Optical measurement model acquisition method, optical measurement method and device, and storage medium

Abstract

The application discloses an optical measurement model acquisition method, an optical measurement method and device, and a storage medium. The optical measurement model acquisition method comprises the following steps: acquiring a training data set comprising training data of samples obtained under different manufacturing schemes and corresponding data labels, wherein the training data comprises spectral data and manufacturing scheme prompt information of the samples, and the data labels are actual structure parameters of the samples; inputting each training data into a machine learning model to predict structure parameters of samples corresponding to the training data, to obtain corresponding predicted structure parameters; and training the machine learning model based on the predicted structure parameters and the corresponding data labels to obtain an optical measurement model. Therefore, the optical measurement model can learn the knowledge of different manufacturing schemes, and subsequently, only by providing corresponding manufacturing scheme prompt information, the model can call relevant knowledge and accurately perform optical measurement, so that only one unified optical measurement model can perform optical measurement for different manufacturing schemes.

Description

Technical Field

[0001] This invention relates to the field of optical measurement technology, specifically to methods for acquiring optical measurement models, optical measurement methods and equipment, and storage media. Background Technology

[0002] In the semiconductor manufacturing industry, to ensure the quality of produced semiconductor devices, it is necessary to measure them to verify whether they meet quality requirements. Measurements of structural parameters are often performed using optical measurement methods. The basic principle is to collect the reflected / scattered light from the sample under test, obtain the spectral data of the zero-order diffraction light, and analyze the measured spectral data using a physical simulation model of the sample to obtain the structural parameters of the sample.

[0003] In recent years, machine learning technology has been increasingly applied to the field of semiconductor virtual metrology. Some research and industrial practices have attempted to use data-driven models, such as machine learning models, to replace or enhance physical simulation models, learning to predict structural parameters from sample data. However, existing machine learning solutions typically train specialized models for single processes or product structures. Each process or structural type (e.g., different device types or different material stacks) often requires separate data collection and training of an independent model, lacking cross-process and cross-material versatility. This necessitates retraining a new model when processes change or new materials are introduced, increasing maintenance costs and data requirements. Summary of the Invention

[0004] This application provides a method for obtaining an optical measurement model, an optical measurement method and device, and a computer-readable storage medium, which mainly solves the technical problem that existing technologies for optical measurement using machine learning require training a separate machine learning model for each process or structure.

[0005] Firstly, a method for obtaining an optical measurement model is provided, including:

[0006] Obtain a training dataset, which includes training data and corresponding data labels of samples obtained under various different manufacturing schemes. The training data includes the spectral data of the samples and manufacturing scheme prompts. The data labels are the actual structural parameters of the samples.

[0007] Each training data is input into a machine learning model to predict the structural parameters of the sample corresponding to the training data, and the corresponding predicted structural parameters are obtained.

[0008] The machine learning model is trained based on the predicted structural parameters and the corresponding data labels to obtain an optical measurement model.

[0009] In some embodiments, the machine learning model is a pre-trained large model.

[0010] In some embodiments, training the machine learning model based on the predicted structural parameters and corresponding data labels includes:

[0011] Based on the predicted structural parameters and the corresponding data labels, the parameters of the machine learning model are updated using a parameter fine-tuning method.

[0012] In some embodiments, the machine learning model has a self-attention mechanism.

[0013] In some embodiments, the manufacturing scheme prompt information is input into the machine learning model in text form or in the form of embedded vectors, and / or the spectral data is input into the machine learning model in the form of a discretized numerical sequence.

[0014] In some embodiments, training the machine learning model based on the predicted structural parameters and corresponding data labels includes:

[0015] If the number of predicted structural parameters corresponding to the prompts from different manufacturing schemes is the same, then the machine learning model is trained based on the following objective function:

[0016]

[0017] Where N represents the number of training data points, and i represents the i-th training data point. This represents the predicted structure parameters output by the machine learning model for the i-th training data. This indicates the corresponding data label;

[0018] If the number of predicted structural parameters corresponding to the prompts from different manufacturing schemes is different, then the machine learning model is trained based on the following objective function:

[0019]

[0020] Where d represents the maximum value among the number of predicted structural parameters corresponding to each manufacturing scheme prompt information, and the machine learning model outputs d predicted structural parameters for all training data; This represents the j-th predicted structure parameter output by the machine learning model for the i-th training data. Indicates the corresponding data label, This represents the weight of the j-th predicted structural parameter output by the machine learning model for the i-th training data; This is the relevance mask for the j-th predicted structure parameter output by the machine learning model for the i-th training data. If this predicted structure parameter is unrelated to the manufacturing scheme hint information in the i-th training data, then... =0, if relevant. =1.

