Method for training photolithography imaging quality prediction model and method for optimizing photolithography process

Abstract

The application provides a training method of a photolithography imaging quality prediction model and an optimization method of a photolithography process, and belongs to the technical field of photolithography manufacturing. The training method comprises the following steps: acquiring a plurality of mask defect parameter sets, wherein the mask defect parameter set comprises a plurality of mask defect parameters; performing three-dimensional Gaussian modeling based on the mask defect parameter set to obtain an extreme ultraviolet photolithography reflective mask model; performing strict electromagnetic field simulation on the extreme ultraviolet photolithography reflective mask model to obtain a photolithography imaging parameter set, wherein the photolithography imaging parameter set comprises a plurality of photolithography imaging parameters; obtaining a plurality of training sample pairs based on the plurality of mask defect parameter sets and the plurality of photolithography imaging parameter sets corresponding to the plurality of mask defect parameter sets respectively; taking the mask defect parameter set of the training sample pair as an input sample, taking the photolithography imaging parameter set of the training sample pair as a label, and training an initial network model based on the plurality of training sample pairs to obtain a photolithography imaging quality prediction model.

Description

Technical Field

[0001] This application relates to the field of photolithography manufacturing technology, and more specifically, to a training method for a photolithography imaging quality prediction model and a method for optimizing photolithography processes. Background Technology

[0002] Extreme ultraviolet lithography (EUVL) uses reflective masks, and the imaging quality directly affects the process window and chip yield. However, existing masks inevitably contain amplitude defects located near the top layer of the multilayer film and phase defects distributed within the multilayer film. These mask defects greatly affect the lithography imaging quality.

[0003] Currently, the assessment of the impact of mask defects on lithography imaging quality mainly relies on single-factor experiments or empirical rules. Therefore, there is an urgent need for an algorithmic characterization method that can systematically analyze the influence of multi-factor mask defects on lithography imaging quality, enabling quantitative prediction and sensitivity analysis of key imaging indicators, and providing scientific guidance for mask defect control, process window optimization, and advanced node chip manufacturing. Summary of the Invention

[0004] In view of this, this application provides a training method for a lithography imaging quality prediction model and an optimization method for lithography processes.

[0005] One aspect of this application provides a training method for a lithography imaging quality prediction model, comprising: acquiring multiple mask defect parameter sets, wherein the mask defect parameter sets include multiple mask defect parameters; performing three-dimensional Gaussian modeling based on the mask defect parameter sets to obtain an extreme ultraviolet (EUV) lithography reflective mask model; performing rigorous electromagnetic field simulation on the EUV lithography reflective mask model to obtain a lithography imaging parameter set, wherein the lithography imaging parameter set includes multiple lithography imaging parameters; obtaining multiple training sample pairs based on the multiple mask defect parameter sets and their respective corresponding multiple lithography imaging parameter sets; using the mask defect parameter sets of the training sample pairs as input samples and the lithography imaging parameter sets of the training sample pairs as labels, training an initial network model based on the multiple training sample pairs to obtain a lithography imaging quality prediction model.

[0006] According to an embodiment of this application, the plurality of mask defect parameters include defect surface height, defect bottom height, defect surface full width at half maximum (FWHM), and defect bottom full width at half maximum (FWHM). The defect surface height is less than the defect bottom height, and the defect surface full width at half maximum (FWHM) is greater than the defect bottom full width at half maximum (FWHM), so that the extreme ultraviolet lithography reflective mask model obtained based on the set of mask defect parameters is a non-standard three-dimensional Gaussian geometry that is wider at the top and narrower at the bottom.

[0007] According to an embodiment of this application, the training method further includes: constructing multiple composite derived features using multiple mask defect parameters of the mask defect parameter set; constructing an enhanced mask defect parameter set corresponding to the mask defect parameter set using the multiple mask defect parameters and the multiple composite derived features; and training the initial network model using the enhanced mask defect parameter set as input samples and the lithography imaging parameter set corresponding to the mask defect parameter set as labels to obtain an enhanced lithography imaging quality prediction model.

[0008] According to embodiments of this application, the plurality of mask defect parameters include a defect termination film layer number, a defect initiation film layer number, a defect surface half-peak full width, a defect bottom half-peak full width, a defect surface height, a defect bottom height, and a defect center lateral offset. The plurality of composite derived features include a defect longitudinal depth feature, a defect morphology taper feature, and a positional coupling feature. The construction of the plurality of composite derived features using the plurality of mask defect parameters in the set of mask defect parameters includes: subtracting the defect termination film layer number from the defect initiation film layer number to obtain a defect longitudinal depth feature; subtracting the defect surface half-peak full width from the defect bottom half-peak full width to obtain a first taper value; subtracting the defect surface height from the defect bottom height to obtain a second taper value; dividing the first taper value from the second taper value to obtain a defect morphology taper feature; and dividing the defect center lateral offset from the preset single-period line empty pattern half-width to obtain a positional coupling feature.

