Large-aperture diffraction structure dose-path joint optimization and splicing method

Abstract

The invention discloses a large-aperture diffraction structure dose-path joint optimization and splicing method, and relates to the technical field of diffraction structure manufacturing, and the method comprises the steps: carrying out the geometric analysis of a design layout of a large-aperture diffraction structure, obtaining the spatial layout data of a sub-field and an overlapping region, and extracting the key feature points related to the morphology control; taking the spatial layout data and the key feature points as input, training a machine learning model taking the splicing area morphology smoothness as an optimization target, and obtaining a morphology prediction model; taking the ergodic deduction result of the field-division scanning path sequence and the parameterized adjustment result of the exposure dose distribution of the overlapping area as joint optimization variables, substituting the joint optimization variables into the morphology prediction model, and carrying out iterative calculation until convergence to obtain an optimal path and dose distribution; and fusing the data to generate a joint control instruction set. According to the method, the morphology prediction accuracy and the splicing area smoothness can be improved, and the splicing efficiency and quality are optimized.

Description

Technical Field

[0001] This invention belongs to the field of diffraction structure manufacturing technology, specifically a method for dose-path joint optimization and splicing of large-aperture diffraction structures. Background Technology

[0002] Large-aperture diffraction structures are widely used in optical systems. During their manufacturing process, the overall structure needs to be segmented and spliced. In existing technologies, the splicing and manufacturing of large-aperture diffraction structures are usually completed by optimizing the segmented scanning path or adjusting the exposure dose of the overlapping area separately. At the same time, the morphology is predicted by conventional machine learning models or traditional modeling methods. In the modeling process, single or few input parameters are often used, without fully combining the spatial characteristics of the segmented and overlapping areas and key information related to the morphology.

[0003] Existing technical solutions have shortcomings in practical applications. When optimizing the scanning path alone, it can easily lead to abrupt changes in morphology in the overlapping area due to dose mismatch. When adjusting the exposure dose alone, it can cause uneven morphology within the field segment and the stitching area due to unreasonable path sequence. At the same time, conventional morphology prediction models, due to one-sided input parameters, cannot accurately capture the spatial correlation and morphology influencing factors between the field segment and the overlapping area, resulting in large prediction deviations, long iterative optimization cycles, and difficulty in meeting the requirements of large-aperture diffraction structures for stitching accuracy and morphology smoothness.

[0004] The core of the stitching quality of large-aperture diffraction structures depends on the smoothness of the stitching area morphology. How to achieve synergistic optimization of scanning path and exposure dose, and how to improve the accuracy of morphology prediction through more comprehensive input parameters, thereby solving the problems of uneven stitching area morphology, large prediction deviation, and low iteration efficiency, have become key challenges that urgently need to be solved in the current manufacturing process of large-aperture diffraction structures. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; To this end, the present invention proposes a dose-path joint optimization and splicing method for large-aperture diffraction structures, including: Geometric analysis was performed on the design layout of the large-aperture diffraction structure to generate spatial layout data of the field division and overlap regions. Key feature points related to morphology control are extracted from the design layout of the large-aperture diffraction structure. The spatial layout data of the field division and overlap area, as well as the key feature points, are used as input to train a machine learning model with the morphology smoothness of the splicing area as the optimization objective, and a trained morphology prediction model is obtained. Using the pre-trained morphology prediction model, the scanning path sequence within each sub-field data block is traversally deduced, and the exposure dose distribution within the overlapping scanning area is simultaneously parameterized. The deduced scanning path sequence within the sub-field and the parameterized overlapping area exposure dose distribution are used as joint optimization variables and substituted into the pre-trained morphology prediction model for iterative calculation until the morphology smoothness index output by the pre-trained morphology prediction model converges, thus obtaining the optimal scanning path sequence and optimal overlapping area dose distribution corresponding to each sub-field data block. The optimal scanning path sequence and optimal overlap region dose distribution corresponding to each field data block are fused to generate a complete set of dose-path joint control instructions for the large-aperture diffraction structure.

[0006] Furthermore, the geometric analysis of the design layout of the large-aperture diffraction structure to form spatial layout data of the field division and overlap regions includes: Geometric analysis is performed on the design layout of the large-aperture diffraction structure to identify non-periodic or highly complex regions. Based on the dimensions of the design layout, a seamless field segmentation operation is performed, dividing the entire design layout into several field data blocks with fixed spatial relationships. Simultaneously, an overlapping scanning area is planned for each field data block to eliminate splicing errors, forming spatial layout data for the fields and overlapping areas. Specifically, this includes: The edge contour of the design layout of the large-aperture diffraction structure is extracted to obtain the outer boundary data of the design layout, and the area and aspect ratio of the region enclosed by the outer boundary data are calculated. Based on the area and aspect ratio, and combined with the maximum single processing field of view of the direct writing device, the number of sub-field data blocks and the reference side length of each sub-field data block are determined, wherein the reference side length is less than the maximum single processing field of view. Using the external boundary data of the design layout as a reference, the entire design layout is divided into rectangular grids according to the determined number of field data blocks and the reference side length, generating an initial set of field data blocks. There are uncovered gaps between the field data blocks in the initial set of field data blocks. The initial set of field data blocks is subjected to edge expansion processing, extending the edge of each field data block outward by a preset overlap distance, so that adjacent field data blocks overlap. The overlapping part is the overlapping scanning area. At the same time, the part that exceeds the outer boundary data range of the design layout is trimmed to form the final spatial layout data of the field and the overlapping area.