[0021] Secondly, an optical measurement method is provided, including:

[0022] Obtain spectral data and manufacturing process information from the sample being tested;

[0023] The spectral data and manufacturing scheme information of the sample under test are input into the optical measurement model obtained by the optical measurement model acquisition method of the first aspect described above, and the structural parameters of the sample under test are obtained.

[0024] In some embodiments, if the training dataset used to obtain the optical measurement model does not include the manufacturing scheme prompt information for the sample under test, the process further includes the following steps before obtaining the spectral data and manufacturing scheme prompt information of the sample under test:

[0025] Set a soft hint parameter;

[0026] Acquire the spectral data of the sample obtained under the manufacturing scheme used to manufacture the sample under test as supplementary spectral data, and obtain the corresponding supplementary data label;

[0027] The supplementary spectral data and the soft cue parameters are input into the optical measurement model to obtain the corresponding predicted structural parameters; the parameters of the optical measurement model are frozen, and the soft cue parameters are updated based on the predicted structural parameters and the corresponding supplementary data labels until a preset stopping condition is reached. The final soft cue parameters are used as the manufacturing scheme prompt information for the sample under test.

[0028] The optical measurement method in some embodiments further includes:

[0029] Acquire new training data and corresponding data labels to form an incremental training data pool;

[0030] The optical measurement model is periodically retrained using the incremental training data pool and the training dataset.

[0031] In some embodiments, the manufacturing scheme prompt information includes at least one of manufacturing process information, material information, and device structure information, and the structural parameters include at least one of linewidth, height, sidewall angle, and aperture.

[0032] Thirdly, an optical measurement device is provided, comprising:

[0033] A light source is used to emit a detection beam that illuminates the surface of the sample being tested.

[0034] A spectrometer is used to collect reflected and / or scattered light from the surface of the sample under test, and to measure the intensity of each wavelength in the reflected and / or scattered light to form spectral data.

[0035] A processor, the processor being configured to perform the method described in the first or second aspect above.

[0036] Fourthly, a computer-readable storage medium is provided, wherein a computer-executable program or instructions are stored therein, and the computer-executable program or instructions, when executed by a processor, are used to implement the method described in the first or second aspect above.

[0037] The optical measurement model acquisition method, optical measurement method and equipment, and computer-readable storage medium provided in this application use training data and corresponding data labels of samples obtained under various different manufacturing schemes to train a single machine learning model to obtain the optical measurement model. Manufacturing scheme prompts are introduced into the training data to provide contextual information of the manufacturing scheme to the machine learning model, enabling the optical measurement model to learn the measurement experience and knowledge of different manufacturing schemes. Subsequently, when using the optical measurement model to perform optical measurements on the sample under test, only the corresponding manufacturing scheme prompts need to be provided as guidance, and the optical measurement model can call the relevant knowledge to accurately perform the optical measurement task of the sample under test. Only a unified optical measurement model is needed to cover the optical measurement requirements of different manufacturing schemes, without the need to train a new model separately for each manufacturing scheme. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the overall flow of the technical solution of this application;

[0039] Figure 2 This is a flowchart of a method for obtaining an optical measurement model according to some embodiments;

[0040] Figure 3 These are flowcharts of optical measurement methods according to some embodiments;

[0041] Figure 4 This is a flowchart of some embodiments for further training of the optical measurement model for new manufacturing schemes;

[0042] Figure 5 These are schematic diagrams of the structure of optical measurement devices according to some embodiments. Detailed Implementation

[0043] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of this application. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to this application are not shown or described in the specification. This is to avoid obscuring the core parts of this application with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0044] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0045] The serial numbers assigned to components or physical quantities in this document, such as "first," "second," etc., are used only to distinguish the described objects and have no sequential or technical meaning. They should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. "Multiple" means two or more. Unless otherwise specified, "connection" or "linkage" in this application includes both direct and indirect connections (linkages).

[0046] The inventors explored and analyzed the existing technology and found that there are still some shortcomings in the existing technology, mainly as follows.