[0009] Another aspect of this application provides a method for optimizing a photolithography process, comprising: determining multiple influence weight values ​​corresponding to multiple mask defect parameters in a mask defect parameter set through scrambling importance analysis based on a photolithography imaging quality prediction model, wherein the photolithography imaging quality prediction model is obtained according to the training method described above; and determining an optimization strategy for the photolithography process based on the multiple influence weight values.

[0010] According to an embodiment of this application, the above-mentioned determination of multiple influence weight values ​​corresponding to multiple mask defect parameters in a mask defect parameter set through scrambling importance analysis based on the lithography imaging quality prediction model includes: for each target defect parameter among the multiple mask defect parameters, randomly generating multiple variable defect values ​​based on a first preset sampling step size and a first preset value range of the target defect parameter; determining a random mask defect parameter set based on the variable defect values ​​and multiple preset benchmark defect values ​​corresponding to multiple other defect parameters, wherein the other defect parameters are mask defect parameters other than the target defect parameter among the multiple mask defect parameters; determining multiple random mask defect parameter sets corresponding to each of the multiple variable defect values ​​based on the multiple variable defect values; and obtaining the influence weight value of the target defect parameter based on the multiple random mask defect parameter sets.

[0011] According to an embodiment of this application, obtaining the influence weight value of the target defect parameter based on multiple sets of random mask defect parameters includes: inputting the multiple sets of random mask defect parameters into the lithography imaging quality prediction model to obtain multiple sets of random lithography imaging parameters; performing weighted calculations on the multiple random lithography imaging parameter values ​​in the multiple sets of random lithography imaging parameters to determine the amplitude corresponding to the multiple sets of random lithography imaging parameters; and determining the influence weight value of the target defect parameter based on the degree of dispersion between the multiple amplitudes corresponding to each of the multiple sets of random lithography imaging parameters.

[0012] According to an embodiment of this application, the optimization method further includes: constructing multiple test sample pairs; inputting the mask defect parameter set of the test sample pairs into the lithography imaging quality prediction model to obtain a predicted lithography imaging parameter set; determining a prediction error value based on the difference between the lithography imaging performance parameter set corresponding to the mask defect parameter set in the test sample pairs and the predicted lithography imaging performance parameter set; when the total prediction error value determined by the multiple prediction error values ​​is greater than a first preset threshold, using a genetic algorithm to determine multiple target value ranges corresponding to the multiple mask defect parameters in the mask defect parameter set; and determining an optimization strategy for the lithography process based on the multiple target value ranges.

[0013] According to an embodiment of this application, the method of determining the target value ranges corresponding to the mask defect parameters of the mask defect parameter set using a genetic algorithm includes: dividing the mask defect parameters into multiple important defect parameters and multiple minor defect parameters based on the multiple influence weight values ​​corresponding to the multiple mask defect parameters; randomly generating multiple target defect values ​​based on a first preset sampling step size and a first preset value range of the important defect parameters; randomly generating multiple minor defect values ​​based on a second preset sampling step size and a second preset value range of the minor defect parameters; determining an initial mask defect parameter set based on the multiple target defect values ​​corresponding to the multiple important defect parameters and the multiple minor defect values ​​corresponding to the multiple minor defect parameters; determining multiple initial mask defect parameter sets based on the multiple target defect values ​​corresponding to the important defect parameters; and determining the multiple target value ranges corresponding to the multiple mask defect parameters using a genetic algorithm based on the multiple initial mask defect parameter sets.

[0014] According to an embodiment of this application, the above-mentioned determination of multiple target value ranges corresponding to multiple mask defect parameters based on multiple initial mask defect parameter sets using a genetic algorithm includes: using multiple initial mask defect parameter sets as an initial population, multiple first preset value ranges corresponding to multiple important defect parameters as constraints of the search boundary, the above-mentioned lithography imaging quality prediction model as a fitness function, and using a genetic algorithm to iteratively optimize the above initial population based on preset fitness threshold conditions to determine a target population, wherein the target population includes multiple target mask defect parameter sets; determining multiple optimized value ranges corresponding to multiple important defect parameters based on multiple target mask defect parameter sets; and determining multiple target value ranges corresponding to multiple mask defect parameters based on multiple optimized value ranges corresponding to multiple important defect parameters and multiple second preset value ranges corresponding to multiple minor defect parameters.

[0015] According to the embodiments of this application, an extreme ultraviolet (EUV) lithography reflective mask model is obtained by constructing a set of mask defect parameters and performing three-dimensional Gaussian modeling. A rigorous electromagnetic field simulation is then performed on the EUV lithography reflective mask model to obtain a set of lithography imaging parameters. This allows for the construction of multiple training sample pairs, which are then used to train the initial network model to obtain a lithography imaging quality prediction model. This establishes a mapping relationship between multi-dimensional mask defect parameters and lithography imaging performance indicators, enabling a comprehensive assessment of the impact of complex mask defects and overcoming the limitations of existing single-factor or empirical rules. Attached Figure Description

[0016] The above and other objects, features and advantages of this application will become clearer from the following description of embodiments of this application with reference to the accompanying drawings.

