention;

[0038] Figure 4 Comparison chart of target distribution between generated virtual samples and real samples. DETAILED DESCRIPTION

[0039] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention.

[0040] The accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0041] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the steps. For example, some steps may be decomposed, while some steps may be combined or partially combined, so the actual execution order may change according to the actual situation.

[0042] like Figure 1 As shown in FIG. 1 , the distortion correction matching method of the unsupervised scanning electron microscope image mainly includes the following steps:

[0043] S1, collects real-time data of the fault detection and classification system (FDC system) in the wafer manufacturing process, including process data and measurement data of key semiconductor manufacturing steps, and regards the process data as process variables and the measurement data as target variables.

[0044] The key semiconductor manufacturing steps include a chemical vapor deposition step (CVD for short), an etching step (ETCH for short), and a chemical mechanical polishing step (CMP for short).

[0045] The critical semiconductor manufacturing step process data (process variables) are a series of sensor data.

[0046] In one embodiment of the present invention, for example, a CVD process collects lateral / top silane ( ) parameters such as gas flow, top / bias / side RF reflection power, deposition time, front pressure, etc. are used as process data, and the final deposition thickness is used as measurement data; for example, in the ETCH process, taking the polysilicon etching process as an example, high-energy plasma is used to impact the surface of the semiconductor material to induce a chemical reaction to remove the material, mainly through the chemical reaction between the fluorine-containing gas and the polysilicon, including , and For the participation of the process, all sensor data involved in the process are collected as process data, including parameters such as gas type, gas flow, pressure and power during etching, and then the key dimensions measured after etching are collected as measurement data. For example, in the CMP process, all sensor data in the process are collected as process data, including parameters such as chamber ID, wafer ID, chamber pressure, wafer rotation rate, slurry flow, processing stage, and then the average material removal rate of the process is collected as measurement data.

[0047] S2, takes the process data and measurement data of the key semiconductor manufacturing steps as the initial training sample set, iteratively trains the regression enhanced diffusion model (REDM model for short), and updates the training sample set after each round of iterative training.

[0048] like Figure 2 As shown, the REDM model includes four parts, namely, diffusion noise adding module, return noise removing module, regression enhancement module and sample screening module.

[0049] (1) Diffusion noise module:

[0050] In the diffusion noise adding process, the diffusion noise adding module gradually adds noise to the process data matrix of the training sample. , from the data distribution Starting from the predefined Gaussian distribution The process data matrix will gradually become blurred when the sampling noise is in the middle. When t is large enough, the process data matrix will become a pure Gaussian noise data matrix. This process is defined as:

[0051]

[0052] in, represents all process data matrix sequences from time 1 to T, Represents a given initial state The probability distribution of the matrix sequence, represents the time step, represents the total number of time steps, represents the process data matrix at time step t, Represents the process data matrix at time step t-1.

[0053] The noisy process data matrix is ​​gradually generated through the diffusion process .

[0054] (2) Backhaul denoising module:

[0055] The backpropagation denoising process uses the random noise data matrix as the latent variable Perform step-by-step denoising to predict the noise component added in each step of the diffusion process and gradually generate a noise feature matrix as virtual process data . Noise distribution is usually unknown, which is represented by the parameter The output of this neural network is , used to predict the virtual process data matrix The "real" noise component The loss function of the backhaul denoising process can be described as:

[0056]

[0057] in, represents the loss between the predicted noise and the actual added noise in the backhaul denoising process, Represents the initial data ,noise , time step expectations, represents the real noise component, Represents the virtual process data according to step t The predicted noise component.

[0058] In the present invention, the feedback denoising module is regarded as a soft sensor technology data generation model. The trained feedback denoising module can be used to generate process data of virtual semiconductor key manufacturing steps, that is, virtual samples are used as soft sensor technology data for subsequent training of soft sensor models.

[0059] (3) Regression enhancement module:

[0060] Traditional generative models perform poorly in capturing the complex mapping relationship between process variables and target variables in key semiconductor manufacturing steps, resulting in the inability of the generated virtual samples to accurately reflect changes in actual production. The present invention modifies the loss function of the generative model, hoping that the generated samples are more conducive to the soft sensor model.

[0061] In the present invention, the regression enhancement module is regarded as a soft sensor technology data discrimination model, which receives the virtual process data matrix generated by the feedback denoising module and the real process data matrix , generate virtual measurement data.