[0047] Poor model generalization and reusability: Traditional optical measurement models (whether physical simulation models or small machine learning models) are highly targeted but lack versatility. Each model is usually limited to a specific process node or structure, making it difficult to apply directly to new situations. Different manufacturing solutions (including different product lines, different manufacturing processes, different materials and / or device structures) require the maintenance of their own model libraries, making it impossible to achieve "one model for multiple uses".

[0048] Long adaptation cycle for new processes: When new manufacturing processes, materials or device structures are introduced, there is a lack of optical measurement models that can be quickly adapted; existing methods often require experts to rebuild the model based on new data, which is time-consuming and relies on a large number of labeled samples.

[0049] High overhead of repetitive training: The inventors believe that there are common physical laws and measurement characteristics among different manufacturing processes, materials and device structures, but existing methods have not recognized this and it is difficult to share this knowledge; each time a new model is built, it needs to be trained from scratch, which wastes the value of existing data and increases computational overhead; and for advanced processes, obtaining sufficient training data is expensive and time-consuming.

[0050] In summary, existing technologies lack generalizability to meet the increasingly diverse optical measurement needs in the semiconductor field, resulting in bottlenecks in development efficiency.

[0051] Therefore, this application proposes a solution based on manufacturing scheme prompts to effectively address the diverse needs and development efficiency bottlenecks in semiconductor optical measurement. The core technical problem to be solved is: how to build a unified machine learning model that can accurately perform optical measurement tasks under different manufacturing schemes.

[0052] To address the aforementioned technical problems, this application provides a solution by introducing manufacturing scheme prompts into semiconductor optical metrology. Specifically, it constructs a unified machine learning model and assigns specific prompts as contextual input for each manufacturing scheme, guiding the model to employ corresponding knowledge and reasoning logic during inference. Compared to existing solutions that train models separately for each manufacturing scheme, this application adds manufacturing scheme prompts to the model's input, thus achieving flexible adaptation to different manufacturing schemes without modifying model parameters. Optical metrology includes OCD (Optical Critical Dimension) measurement, among others.

[0053] Figure 1 A flowchart illustrating the overall process of semiconductor optical measurement according to the method of this application in one embodiment is shown. First, training data is used to obtain an optical measurement model with structural parameter prediction capabilities. Then, the optical measurement model is deployed to a practical application environment to perform optical measurements on the samples being measured during the production process. For example... Figure 1As shown, information such as manufacturing process, product category, and structure can be obtained from incoming material information to serve as manufacturing scheme prompts. Spectral data is also acquired and, after preprocessing, combined with the prompts to form training data. The data labels are the structural parameters of the corresponding samples. The training data is used for model training, with the prompts driving the regression task of structural parameters. The training process can employ Soft Prompt Tuning and the LoRA algorithm. After training, a usable optical measurement model is obtained and deployed in a real-world application environment. It receives spectral data from the tested sample and manufacturing scheme prompts, performs inference and prediction, and obtains the structural parameters of the tested sample. After training, the model can continue to learn and update using newly acquired spectral data and manufacturing scheme prompts from actual production.

[0054] This application provides a method for obtaining an optical measurement model, which can perform optical measurement tasks such as OCD. Please refer to... Figure 2 The method for obtaining the optical measurement model in some embodiments includes steps 110 to 130, which are described in detail below.

[0055] Step 110: Obtain the training dataset.

[0056] The training dataset includes training data and corresponding data labels for samples obtained under various manufacturing schemes. The training data includes the spectral data of the samples and manufacturing scheme prompts, while the data labels are the actual structural parameters of the samples.

[0057] Spectral data includes the spectral reflectance and / or scattering data of the sample, such as reflectance spectra or multi-angle scattering intensity values ​​within a certain wavelength range. The obtained raw spectral data can be preprocessed (such as denoising and normalization) before being embedded into the model input.

[0058] The manufacturing plan includes the manufacturing process, materials, and device structure specified for manufacturing semiconductor devices. Manufacturing plan information includes at least one of the following: manufacturing process information, material information, and device structure information. Manufacturing process information includes the process node name (e.g., "7nm FinFET process"), and material information includes the material stack-up order and material type (photoresist, etc.). , Device structure information includes device name (e.g., FinFET), pattern type (line / hole / fin structure, etc.) and target linewidth.

[0059] Structural parameters include at least one of linewidth, height, sidewall angle, and aperture. For example, for OCD measurements, structural parameters include the sample's critical dimensions (linewidth, height, etc.). Each manufacturing process and sample has its own set of structural parameters that need to be measured, and these parameters may differ for different manufacturing processes and samples.