[0017] Figure 1 A flowchart illustrating a training method for a lithographic imaging quality prediction model according to an embodiment of this application is shown.

[0018] Figure 2 A schematic diagram of the structure of an extreme ultraviolet lithography reflective mask model according to an embodiment of this application is shown.

[0019] Figure 3 A flowchart illustrating an optimization method for a photolithography process according to an embodiment of this application is shown.

[0020] Figure 4 A schematic diagram illustrating the effect of multiple mask defect parameters on offset according to an embodiment of this application is shown.

[0021] Figure 5 A schematic diagram illustrating the influence of multiple mask defect parameters on the logarithmic slope of a normalized image according to embodiments of this application is shown.

[0022] Figure 6 A schematic diagram illustrating the influence of multiple mask defect parameters on critical dimensions according to embodiments of this application is shown.

[0023] Figure 7 A schematic diagram showing the prediction accuracy of the offset by the lithographic imaging quality prediction model according to an embodiment of this application is illustrated.

[0024] Figure 8 A schematic diagram showing the prediction accuracy of the lithographic imaging quality prediction model for the normalized image logarithmic slope according to an embodiment of this application is illustrated.

[0025] Figure 9 A schematic diagram showing the prediction accuracy of the lithographic imaging quality prediction model for critical dimensions according to an embodiment of this application is illustrated. Detailed Implementation

[0026] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0027] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms "comprising," "including," etc., as used herein indicate the presence of the above-described features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0028] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0029] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0030] With the continuous advancement of nanoscale manufacturing processes, EUVL, as a mainstream advanced lithography solution, plays a decisive role in process window and chip yield due to the quality of its mask imaging. However, it is difficult to achieve "zero defects" in masks in existing lithography technologies. Phase-type defects in multilayer films can cause critical size deviations, image quality degradation, and local pattern distortion during imaging, thereby directly limiting the process window and reducing chip yield. Existing technologies rely heavily on single-factor experiments or empirical rules when assessing the impact of defects, considering only the influence of a single geometric parameter of the defect, such as size, depth, or location, on imaging, lacking quantitative analysis of the combined effects of multiple factors. This makes it impossible to systematically understand the degree of influence of different types, locations, and burial depths of defects on imaging indicators, and also makes it impossible to predict the exposeability of defects and their limitations on the lithography process window. Furthermore, existing technologies lack a clear quantitative model for the imaging effect of multiple defect superpositions, resulting in a lack of scientific basis for mask design and process optimization, which can easily lead to overly conservative design or local failures.

[0031] Therefore, it is necessary to address the problems of unsystematic mask defect characterization, insufficient sensitivity analysis, and lack of quantitative prediction in existing technical solutions.

[0032] Figure 1 A flowchart illustrating a training method for a lithographic imaging quality prediction model according to an embodiment of this application is shown.

[0033] like Figure 1 As shown, the training method for the lithography imaging quality prediction model includes operations S110~S150.

[0034] In operation S110, multiple mask defect parameter sets are obtained, wherein the mask defect parameter sets include multiple mask defect parameters.

[0035] Sample values ​​for multiple mask defect parameters can be generated using a multi-factor random sampling method to ensure balanced coverage of each input dimension within its defined range, while also guaranteeing the uniformity of the high-dimensional combination space. These multiple mask defect parameters can be used to characterize features such as the lateral size, vertical height, burial depth, and position of the mask defect along the graphic direction.

[0036] In operation S120, a three-dimensional Gaussian model is performed based on the mask defect parameter set to obtain an extreme ultraviolet lithography reflective mask model.

[0037] Extreme ultraviolet (EUV) lithography reflective masks are multilayer film structures, typically composed of alternating depositions of high-refractive-index and low-refractive-index materials, such as Mo / Si or Mo / Be multilayer films. These multilayer films may contain amplitude-type or phase-type defects on their surface and internally. For example, a multilayer film structure can be composed of 40 alternating Mo layers and 40 Si layers, totaling 80 single-layer films. The geometric morphology of mask defects includes hemispherical, Gaussian, anti-Gaussian, or other definable three-dimensional structures. This application uses a Gaussian shape to characterize mask defects. Its lateral dimensions, longitudinal height, embedment depth, and position in the pattern direction can be set according to the actual error range during mask manufacturing.

[0038] During operation of S130, a rigorous electromagnetic field simulation was performed on the extreme ultraviolet lithography reflective mask model to obtain a set of lithography imaging parameters, which includes multiple lithography imaging parameters.

[0039] Simulation of extreme ultraviolet (EUV) lithography reflective masks can be performed using rigorous lithography simulation software. The simulation light source is an EUV source with a wavelength of 13.5 nm, employing ring illumination with a central incident angle of 6°, and an internal coherence factor σ. in The external coherence factor σ is 0.6. out The polarization value is 0.9, and the polarization state is y-direction polarization. The target mask test pattern used is a 16nm / 16nm line-space pattern. The simulation process comprehensively considers interlayer reflection, phase characteristics, optical transfer function, optical distortion, and astigmatism effects of multilayer films. The simulation calculates diffraction, interference, and local phase perturbation at mask defects by solving Maxwell's equations, and finally outputs the spatial distribution results of the latent image on the wafer surface, completing the simulation calculation for a single sample.