[0062] In a specific implementation of the present invention, the regression enhancement module includes a feature encoder, a pre-trained multi-head attention module, and a linear prediction layer, wherein the encoder can be a transformer encoder, and the linear prediction layer is a multi-layer perceptron. The real process data and the virtual process data are respectively encoded by the encoder to obtain their feature embedding representations in the feature space, and then the feature embedding representations are sent to the pre-trained multi-head attention module to extract features. The ability of the model to capture complex mapping relationships is enhanced by introducing a multi-layer head attention mechanism, and finally the virtual measurement data is generated by the linear prediction layer. and By comparing virtual measurement data with real measurement data Compare to calculate the loss value:

[0063]

[0064]

[0065] in, It is virtual measurement data generated based on the prediction of real process data. With real measurement data The loss between It is virtual measurement data generated based on virtual process data prediction. With real measurement data The loss between It is the real process data and the expectation of the real measurement data y, Is virtual process data and the expectation of the actual measurement data y.

[0066] In order to prevent limiting performance and reducing stability, the present invention adds a gradient penalty to enforce the Lipschitz constraint. The loss function is calculated as follows:

[0067]

[0068] Among them, the gradient of the discriminant network is expressed as , represents the probability distribution between virtual process data and real process data, Indicates that from this distribution Sampling expectations.

[0069] The loss function of the backhaul denoising module and the regression enhancement module is finally obtained as:

[0070]

[0071]

[0072] in, and are the total losses of the backhaul denoising module and the regression enhancement module, respectively. is the loss between the predicted noise and the actual added noise in the backhaul denoising process, It is the loss between the virtual measurement data generated based on the prediction of the real process data and the real measurement data. It is the loss between the virtual measurement data generated by the virtual process data prediction and the real measurement data. is the Lipschitz constraint loss, is the loss coefficient.

[0073] (4) Sample screening module:

[0074] In practical scenarios, even with a well-trained generative model, ensuring the quality of all virtual samples is challenging. Virtual samples that deviate from the original sample distribution may harm the performance of subsequent soft sensor models, and samples that are too similar are not good for enhancing the prediction performance of soft sensor models. Maintaining the diversity of the samples themselves is also a key aspect. After each round of generative training is completed, the generated virtual samples with the same number of training set samples are divided into subsets to form a virtual sample set And stored in the historical sample library, where Indicates A virtual sample subset, represents the training set, is a virtual sample subset Here, the number of samples in each virtual sample subset may be the same or different, and this embodiment takes the same as an example.

[0075] Calculate the Sliced-Wasserstein distance between each virtual sample subset and the training set:

[0076]

[0077] in, represents the Sliced-Wasserstein distance, express The unit sphere in dimensional space, is a direction vector chosen uniformly at random from the unit sphere, and Represent virtual sample subsets and training set The original multidimensional probability distribution of represents a linear projection, represents the inferred measure, Represents the distribution μ through mapping Push to a new space. Similarly, Indicates that the distribution By mapping Push to a new space. represents the one-dimensional Wasserstein distance, Express From the unit sphere Expected sampling.

[0078] Since it is not feasible to perform projection on all directions on the sphere, this embodiment uses 50 random projections to approximately calculate the SWD.

[0079] Calculate the cosine similarity within each virtual sample subset:

[0080]

[0081] in, Respectively represent The j-th and k-th virtual samples in the virtual sample subsets.

[0082] Based on the SWD and SIM scores of each virtual sample subset, a non-dominated sorting (NDS) is performed in the historical sample library to identify the Pareto frontier of each level and obtain a Pareto frontier sequence, where each virtual sample subset corresponds to a Pareto frontier level. For each subset, its crowding distance in the target space is calculated to measure the sparsity of the subset in the target space. The larger the crowding distance, the more sparse the subset is.

[0083] When screening virtual sample subsets, virtual sample subsets are selected from the first level of the Pareto frontier. If the number of virtual sample subsets at the current level does not meet the screening number requirement, the selection continues from the next level. For virtual sample subsets at the same Pareto frontier level, subsets with larger crowding distances are selected first, thereby enhancing the diversity of virtual sample sets until the screening number requirement is met. In each training iteration, the selected virtual sample subset is added to the training set to form a synthetic sample.

[0084] (3) Using random Gaussian noise as input, the feedback denoising module of the trained regression enhanced diffusion model is used to generate virtual process data, and the generated results are used as soft sensor technology data for key semiconductor manufacturing steps.