[0060] Training data can be obtained through various means. In one embodiment, simulation tools (such as Rigorous Coupled Wave Analysis (RCWA)) can be used to generate simulated spectral data based on the manufacturing scheme, and the simulated spectral data can be used to construct a data-labeled simulation training dataset. In another embodiment, training data can be constructed using measured data from existing products. For example, spectral data of actual products can be collected from optical measurement devices such as OCD measurement devices, and the corresponding structural parameters such as linewidth / thickness can be obtained by calibration using precision offline measurement devices such as SEM (Scanning Electron Microscope) and AFM (Atomic Force Microscope). Preferably, simulation data and measured data can be merged to form a training dataset, thereby covering a wide range of manufacturing scheme variations.

[0061] Step 120: Input each training data into a machine learning model to predict the structural parameters of the sample corresponding to the training data, and obtain the corresponding predicted structural parameters.

[0062] The machine learning model can employ various existing machine learning models. Existing technologies typically use small-scale, dedicated models for optical measurements. The inventors believe that small-scale machine learning models may struggle to capture common knowledge across manufacturing processes, and the models may lack the ability to generalize to new situations. Therefore, improvements are needed to better realize the concept of a single model covering different manufacturing processes as proposed in this application.

[0063] Therefore, in some embodiments of this application, the machine learning model used is a pre-trained large model, such as DeepSeek or LLaMA. Large models possess powerful learning capabilities, capable of integrating measurement experience and knowledge from different manufacturing schemes, and mastering the common physical laws governing various manufacturing schemes. With the help of prompts from different manufacturing schemes, the model can invoke different knowledge to predict structural parameters. By introducing a combination of a large model and manufacturing scheme prompts into optical measurement, flexible adaptation of a single optical measurement model to different manufacturing schemes is achieved.

[0064] Step 130: Train the machine learning model based on the predicted structural parameters and the corresponding data labels to obtain the optical measurement model.

[0065] An objective function can be constructed based on the predicted structural parameters and the corresponding data labels. The machine learning model can then be trained based on this objective function, which represents the degree of difference between the predicted structural parameters and the corresponding data labels, enabling the machine learning model to learn to output predicted structural parameters that are closer to the data labels.

[0066] Since different manufacturing schemes require different structural parameters to be measured, the number of predicted structural parameters corresponding to the prompts for different manufacturing schemes in the training dataset may not be the same. In some embodiments, if the number of predicted structural parameters corresponding to the prompts for different manufacturing schemes is the same, the machine learning model is trained based on the following objective function:

[0067]

[0068] Where N represents the number of training data points, and i represents the i-th training data point. This represents the predicted structure parameters output by the machine learning model for the i-th training data. This indicates the corresponding data label.

[0069] If the number of predicted structural parameters corresponding to the prompts from different manufacturing schemes is different, then the machine learning model is trained based on the following objective function:

[0070]

[0071] Where d represents the maximum value among the number of predicted structural parameters corresponding to the prompt information of each manufacturing scheme, and the machine learning model outputs d predicted structural parameters for all training data. This represents the j-th predicted structure parameter output by the machine learning model for the i-th training data. Indicates the corresponding data label, This represents the weight of the j-th predicted structural parameter output by the machine learning model for the i-th training data. Since the machine learning model outputs d predicted structural parameters for all training data, however, not all manufacturing scheme prompts in the training data correspond to d predicted structural parameters. For some manufacturing schemes, some structural parameters do not need to be predicted, i.e., they are irrelevant to the prompt information for that manufacturing scheme. Therefore, this embodiment uses a correlation mask to mask irrelevant predicted structural parameters. This is the relevance mask of the j-th predicted structure parameter output by the machine learning model for the i-th training data. If this predicted structure parameter is unrelated to the manufacturing scheme hint information in the i-th training data, then... =0, if relevant. =1.

[0072] The training process primarily involves updating the parameters of the machine learning model using the backpropagation method. In some embodiments, the steps of training the machine learning model based on the predicted structural parameters and corresponding data labels include: updating the parameters of the machine learning model using a parameter fine-tuning method based on the predicted structural parameters and corresponding data labels. The parameter fine-tuning method can be the LoRA algorithm, the QLORA algorithm, etc.