[0040] The set of lithography imaging parameters can specifically include the spatial image distribution of the latent image on the wafer surface and the corresponding key output indicators, used to characterize imaging performance. For example, multiple lithography imaging parameters may include the critical dimension (CD), the normalized image log-slope (NILS), and the offset. These multiple lithography imaging parameters can be automatically extracted using rigorous simulation software to quantitatively reflect the impact of mask defects on the imaging plane, ensuring the accuracy and repeatability of the data.

[0041] In operation S140, multiple training sample pairs are obtained based on multiple sets of mask defect parameters and their corresponding multiple sets of lithographic imaging parameters.

[0042] Furthermore, a database can be established based on multiple training samples to link mask defects with imaging results, storing the mask defect parameter set with the corresponding lithographic imaging parameter set. The database structure supports multi-factor queries, statistical analysis, and data standardization. It can also perform data cleaning, outlier removal, and normalization according to the needs of subsequent model training, ensuring high quality and consistency of input and output data, and providing a reliable data foundation for subsequent algorithm-based multi-factor regression modeling.

[0043] In operation S150, the mask defect parameter set of the training sample pair is used as the input sample, and the lithography imaging parameter set of the training sample pair is used as the label. The initial network model is trained based on multiple training sample pairs to obtain the lithography imaging quality prediction model.

[0044] In the embodiments of this application, the Least Squares Boosting Regression Tree (LSBoost) algorithm can be used to model the nonlinear relationship between the mask defect parameter set and the lithography imaging parameter set. LSBoost generates weak regression trees step by step through iterative least squares residuals and combines them to form a strong regression prediction model, which can effectively capture the complex nonlinear effects of multiple mask defect parameters on multiple lithography imaging parameters. The lithography imaging quality prediction model can be used to guide mask manufacturing defect control, process window optimization, and chip imaging performance improvement.

[0045] Through the embodiments of this application, an extreme ultraviolet (EUV) lithography reflective mask model is obtained by constructing a set of mask defect parameters and performing three-dimensional Gaussian modeling. A rigorous electromagnetic field simulation is then performed on the EUV lithography reflective mask model to obtain a set of lithography imaging parameters. This allows the construction of multiple training sample pairs, which are then used to train the initial network model to obtain a lithography imaging quality prediction model. A mapping relationship between multi-dimensional mask defect parameters and lithography imaging performance indicators is established, enabling a comprehensive assessment of the impact of complex mask defects and overcoming the limitations of existing single-factor or empirical rules.

[0046] To conform to the smooth characteristics of mask defects, multiple mask defect parameters are used, including the defect surface height htop, the defect bottom height hbot, the full width at half maximum (Wtop) of the defect surface, and the full width at half maximum (Wbot) of the defect bottom. The defect surface height htop is less than the defect bottom height hbot, and the full width at half maximum (Wtop) of the defect surface is greater than the full width at half maximum (Wbot) of the defect bottom. This ensures that the extreme ultraviolet lithography reflective mask model obtained based on the set of mask defect parameters is a non-standard three-dimensional Gaussian geometry that is wider at the top and narrower at the bottom.

[0047] Figure 2 A schematic diagram of the structure of an extreme ultraviolet lithography reflective mask model according to an embodiment of this application is shown.

[0048] like Figure 2 As shown, the mask defects of the extreme ultraviolet lithography reflective mask model include the defect surface height htop, the defect bottom height hbot, the full width at half maximum (Wt) of the defect surface wtop, and the full width at half maximum (Wbot) of the defect bottom, which are non-standard three-dimensional Gaussian geometries that are wider at the top and narrower at the bottom.

[0049] Furthermore, necessary derived features can be extracted from multiple mask defect parameters to improve the ability of the lithography imaging quality prediction model to capture the influence of defects. For example, the defect size can be combined with the mask layer number to form a composite feature, which can be used to reflect the nonlinear influence of the spatial distribution of defects on the imaging results.

[0050] In one specific embodiment of this application, multiple mask defect parameters include the defect termination film layer number, the defect initiation film layer number, the full width at half maximum (FWHM) of the defect surface, the full width at half maximum (FWHM) of the defect bottom, the defect surface height, the defect bottom height, and the lateral offset of the defect center. Multiple composite derived features include the defect longitudinal depth feature, the defect morphology taper feature, and the position coupling feature. The training method further includes: subtracting the defect termination film layer number from the defect initiation film layer number to obtain the defect longitudinal depth feature; subtracting the full width at half maximum (FWHM) of the defect surface from the full width at half maximum (FWHM) of the defect bottom to obtain the first taper value; and subtracting the defect surface height from the full width at half maximum (FWHM) of the defect bottom. Subtracting the height from the bottom of the defect yields the second taper value; dividing the first taper value by the second taper value yields the taper feature of the defect morphology; dividing the lateral offset of the defect center by the half-width of the preset single-cycle line empty pattern yields the position coupling feature; using multiple mask defect parameters and multiple composite derived features, an enhanced mask defect parameter set corresponding to the mask defect parameter set is constructed; using the enhanced mask defect parameter set as input samples and the lithography imaging parameter set corresponding to the mask defect parameter set as labels, the initial network model is trained to obtain the enhanced lithography imaging quality prediction model.