[0085] In order to quantitatively evaluate the performance of the REDM model, this paper compares it with existing models. Due to the large number of generative models, only the leading methods in each generative model paradigm are selected for evaluation, including TVAE, CTGAN, SMOTE and TabDDPM as baseline models.

[0086] REDM model parameter setting: The number of iterations is set to 30 generations, and each virtual sample generated is divided into 10 parts for sample selection. Figure 3 The evolution of the selected subset during the iteration process can be observed. Red represents the currently selected subset, yellow represents the historically selected subset, and blue represents the unselected subset. It can be seen that as the iteration proceeds, the SWD and SIM scores of the virtual sample subset gradually evolve to the lower left corner. At the same time, the red samples are dispersed in the Pareto frontier of the final virtual sample subset, indicating that as the iteration proceeds, the generated virtual samples are constantly learning the data distribution of the real samples, and the distribution distance between the two is gradually shrinking. At the same time, the final selected virtual sample subset is more diverse and is not just a copy of the real sample.

[0087] Machine learning utility evaluation: Using machine learning utility as the main evaluation indicator, quantify the performance of the regression model trained on synthetic data on the real test set. Three evaluation protocols, random forest, Xgboost and Catboost, are used for calculation. This example collects data sets of CVD, ETCH, and CMP processes, named CVD-1, CVD-2, CVD-3, CVD-4, ETCH-1, CMP-1, CMP-2, and CMP-3 respectively.

[0088] Generated virtual sample quality assessment: Use the kernel density estimation algorithm to obtain the probability density curve of the virtual sample to effectively compare the probability density curves of the virtual sample and the real sample. Figure 4 We can observe the target variable distribution of the data generated by REDM, CTGAN, TVAE and SMOTE and the real data. The first 8 columns represent the distribution difference between the target variables of the generated samples and the real samples of a single dataset under the training of the four generative models, and the last column represents the average distribution distance WD (Wasserstein Distance) between the generated samples and the real sample target variables of the generative model in the eight datasets. It can be seen that SMOTE has the smallest WD and the highest distribution similarity. This is because SMOTE is based on the principle of interpolation, which is equivalent to the copy result of the real data, while REDM's WD is smaller than CTGAN and TVAE, which shows that the samples generated by REDM have higher similarity than CTGAN and TVAE, and higher diversity than SMOTE.

[0089] Comparison of generative models: This experiment divides REDM into two variants: REDM-I and REDM-II. REDM-I does not include the sample screening module, while REDM-II includes the full regression enhanced diffusion model. By sampling synthetic datasets for each generative model, the R² score of the regression model is evaluated using the real test set.

[0090] Tables 1, 2, and 3 show the machine learning utility values ​​under different evaluation protocols. The results are generated by averaging five random seeds.

[0091] Table 1

[0092]

[0093] Table 2

[0094]

[0095] Table 3

[0096]

[0097] In each evaluation protocol, REDM-I outperforms TVAE, CTGAN, SMOTE, and TabDDPM on most datasets, indicating that incorporating the mapping relationship between features and target variables into the generative model can improve the machine learning utility of synthetic samples. REDM-II further improves this improvement through sample screening. As a lightweight generation method, SMOTE is competitive in data similarity, but its generated samples often "copy" real data, making it difficult to further improve machine learning utility. In contrast, the data generated by REDM has certain noise and transformation, which helps the regression model to be more robust to noise and transformation, resulting in better performance in real tests. In the baseline regression model test, CatBoost performed best. In actual semiconductor manufacturing scenarios, we usually only consider this best regression model. REDM-II improves the R² score of the model by 4.53% on average.

[0098] The present invention also provides a soft sensor technology data generation system based on a diffusion model, which is used to implement the above-mentioned embodiments. The terms "module", "unit", etc. used below can implement a combination of software and / or hardware for a predetermined function. Although the system described in the following embodiments is preferably implemented in software, it is also possible to implement hardware, or a combination of software and hardware.

[0099] This embodiment provides a soft sensor technology data generation system based on a diffusion model, including:

[0100] A wafer manufacturing process data collection module is used to collect data in the wafer manufacturing process, including process data and measurement data of key semiconductor manufacturing steps, and use the collected process data and measurement data as a training set;

[0101] A regression enhanced diffusion model module, which includes a diffusion denoising module, a backhaul denoising module and a regression enhancement module based on a multi-head attention mechanism. The diffusion denoising module gradually adds random noise to the process data of the training set to obtain pure Gaussian noise data. The backhaul denoising module gradually denoises and restores the pure Gaussian noise data to generate virtual process data. The regression enhancement module generates virtual measurement data based on the real process data and the virtual process data respectively.