[0073] This embodiment employs a parameter fine-tuning method to update the parameters of a machine learning model. This method updates only a small subset of parameters while keeping most of the model's parameters frozen, thus improving training efficiency. For example, the LoRA algorithm updates only the inserted low-rank matrix while keeping most of the model's parameters frozen, reducing the number of parameters updated. This is particularly important for large models with a large number of parameters and pre-existing general knowledge. Updating all parameters synchronously would be costly and inefficient. The parameter fine-tuning method reduces training overhead while preserving the original model's general representational capabilities to the greatest extent possible.

[0074] The machine learning models in some embodiments of this application have a self-attention mechanism, such as the large model. The self-attention mechanism of the machine learning model processes the manufacturing scheme prompt information together with the spectral data. The manufacturing scheme prompt information participates in the calculation of the attention weight in the self-attention mechanism, thereby modulating the representation of the spectral data in subsequent layers, enhancing the features related to the manufacturing scheme prompt information, and achieving the effect of "the same weight, different manufacturing scheme contexts follow different inference paths".

[0075] Since this application combines spectral data with manufacturing scheme prompts as input to the machine learning model, it involves multi-source heterogeneous data, including spectral data, process formula information, and metadata of structure and materials. It is difficult for machine learning models to process this data. Therefore, in some embodiments of this application, the input data has been modified to convert these unstructured or multimodal data into an input form acceptable to the machine learning model.

[0076] In some embodiments, manufacturing suggestion information is input into the machine learning model in text form or as an embedded vector.

[0077] Specifically, information such as manufacturing processes, materials, and device structures included in the manufacturing plan can be converted into structured text and used as input prompts for the machine learning model. For example, in one embodiment, the manufacturing plan prompts could be constructed as follows: "Process = 7nm; Device = FinFET; Stacked Structure = "On Si; target linewidth ≈ 20nm…" This can be summarized manually or automatically generated using templates, integrating key differences between various manufacturing schemes into the prompt text to provide intuitive context for the machine learning model. Alternatively, information such as manufacturing processes, materials, and device structures included in the manufacturing scheme can be converted into embedding vectors as input prompts for the machine learning model. These embedding vectors can be obtained through pre-training.

[0078] In some embodiments, spectral data is input into the machine learning model in the form of a discretized numerical sequence.

[0079] Typically, the collected spectral data is a continuous numerical sequence. Since some machine learning models cannot directly process continuous numerical sequences, this embodiment discretizes it and inputs it into the machine learning model as a discretized numerical sequence. One approach is to sample the original spectral data into a fixed-length sequence based on wavelength, and directly treat it as a numerical feature vector input into the machine learning model.

[0080] In addition, for optical property parameters of materials (such as the curve of refractive index as a function of wavelength), material databases can be referenced and appended to the manufacturing scheme prompt information in the form of vector codes to help machine learning models consider the influence of materials on the spectrum.

[0081] The final input data is a sequence of spectral data and manufacturing solution prompts, forming an input format acceptable to the machine learning model. The spectral data provides the measurement signal, while the manufacturing solution prompts provide the context of the manufacturing solution. For example, for a sample, the input data might be: Input = [Prompts: Process = 7nm; Structure = FinFET gate; Material = Si / ]+[Spectral data: 0.12, 0.35, 0.48, ...]; The output is the corresponding actual structural parameters, such as [linewidth = 19.8nm, height = 42.1nm].

[0082] The optical measurement model obtained by the above method has learned the mapping relationship between spectral data and structural parameters under different manufacturing schemes, as well as how to adjust the internal calculation according to the manufacturing scheme prompts. After training, the model parameters are fixed, and it is not necessary to train a separate model for each manufacturing scheme.

[0083] Based on the optical measurement model obtained by the above-described method, this application provides an optical measurement method. Please refer to [reference needed]. Figure 3 In some embodiments, the method includes steps 201 to 202, which are described in detail below.

[0084] Step 201: Obtain the spectral data and manufacturing scheme prompts of the sample being tested.

[0085] The sample under test can be a wafer, etc. The spectral data is obtained by actually measuring the sample. The manufacturing scheme prompt information can be prompt text constructed during training or an embedded vector. The manufacturing scheme prompt information provides the optical measurement model with the context of the manufacturing scheme used for the sample under test. The spectral data is concatenated with it or encoded at specific positions, and together with the manufacturing scheme prompt information, it serves as the input data for the optical measurement model.

[0086] Step 202: Input the spectral data of the sample to be tested and the manufacturing scheme prompt information into the optical measurement model obtained by the optical measurement model acquisition method of any embodiment of this application to obtain the structural parameters of the sample to be tested.