[0051] Multiple mask defect parameters can be implemented based on the following 7 physical parameters, and all mask defect parameters are set with value range and geometric constraints in combination with the actual manufacturing error of the extreme ultraviolet mask.

[0052] The defect surface height htop ranges from 0.5 to 4.5 nm, with a sampling step size of 0.5 nm.

[0053] The defect bottom height hbot ranges from 5 to 20 nm, with a sampling step size of 0.5 nm.

[0054] The full width at half maximum (WTOP) of the defect surface ranges from 10 to 40 nm, with a sampling step size of 1 nm.

[0055] The full width at half maximum (WBOT) at the bottom of the defect ranges from 10 to 30 nm, with a sampling step size of 1 nm.

[0056] The defect initiation film layer number is layl, with a value range of 0~79 and a sampling step size of 1.

[0057] The defect termination film layer number is layh, with a value range of 1 to 80 and a sampling step size of 1.

[0058] The lateral offset y of the defect center ranges from −16 to 16 nm, with a sampling step size of 1 nm.

[0059] The longitudinal depth characteristics of the defect are The defect characterizes the number of layers in the multilayer film.

[0060] The defect morphology taper characteristics are as follows Match the three-dimensional contour features of the Gaussian defect, which is wider at the top and narrower at the bottom.

[0061] The positional coupling feature is the division of the lateral offset y of the defect center by the half-width of the single-cycle line space pattern, where the half-width of the single-cycle line space pattern is 16 nm, which is used to associate the spatial positional coupling relationship of the 16 nm / 16 nm target line space pattern.

[0062] Figure 3 A flowchart illustrating an optimization method for a photolithography process according to an embodiment of this application is shown.

[0063] like Figure 3 As shown, the optimization method for the photolithography process includes operations S310~S320.

[0064] In operation S310, based on the lithography imaging quality prediction model, the scrambling importance analysis determines the multiple influence weight values ​​corresponding to each of the multiple mask defect parameters in the mask defect parameter set.

[0065] The lithography imaging quality prediction model is obtained based on the training method described above, which will not be repeated here.

[0066] Randomization importance analysis involves randomly shuffling the specific values ​​of a mask defect parameter and observing the changes in the prediction results of the lithography imaging quality prediction model to quantify the influence of each mask defect parameter on the final imaging quality, thereby obtaining their respective influence weight values.

[0067] In operating S320, an optimization strategy for the photolithography process is determined based on multiple influence weight values.

[0068] Based on multiple influencing weights, the optimization strategy for photolithography can be determined as follows: Mask defect parameters with influencing weights play a decisive role in image quality, so they should be prioritized for control or compensation when adjusting the process, repairing the mask, or setting exposure conditions. Mask defect parameters with low influencing weights can have their tolerance appropriately relaxed. This way, limited process debugging resources can be concentrated on the most critical issues to achieve efficient and targeted improvement of photolithography performance.

[0069] In the embodiments of this application, based on the lithography imaging quality prediction model, multiple influence weight values ​​corresponding to multiple mask defect parameters in the mask defect parameter set are determined by scrambling importance analysis. Specifically, this may include the following operations: For each target defect parameter among the multiple mask defect parameters, multiple variable defect values ​​are randomly generated based on a first preset sampling step size and a first preset value range of the target defect parameter; Based on the variable defect values ​​and multiple preset benchmark defect values ​​corresponding to multiple other defect parameters, a random mask defect parameter set is determined, wherein the other defect parameters are mask defect parameters other than the target defect parameter among the multiple mask defect parameters; Based on the multiple variable defect values, multiple random mask defect parameter sets corresponding to each of the multiple variable defect values ​​are determined; The multiple random mask defect parameter sets are respectively input into the lithography imaging quality prediction model to obtain multiple random lithography imaging parameter sets; The multiple random lithography imaging parameter values ​​in the random lithography imaging parameter sets are weighted and calculated to determine the amplitude corresponding to the random lithography imaging parameter set; Based on the dispersion between the multiple amplitudes corresponding to each of the multiple random lithography imaging parameter sets, the influence weight value of the target defect parameter is determined.