[0102] A regression enhanced diffusion model training module is used to iteratively train the regression enhanced diffusion model using a training set, and update the parameters of the regression enhanced diffusion model according to the loss of virtual measurement data and real measurement data, the return denoising loss, and the Lipschitz constraint loss; and in the process of iteratively training the regression enhanced diffusion model, the training set is updated after each round of training, and the updating method is: the virtual process data generated by the return denoising module is divided into virtual sample subsets, and some virtual sample subsets are selected according to the non-dominated sorting results and congestion results of the virtual sample subsets and added to the training set;

[0103] The soft sensor technology data generation module is used to generate virtual process data using random Gaussian noise as input and the return denoising module of the trained regression enhanced diffusion model, and the generated results are used as soft sensor technology data for key semiconductor manufacturing steps.

[0104] As for the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can be referred to the partial description of the method embodiment. For example, the regression enhancement module includes:

[0105] An encoder module for encoding real process data and virtual process data;

[0106] A pre-trained multi-head attention module, which is used to extract features of the encoded real process data and virtual process data;

[0107] The linear prediction layer module is used to predict the virtual measurement data according to the characteristics of the encoded real process data and the virtual process data.

[0108] The implementation methods of the remaining modules will not be described in detail here. The system embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of the present invention. Those of ordinary skill in the art may understand and implement it without creative work.

[0109] The embodiments of the system of the present invention can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a device or apparatus such as a computer. The system embodiments can be implemented by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, the corresponding computer program instructions in the non-volatile memory are read into the memory by the processor of any device with data processing capabilities and run.

[0110] An embodiment of the present invention further provides an electronic device, including a memory and a processor;

[0111] The memory is used to store computer programs;

[0112] The processor is used to implement the above-mentioned method for generating data of soft sensor technology based on diffusion model when executing the computer program.

[0113] An embodiment of the present invention further provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the above-mentioned method for generating data of soft sensor technology based on a diffusion model is implemented.

[0114] The computer-readable storage medium may be an internal storage unit of any device with data processing capability described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium may also be an external storage device of any device with data processing capability, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capability. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capability, and may also be used to temporarily store data that has been output or is to be output.

[0115] Obviously, the above-described embodiments and drawings are only some examples of the present application. For those of ordinary skill in the art, the present application can also be applied to other similar situations based on these drawings without creative work. In addition, it is understandable that, although the work done in this development process may be complicated and lengthy, for those of ordinary skill in the art, certain changes in design, manufacturing or production based on the technical content disclosed in this application are only conventional technical means and should not be regarded as insufficient content disclosed in this application. Without departing from the concept of the present application, several variations and improvements can also be made, which all belong to the scope of protection of the present application. Therefore, the scope of protection of the present application shall be subject to the attached claims.

Claims

1. A method for generating data of soft sensor technology based on diffusion model, characterized in that: The following steps are involved: (1) Collect data from the wafer manufacturing process, including process data and measurement data of key semiconductor manufacturing steps, and use the collected process data and measurement data as training sets; (2) Iteratively train the regression enhanced diffusion model using the training set; The regression enhancement diffusion model includes a diffusion denoising module, a return denoising module and a regression enhancement module based on a multi-head attention mechanism. The diffusion denoising module gradually adds random noise to the process data of the training set to obtain pure Gaussian noise data, and the return denoising module gradually denoises and restores the pure Gaussian noise data to generate virtual process data. The regression enhancement module generates virtual measurement data according to the real process data and virtual process data, and updates the parameters of the regression enhancement diffusion model according to the loss of the virtual measurement data and the real measurement data, the return denoising loss and the Lipschitz constraint loss; In the process of iterative training of the regression enhanced diffusion model, the training set is updated after each round of training. The updating method is as follows: the virtual process data generated by the feedback denoising module is divided into virtual sample subsets, and some virtual sample subsets are selected according to the non-dominated sorting results and crowding results of the virtual sample subsets and added to the training set; (3) Using random Gaussian noise as input, the feedback denoising module of the trained regression enhanced diffusion model is used to generate virtual process data, and the generated results are used as soft sensor technology data for key semiconductor manufacturing steps.