[0087] If an optical measurement model possesses a self-attention mechanism, the manufacturing scheme prompts are incorporated into the input. When calculating attention weights, the optical measurement model takes these prompts into account. This means the model's attention head focuses on the manufacturing scheme prompts, assigning higher attention weights to the information contained within them. This influences the reasoning direction of subsequent layers, causing the optical measurement model to focus on features more relevant to the structural parameters to be predicted. For example, large models can leverage their powerful contextual learning capabilities to dynamically adjust internal feature representations and reasoning logic in a single forward inference by focusing on the prompts in the input data, ensuring the output aligns with the current manufacturing scheme context.

[0088] For example, when the manufacturing specification describes the sample as a "DRAM deep hole" structure, the early layers of the optical measurement model may activate feature modes associated with high aspect ratio hole measurements, interpreting specific peak shapes in the spectrum and focusing more on energy attenuation characteristics in the long wavelength range to infer the aperture. However, when the manufacturing specification includes "FinFET_7nm," the optical measurement model tends to focus on interference peaks in the short wavelength range to predict the linewidth. The internal inference direction of the optical measurement model adjusts accordingly based on the different manufacturing specification specifications, but still shares the same set of parameters. This mechanism effectively enables the generation of measurement results specific to the manufacturing specification by combining different contexts in the forward computation of the optical measurement model.

[0089] To improve the performance of the optical measurement model and maintain its adaptability to new manufacturing schemes, some embodiments of this application introduce a continuous learning mechanism that allows the optical measurement model to be continuously updated and optimized over time.

[0090] Please refer to Figure 4In some embodiments, if the training dataset used to obtain the optical measurement model does not include manufacturing scheme hints for the sample under test, that is, the manufacturing scheme used for the sample under test is a new manufacturing scheme for the optical measurement model, then the following steps are included before obtaining the spectral data and manufacturing scheme hints for the sample under test:

[0091] Step 301: Set a soft cue parameter, such as a soft cue vector. The soft cue parameter is a trainable parameter that can be updated through training.

[0092] Step 302: Obtain the spectral data of the sample obtained under the manufacturing scheme used to manufacture the sample under test as supplementary spectral data, and obtain the corresponding supplementary data label; the supplementary spectral data and the corresponding supplementary data label can also be obtained through simulation or actual measurement;

[0093] Step 303: Input the supplementary spectral data and soft cue parameters into the optical measurement model to obtain the corresponding predicted structural parameters; freeze the parameters of the optical measurement model, update the soft cue parameters based on the predicted structural parameters and the corresponding supplementary data labels until the preset stopping condition is reached, and use the final soft cue parameters as the manufacturing scheme prompt information for the sample under test.

[0094] Updating the soft cue parameters can also be done by constructing an objective function based on the predicted structural parameters and corresponding supplementary data labels, and then using backpropagation. The constructed objective function aims to enable the optical measurement model to output more accurate predicted structural parameters under the soft cue parameters.

[0095] Similarly, in one embodiment of the optical measurement model acquisition method of this application, after obtaining the optical measurement model, if a new manufacturing scheme emerges, i.e., the training dataset used to acquire the optical measurement model does not include the manufacturing scheme prompt information of the new manufacturing scheme, a soft prompt parameter can be set for training to adapt to the new manufacturing scheme. That is, after obtaining the optical measurement model, the method can also include the following steps: setting a soft prompt parameter; acquiring the spectral data of the sample obtained under the new manufacturing scheme as supplementary spectral data, and acquiring the corresponding supplementary data labels; inputting the supplementary spectral data and the soft prompt parameter into the optical measurement model to obtain the corresponding predicted structure parameters; freezing the parameters of the optical measurement model, updating the soft prompt parameter based on the predicted structure parameters and the corresponding supplementary data labels until a preset stopping condition is reached, and using the final soft prompt parameter as the manufacturing scheme prompt information of the new manufacturing scheme.

[0096] The above embodiments use a soft cue parameter to represent new manufacturing scheme prompts. Updating and optimizing this parameter allows the model to effectively incorporate information from the new manufacturing scheme. This new manufacturing scheme prompt, along with the spectral data of the sample being tested, is input into the optical measurement model, providing it with the context of the new manufacturing scheme. This allows the optical measurement model to immediately gain better adaptability to the new manufacturing scheme. Since soft cue parameters typically have only tens or hundreds of parameters, updating them requires far fewer parameters and less memory than a full parameter update of the optical measurement model. This approach is particularly suitable for situations where data is scarce and rapid deployment is needed in the early stages of a new manufacturing scheme's implementation, enabling rapid adaptation to small datasets.