[0070] The logic of Permutation Importance (PI) analysis in processing the lithography imaging quality prediction model is as follows: While keeping all mask defect parameters except the target defect parameters constant, each target defect parameter is randomly scrambled sequentially. The lithography imaging quality prediction model then predicts the magnitude of change in the corresponding set of lithography imaging parameters to quantify the sensitivity of the lithography imaging parameter set to imaging performance. Furthermore, the scrambling operation can be repeated multiple times and averaged to obtain the comprehensive influence weight of multiple mask defect parameters on lithography imaging performance indicators, thereby identifying the key structural parameters that have the most significant impact on imaging.

[0071] For example, multiple mask defect parameters include the defect termination layer number, the defect initiation layer number, the full width at half maximum (FWHM) of the defect surface, and the full WHM of the defect bottom. The defect initiation layer number, the FWHM of the defect surface, and the full WHM of the defect bottom are all set to fixed values, while the defect termination layer number takes 20 specific values ​​within a preset range, resulting in 20 sets of mask defect parameters. In each set, the specific values ​​for the defect initiation layer number, the FWHM of the defect surface, and the full WHM of the defect bottom are fixed; only the value of the defect termination layer number varies. These 20 sets of mask defect parameters are then input into a lithography imaging quality prediction model, resulting in 20 sets of lithography imaging parameters. Each set corresponds to a quantized value. If the amplitude differences among these 20 quantized values ​​are very large, it indicates that the defect termination layer number has a very large influence weight on the mask defect parameter.

[0072] Figure 4 A schematic diagram illustrating the effect of multiple mask defect parameters on offset according to an embodiment of this application is shown. Figure 5 A schematic diagram illustrating the influence of multiple mask defect parameters on the logarithmic slope of a normalized image according to embodiments of this application is shown. Figure 6 A schematic diagram illustrating the influence of multiple mask defect parameters on critical dimensions according to embodiments of this application is shown.

[0073] pass Figure 4-6 The results show that the offset in the imaging results is mainly affected by the lateral offset y of the defect center; NILS is mainly affected by the lateral offset y of the defect center and the defect surface height htop; CD is mainly affected by the lateral offset y of the defect center and the defect surface height htop. Similarly, the above methods can also be applied to analyze and model other lithographic mask buried defects.

[0074] Furthermore, the lithography imaging quality prediction model obtained by the training method according to the embodiments of this application can be tested. If the lithography imaging quality prediction model does not meet the testing requirements, a genetic algorithm can be further used to optimize the value range of multiple mask defect parameters. The specific method includes: constructing multiple test sample pairs; inputting the mask defect parameter set of the test sample pairs into the lithography imaging quality prediction model to obtain a predicted lithography imaging parameter set; determining the prediction error value based on the difference between the lithography imaging performance parameter set corresponding to the mask defect parameter set in the test sample pair and the predicted lithography imaging performance parameter set; when the total prediction error value determined by the multiple prediction error values ​​is greater than a first preset threshold, dividing the multiple mask defect parameters into multiple important defect parameters and multiple minor defect parameters based on the multiple influence weight values ​​corresponding to each of the multiple mask defect parameters; randomly generating multiple target defect values ​​based on the first preset sampling step size and the first preset value range of the important defect parameters; and randomly generating multiple minor defects based on the second preset sampling step size and the second preset value range of the minor defect parameters. The process involves: determining an initial set of mask defect parameters based on multiple target defect values ​​corresponding to multiple important defect parameters and multiple secondary defect values ​​corresponding to multiple minor defect parameters; determining multiple initial sets of mask defect parameters based on multiple target defect values ​​corresponding to important defect parameters; using these initial sets of mask defect parameters as an initial population, multiple first preset value ranges corresponding to multiple important defect parameters as constraints on the search boundary, and a lithography imaging quality prediction model as a fitness function; iteratively optimizing the initial population using a genetic algorithm based on preset fitness threshold conditions to determine a target population, where the target population includes multiple sets of target mask defect parameters; determining multiple optimized value ranges corresponding to multiple important defect parameters based on these multiple sets of target mask defect parameters; determining multiple target value ranges corresponding to multiple mask defect parameters based on the multiple optimized value ranges corresponding to multiple important defect parameters and multiple second preset value ranges corresponding to multiple minor defect parameters; and determining an optimization strategy for the lithography process based on these multiple target value ranges.

[0075] In the embodiments of this application, test sample pairs can be obtained in the same way as training sample pairs. Alternatively, multiple sample pairs can be determined and divided proportionally into a training set corresponding to the training sample pairs and a test set corresponding to the test sample pairs. This ensures that the training data is used for model fitting, the test data is used for model performance verification, and the risk of overfitting is avoided. The division ratio can be selected according to the number of samples and modeling requirements, commonly 70-80% for the training set and 20-30% for the test set.

[0076] The specific result of the total prediction error is used to evaluate the model's performance, and can be expressed as the mean absolute error (MAE). ), root mean square error ( ) and coefficient of determination ( The lithography imaging quality prediction model is quantified using indicators such as (n, n is the number of test sample pairs), thereby obtaining the fitting effect and prediction accuracy of the model on the test set. This is used to determine whether the model meets the requirements for engineering applications. This represents the quantized value (corresponding to the true value) of the set of lithographic imaging performance parameters for the i-th test sample pair. This represents the quantized value (corresponding to the predicted value) of the set of predicted lithographic imaging performance parameters for the i-th test sample pair. This is the mean of the true values.