2. The method for generating data of soft sensor technology based on diffusion model according to claim 1, characterized in that: The key semiconductor manufacturing steps include chemical vapor deposition step, etching step and chemical mechanical polishing step.

3. The method for generating data of soft sensor technology based on diffusion model according to claim 1, characterized in that: The process data of the key semiconductor manufacturing steps are a series of sensor values.

4. The method for generating data of soft sensor technology based on diffusion model according to claim 1, characterized in that: The regression enhancement module includes an encoder, a pre-trained multi-head attention module, and a linear prediction layer; the real process data and the virtual process data are encoded by the encoder respectively, and then the features are extracted by the pre-trained multi-head attention module, and finally the virtual measurement data is generated by the linear prediction layer.

5. The method for generating data of soft sensor technology based on diffusion model according to claim 1, characterized in that: During the iterative training of the regression enhanced diffusion model, the loss functions of the backhaul denoising module and the regression enhancement module are: ； ； in, and are the total losses of the backhaul denoising module and the regression enhancement module, respectively. is the loss between the predicted noise and the actual added noise in the backhaul denoising process, It is the loss between the virtual measurement data generated based on the prediction of the real process data and the real measurement data. It is the loss between the virtual measurement data generated by the virtual process data prediction and the real measurement data. is the Lipschitz constraint loss, is the loss coefficient.

6. The method for generating data of soft sensor technology based on diffusion model according to claim 1, characterized in that: The method of selecting a portion of the virtual sample subset according to the non-dominated sorting result and the crowding result of the virtual sample subset includes: Calculate the Sliced-Wasserstein distance between each virtual sample subset and the training set; Calculate the sample similarity score within each virtual sample subset; Based on the Sliced-Wasserstein distance and sample similarity score of each virtual sample subset, the virtual sample subsets are non-dominated sorted to obtain a Pareto front sequence, where each virtual sample subset corresponds to a Pareto front level; Calculate the crowding distance of each virtual sample subset in the Pareto front sequence; Starting from the first level of the Pareto frontier, virtual sample subsets are selected. If the number of virtual sample subsets at the current level does not meet the screening quantity requirement, the selection continues from the next level. For virtual sample subsets at the same Pareto frontier level, subsets with larger crowding distances are selected first until the screening quantity requirement is met.

7. A soft sensor technology data generation system based on a diffusion model, characterized in that: include: A wafer manufacturing process data collection module is used to collect data in the wafer manufacturing process, including process data and measurement data of key semiconductor manufacturing steps, and use the collected process data and measurement data as a training set; A regression enhanced diffusion model module, which includes a diffusion denoising module, a backhaul denoising module and a regression enhancement module based on a multi-head attention mechanism. The diffusion denoising module gradually adds random noise to the process data of the training set to obtain pure Gaussian noise data, and the backhaul denoising module gradually denoises and restores the pure Gaussian noise data to generate virtual process data. The regression enhancement module generates virtual measurement data based on real process data and virtual process data respectively; A regression enhanced diffusion model training module is used to iteratively train the regression enhanced diffusion model using a training set, and update the parameters of the regression enhanced diffusion model according to the loss of virtual measurement data and real measurement data, the return denoising loss, and the Lipschitz constraint loss; and in the process of iteratively training the regression enhanced diffusion model, the training set is updated after each round of training, and the updating method is: the virtual process data generated by the return denoising module is divided into virtual sample subsets, and some virtual sample subsets are selected according to the non-dominated sorting results and congestion results of the virtual sample subsets and added to the training set; The soft sensor technology data generation module is used to generate virtual process data using random Gaussian noise as input and the return denoising module of the trained regression enhanced diffusion model, and the generated results are used as soft sensor technology data for key semiconductor manufacturing steps.

8. The soft sensor technology data generation system based on diffusion model according to claim 7 is characterized in that: The regression enhancement module includes: An encoder module for encoding real process data and virtual process data; A pre-trained multi-head attention module, which is used to extract features of the encoded real process data and virtual process data; The linear prediction layer module is used to predict the virtual measurement data according to the characteristics of the encoded real process data and the virtual process data.

9. An electronic device, characterized in that: It comprises a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement the method for generating data of soft sensor technology based on diffusion model according to any one of claims 1 to 6.

10. A machine-readable storage medium, characterized in that: The machine-readable storage medium stores machine-executable instructions, which, when called and executed by a processor, are used to implement the method for generating data using soft sensor technology based on a diffusion model as described in any one of claims 1 to 6.

Images