[0097] The methods for acquiring optical measurement models and optical measurement methods in some embodiments of this application further include: acquiring new training data and corresponding data labels to form an incremental training data pool; and periodically retraining the optical measurement model using the incremental training data pool and the training dataset.

[0098] Specifically, spectral data obtained from actual production and actual structural parameter measurement results (e.g., structural parameters measured by SEM during offline verification) can be continuously collected. This newly acquired spectral data, along with existing or newly constructed corresponding manufacturing scheme prompts, constitutes new training data. The actual structural parameter measurement results serve as corresponding data labels. This new training data and corresponding data labels are aggregated into an incremental training data pool. The incremental training data pool (new training data and corresponding data labels) is periodically fused with the aforementioned training dataset (past training data and corresponding data labels) to retrain the optical measurement model. For example, the incremental training data pool and the aforementioned training dataset can be fused quarterly, and the model can be retrained using efficient parameter fine-tuning methods such as the LoRA algorithm. Retraining can be performed offline.

[0099] By periodically retraining the optical measurement model with new training data and corresponding data labels, as well as with past training data and corresponding data labels, the accuracy of the optical measurement model in predicting structural parameters can be continuously improved.

[0100] This application also provides an optical measurement device, please refer to... Figure 5 The optical measurement device includes a light source 1, a spectrometer 2, and a processor 3.

[0101] The light source 1 is used to emit a detection beam that illuminates the surface of the sample 4 being tested. The detection beam can be visible light or ultraviolet light, etc.

[0102] The spectrometer 2 is used to collect reflected and / or scattered light from the surface of the sample 4 and measure the intensity of each wavelength in the reflected and / or scattered light to form spectral data.

[0103] The processor 3 is used to execute the method for acquiring an optical measurement model or the optical measurement method according to any embodiment of this application.

[0104] The method for obtaining the optical measurement model, the optical measurement method, and the device according to the above embodiments have the following advantages compared to the prior art.

[0105] First, by introducing manufacturing scheme prompts into the training data, the machine learning model is provided with contextual information about the manufacturing schemes. This allows the acquired optical measurement model to learn the measurement experience and knowledge of different manufacturing schemes. Subsequently, when using the optical measurement model to perform optical measurements on the sample, only the corresponding manufacturing scheme prompts need to be provided as guidance. The optical measurement model can then call upon the relevant knowledge to accurately perform the optical measurement task on the sample. Only a unified optical measurement model is needed to handle the optical measurement tasks of different manufacturing schemes, achieving "one model for multiple uses" without the need to train a new model separately for each manufacturing scheme. This greatly improves the reusability and deployment efficiency of the optical measurement model and reduces the model development and maintenance costs.

[0106] Secondly, by using manufacturing scheme prompts to guide model behavior, frequent adjustments to the model's internal parameters are avoided. When a new manufacturing scheme emerges, only a small number of new parameters are needed to formulate corresponding soft cue parameters, and these soft cue parameters are then optimized or slightly fine-tuned. These soft cue parameters can then be used as prompts for the new manufacturing scheme, enabling the optical measurement model to quickly adapt to the new manufacturing scheme. This significantly shortens the development cycle of measurement methods for new manufacturing schemes and reduces the requirements for data scale.

[0107] Furthermore, by integrating historical data and experience from multiple manufacturing schemes into a single model and activating relevant parts with prompts during inference, knowledge sharing among different manufacturing schemes is achieved. The model adapts to multiple manufacturing schemes while maintaining a grasp of general physical laws, and can be continuously extended to new manufacturing schemes through prompts or fine-tuning with a small amount of new data, solving the problems of model forgetting and transfer difficulties. Because the large model retains the general physical laws of each manufacturing scheme and can achieve the accuracy level of a dedicated model under the guidance of manufacturing scheme prompts, it balances accuracy and generalization.

[0108] Furthermore, some embodiments apply parameter fine-tuning methods to the field of optical measurement, significantly lowering the barrier to joint training of multiple manufacturing schemes. Combined with training strategies incorporating manufacturing scheme prompts, a generalizable optical measurement model was successfully trained. This training paradigm is not available in the development of traditional machine learning measurement models with fewer parameters.