[0077] Table 1 shows the prediction accuracy results of the lithography imaging quality prediction model in the embodiments of this application. It is worth noting that the lithography imaging quality prediction model has an R² of over 0.95 for CD, NILS and offset, with better prediction for offset, reaching 0.9831.

[0078] Table 1: Prediction accuracy results of the lithography imaging quality prediction model.

[0079]

[0080] Figure 7 A schematic diagram showing the prediction accuracy of the offset by the lithographic imaging quality prediction model according to an embodiment of this application is illustrated. Figure 8 A schematic diagram showing the prediction accuracy of the lithographic imaging quality prediction model for the normalized image logarithmic slope according to an embodiment of this application is illustrated. Figure 9 A schematic diagram showing the prediction accuracy of the lithographic imaging quality prediction model for critical dimensions according to an embodiment of this application is illustrated.

[0081] like Figure 7-9 As shown, the predicted values ​​of the visible light lithography imaging quality prediction model are basically distributed around the true values, and the number of results without prediction bias is the largest, as can be seen from the residual distribution and count distribution of the results.

[0082] If the accuracy of the lithography imaging quality prediction model does not meet the preset requirements, or if it is necessary to optimize imaging performance and broaden the process window, a genetic algorithm is initiated to optimize parameters. The genetic algorithm iteratively generates parameter combinations through initial population initialization, selection, crossover, mutation, and fitness evaluation. It uses the trained model to predict corresponding performance indicators and calculates the fitness function, achieving the optimal search for key defect parameter combinations. Iteration continues until the performance indicators are maximized or the convergence condition is met, obtaining multiple optimized value ranges for each of the important defect parameters. This determines the optimizable mask defect control range and process parameters, guiding mask defect management, lithography process window optimization, and improving chip yield.

[0083] Furthermore, the final trained lithography imaging quality prediction model can be visualized, including comparison of predicted and true values, feature importance analysis, and residual distribution plots, to intuitively demonstrate the fitting effect of the lithography imaging quality prediction model and the contribution of each input mask defect parameter to the output index. Through the above steps, a high-precision prediction model of mask defect parameters and output imaging performance index is established, and the generalization ability of the lithography imaging quality prediction model is ensured through cross-validation and optimization.

[0084] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

[0085] The embodiments of this application have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of this application. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Without departing from the scope of this application, those skilled in the art can make various substitutions and modifications, all of which should fall within the scope of this application.

Claims

1. A training method for a lithographic imaging quality prediction model, comprising: Obtain multiple sets of mask defect parameters, wherein the sets of mask defect parameters include multiple mask defect parameters; Three-dimensional Gaussian modeling is performed based on the set of mask defect parameters to obtain an extreme ultraviolet lithography reflective mask model. Rigorous electromagnetic field simulation was performed on the extreme ultraviolet lithography reflective mask model to obtain a set of lithography imaging parameters, wherein the set of lithography imaging parameters includes multiple lithography imaging parameters; Based on the multiple sets of mask defect parameters and their corresponding multiple sets of photolithography imaging parameters, multiple training sample pairs are obtained; Using the set of mask defect parameters of the training sample pairs as input samples and the set of lithography imaging parameters of the training sample pairs as labels, the initial network model is trained based on multiple training sample pairs to obtain a lithography imaging quality prediction model.

2. The training method according to claim 1, characterized in that, The mask defect parameters include defect surface height, defect bottom height, defect surface full width at half maximum (FWHM), and defect bottom full width at half maximum (FWHM). The defect surface height is less than the defect bottom height, and the defect surface full width at half maximum (FWHM) is greater than the defect bottom full width at half maximum (FWHM), so that the extreme ultraviolet lithography reflective mask model obtained based on the set of mask defect parameters is a non-standard three-dimensional Gaussian geometry that is wider at the top and narrower at the bottom.

3. The training method according to claim 1, characterized in that, The training method also includes: Multiple composite derived features are constructed using multiple mask defect parameters from the set of mask defect parameters; By utilizing multiple mask defect parameters and multiple composite derived features, an enhanced mask defect parameter set corresponding to the mask defect parameter set is constructed; Using the enhanced mask defect parameter set as input samples and the lithography imaging parameter set corresponding to the mask defect parameter set as labels, the initial network model is trained to obtain the enhanced lithography imaging quality prediction model.

4. The training method according to claim 3, characterized in that, The multiple mask defect parameters include the defect termination film layer number, the defect initiation film layer number, the full width at half maximum (FWHM) of the defect surface, the full width at half maximum (FWHM) of the defect bottom, the defect surface height, the defect bottom height, and the lateral offset of the defect center. The multiple composite derived features include the defect longitudinal depth feature, the defect morphology taper feature, and the positional coupling feature. The construction of multiple composite derived features using the multiple mask defect parameters of the mask defect parameter set includes: Subtracting the defect termination film number from the defect initiation film number yields the longitudinal depth feature of the defect. The first taper value is obtained by subtracting the full width of the half-peak at the surface of the defect from the full width of the half-peak at the bottom of the defect. Subtracting the height of the defect surface from the height of the defect bottom yields the second taper value; Divide the first taper value by the second taper value to obtain the taper feature of the defect morphology; The positional coupling feature is obtained by dividing the lateral offset of the defect center by the preset half-width of the single-cycle line empty pattern.

5. An optimization method for photolithography process, characterized in that, include: Based on the lithography imaging quality prediction model, multiple influence weight values ​​corresponding to multiple mask defect parameters in the mask defect parameter set are determined by scrambling importance analysis, wherein the lithography imaging quality prediction model is obtained according to the training method as described in any one of claims 1-4; Based on the multiple influence weight values, an optimization strategy for the photolithography process is determined.

6. The optimization method according to claim 5, characterized in that, The lithography-based imaging quality prediction model determines multiple influence weight values ​​for each mask defect parameter in the mask defect parameter set through scrambling importance analysis, including: For each target defect parameter among the multiple mask defect parameters, multiple variable defect values ​​are randomly generated based on the first preset sampling step size and the first preset value range of the target defect parameter; Based on the changing defect value and multiple preset benchmark defect values ​​corresponding to multiple other defect parameters, a random mask defect parameter set is determined, wherein the other defect parameters are mask defect parameters other than the target defect parameter among the multiple mask defect parameters; Based on the multiple changing defect values, determine a set of multiple random mask defect parameters corresponding to each of the multiple changing defect values; Based on multiple sets of random mask defect parameters, the influence weight value of the target defect parameter is obtained.

7. The optimization method according to claim 6, characterized in that, The step of obtaining the influence weight value of the target defect parameter based on multiple sets of random mask defect parameters includes: Multiple sets of random mask defect parameters are input into the lithography imaging quality prediction model to obtain multiple sets of random lithography imaging parameters. A weighted calculation is performed on multiple random lithography imaging parameter values ​​in the random lithography imaging parameter set to determine the amplitude corresponding to the random lithography imaging parameter set. Based on the degree of dispersion among the multiple amplitude values ​​corresponding to each of the multiple sets of random lithography imaging parameters, the influence weight value of the target defect parameter is determined.

8. The optimization method according to claim 5, characterized in that, The optimization method further includes: Construct multiple test sample pairs; The set of mask defect parameters of the test sample pair is input into the lithography imaging quality prediction model to obtain the set of predicted lithography imaging parameters. The prediction error value is determined based on the difference between the set of lithographic imaging performance parameters corresponding to the set of mask defect parameters in the test sample pair and the set of predicted lithographic imaging parameters. When the total prediction error determined by multiple prediction error values ​​is greater than a first preset threshold, a genetic algorithm is used to determine the target value ranges corresponding to each of the multiple mask defect parameters in the mask defect parameter set. Based on the range of multiple target values, an optimization strategy for the photolithography process is determined.

9. The optimization method according to claim 8, characterized in that, The step of using a genetic algorithm to determine the target value ranges corresponding to each of the multiple mask defect parameters in the mask defect parameter set includes: Based on the multiple influence weight values ​​corresponding to each of the multiple mask defect parameters, the multiple mask defect parameters are divided into multiple important defect parameters and multiple minor defect parameters; Based on the first preset sampling step size and the first preset value range of the important defect parameters, multiple target defect values ​​are randomly generated. Based on the second preset sampling step size and the second preset value range of the minor defect parameters, multiple minor defect values ​​are randomly generated; Based on the multiple target defect values ​​corresponding to each of the multiple important defect parameters and the multiple secondary defect values ​​corresponding to each of the multiple secondary defect parameters, an initial mask defect parameter set is determined; Based on the multiple target defect values ​​corresponding to the important defect parameters, multiple initial mask defect parameter sets are determined; Based on the multiple initial mask defect parameter sets, a genetic algorithm is used to determine the multiple target value ranges corresponding to each of the multiple mask defect parameters.

10. The optimization method according to claim 9, characterized in that, The step of determining the target value ranges corresponding to each of the multiple initial mask defect parameter sets using a genetic algorithm includes: Using multiple initial mask defect parameter sets as the initial population, and multiple first preset value ranges corresponding to multiple important defect parameters as constraints of the search boundary, the lithography imaging quality prediction model as the fitness function, and based on the preset fitness threshold conditions, the genetic algorithm is used to iteratively optimize the initial population to determine the target population, wherein the target population includes multiple target mask defect parameter sets; Based on the multiple sets of target mask defect parameters, determine multiple optimized value ranges for each of the multiple important defect parameters; Based on the multiple optimized value ranges corresponding to the multiple important defect parameters and the multiple second preset value ranges corresponding to the multiple minor defect parameters, the multiple target value ranges corresponding to the multiple mask defect parameters are determined.

Images