[0109] In summary, the technical solution of this application improves the intelligence and flexibility of semiconductor optical measurement, which helps to shorten the R&D cycle, reduce data requirements, and improve the response speed to changes in manufacturing solutions.

[0110] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.

[0111] The above examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention.

Claims

1. A method for obtaining an optical measurement model, characterized in that, include: A training dataset is obtained, which includes training data and corresponding data labels of samples obtained under various different manufacturing schemes. The training data includes spectral data of the samples and manufacturing scheme prompt information. The manufacturing scheme prompt information includes at least one of manufacturing process information, material information and device structure information. The data labels are the actual structural parameters of the samples. Each training data is input into a machine learning model to predict the structural parameters of the sample corresponding to the training data, thereby obtaining the corresponding predicted structural parameters; wherein, the manufacturing scheme prompt information is input into the machine learning model in the form of text or embedded vector, and / or the spectral data is input into the machine learning model in the form of a discretized numerical sequence, and the structural parameters include at least one of linewidth, height, sidewall angle and aperture. The machine learning model is trained based on the predicted structural parameters and the corresponding data labels to obtain an optical measurement model.

2. The method for obtaining the optical measurement model as described in claim 1, characterized in that, The machine learning model is a pre-trained large model.

3. The method for obtaining the optical measurement model as described in claim 2, characterized in that, The step of training the machine learning model based on the predicted structural parameters and corresponding data labels includes: Based on the predicted structural parameters and the corresponding data labels, the parameters of the machine learning model are updated using a parameter fine-tuning method.

4. The method for obtaining the optical measurement model as described in claim 1, characterized in that, The machine learning model has a self-attention mechanism.

5. The method for obtaining the optical measurement model as described in claim 1, characterized in that, The step of training the machine learning model based on the predicted structural parameters and corresponding data labels includes: If the number of predicted structural parameters corresponding to the prompts from different manufacturing schemes is the same, then the machine learning model is trained based on the following objective function: Where N represents the number of training data points, and i represents the i-th training data point. This represents the predicted structure parameters output by the machine learning model for the i-th training data. This indicates the corresponding data label; If the number of predicted structural parameters corresponding to the prompts from different manufacturing schemes is different, then the machine learning model is trained based on the following objective function: Where d represents the maximum value among the number of predicted structural parameters corresponding to each manufacturing scheme prompt information, and the machine learning model outputs d predicted structural parameters for all training data; This represents the j-th predicted structure parameter output by the machine learning model for the i-th training data. Indicates the corresponding data label, This represents the weight of the j-th predicted structural parameter output by the machine learning model for the i-th training data; This is the relevance mask for the j-th predicted structure parameter output by the machine learning model for the i-th training data. If this predicted structure parameter is unrelated to the manufacturing scheme hint information in the i-th training data, then... =0, if relevant. =1.

6. An optical measurement method, characterized in that, include: Obtain spectral data and manufacturing process information from the sample being tested; The spectral data and manufacturing scheme prompts of the sample under test are input into the optical measurement model obtained by the optical measurement model acquisition method according to any one of claims 1 to 5 to obtain the structural parameters of the sample under test.

7. The optical measurement method as described in claim 6, characterized in that, If the training dataset used to obtain the optical measurement model does not include the manufacturing scheme prompt information for the sample under test, then the process of obtaining the spectral data and manufacturing scheme prompt information for the sample under test shall include: Set a soft hint parameter; Acquire the spectral data of the sample obtained under the manufacturing scheme used to manufacture the sample under test as supplementary spectral data, and obtain the corresponding supplementary data label; The supplementary spectral data and the soft cue parameters are input into the optical measurement model to obtain the corresponding predicted structural parameters; the parameters of the optical measurement model are frozen, and the soft cue parameters are updated based on the predicted structural parameters and the corresponding supplementary data labels until a preset stopping condition is reached. The final soft cue parameters are used as the manufacturing scheme prompt information for the sample under test.

8. The optical measurement method as described in claim 6, characterized in that, Also includes: Acquire new training data and corresponding data labels to form an incremental training data pool; The optical measurement model is periodically retrained using the incremental training data pool and the training dataset.

9. An optical measurement device, characterized in that, include: A light source is used to emit a detection beam that illuminates the surface of the sample being tested. A spectrometer is used to collect reflected and / or scattered light from the surface of the sample under test, and to measure the intensity of each wavelength in the reflected and / or scattered light to form spectral data. A processor for performing the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer-executable program or instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 8.

Images