[0007] Furthermore, key feature points related to morphology control are extracted from the design layout of the large-aperture diffraction structure. Using the spatial layout data of the field division and overlap regions, along with the key feature points, as input, a machine learning model is trained with the smoothness of the splicing region's morphology as the optimization objective. This yields a trained morphology prediction model, including: Gradient calculation is performed on the design layout of the large-aperture diffraction structure to identify edge regions in the design layout whose gray value change rate exceeds a set threshold, and the key feature points are obtained by uniformly sampling in the edge regions. The spatial layout data of the field and the overlapping area, the key feature points, and the pre-set standard topography reference surface are correlated to construct a multi-dimensional feature vector. The multi-dimensional feature vector includes the field position coordinates, the key feature point coordinates, and the normal vector of the standard topography reference surface. Using actual morphology measurement data accumulated during historical direct writing, the deviation between the actual morphology and the standard morphology reference surface under different scanning path sequences and different combinations of dose distribution in the overlapping area is marked, serving as the supervised learning label for the machine learning model. The Support Vector Regression algorithm is selected as the main structure of the machine learning model. The multidimensional feature vector is used as input and the supervised learning label is used as output. The Support Vector Regression algorithm is trained to obtain the trained shape prediction model that can output the predicted value of shape deviation. The smaller the absolute value of the predicted value of shape deviation, the higher the shape smoothness.

[0008] Furthermore, using the pre-trained morphology prediction model, the scanning path sequence within each sub-field data block is traversally deduced, and the exposure dose distribution within the overlapping scanning area is simultaneously parameterized. The deduced scanning path sequence within the sub-field and the parameterized overlapping area exposure dose distribution are used as joint optimization variables and substituted into the pre-trained morphology prediction model for iterative calculation until the morphology smoothness index output by the pre-trained morphology prediction model converges, thus obtaining the optimal scanning path sequence and optimal overlapping area dose distribution for each sub-field data block, including: Multiple candidate scanning path sequences are defined for the current field data block to be optimized. The candidate scanning path sequences include unidirectional scanning, loop scanning and spiral scanning, and each scanning path sequence corresponds to a path parameter. A dose gradient function is defined for the overlapping scanning area. The dose gradient function includes two control parameters, which are used to control the dose peak in the center region and the dose attenuation slope in the edge region of the overlapping area, respectively. By adjusting the two control parameters, the parameterized adjustment of the exposure dose distribution in the overlapping area is achieved. The path parameters of the candidate scanning path sequence corresponding to the current field data block, the two control parameters of the dose gradient function in the overlapping area, and the spatial layout data and key feature points of the field and the overlapping area are combined to form the current joint optimization variable group. The current set of joint optimization variables is input into the trained shape prediction model to calculate the predicted value of shape deviation under the current set of joint optimization variables. The predicted value of the topographic deviation is compared with the predicted value of the topographic deviation in the previous iteration. If the difference is greater than the preset convergence threshold, the parameters in the joint optimization variable group are adjusted by gradient descent to generate a new joint optimization variable group. The process of substituting and comparing is repeated until the change in the predicted value of the topographic deviation is less than the convergence threshold multiple times in a row. The corresponding scanning path sequence and overlapping area dose distribution at this time are taken as the optimal scanning path sequence and optimal overlapping area dose distribution.

[0009] Furthermore, a dose gradient function is defined for the overlapping scanning area. This function includes two control parameters, which are used to control the dose peak in the central region and the dose attenuation slope in the edge region of the overlapping area, respectively. By adjusting these two control parameters, parameterized adjustment of the exposure dose distribution in the overlapping area is achieved, including: A one-dimensional coordinate system is established with the central axis of the overlapping scanning area as the axis of symmetry. The origin of the one-dimensional coordinate system is located at the center of the overlapping scanning area, and the positive direction points to the interior of the field data block. The dose gradient function is defined as a quadratic function about the one-dimensional coordinate. The quadratic function is in the form of a parabola, with its vertex located at the origin of the one-dimensional coordinate system. The function value at the vertex is the peak dose, and the absolute value of the slope of the quadratic function on both sides of the vertex is the dose attenuation slope. By adjusting the value of the peak dose, the cumulative exposure at the center of the overlapping area is controlled. By adjusting the value of the dose attenuation slope, the rate of change of exposure during the transition from the center of the overlapping area to the interior of the subfield is controlled, thereby generating the overlapping area exposure dose distribution that meets the input requirements of the trained morphology prediction model.

[0010] Furthermore, the optimal scanning path order and optimal overlap region dose distribution corresponding to each field data block are fused to generate a complete set of dose-path joint control instructions for the large-aperture diffraction structure, including: Create a global process control data table, which is indexed by the number of the sub-field data block, and each row records all process parameters corresponding to a sub-field data block; The name of the optimal scanning path sequence, path parameters, two control parameters of the optimal overlap area dose distribution, and the coordinate offset of the field data block in the overall design layout are sequentially filled into the corresponding rows of the process control data table to form structured intermediate data. According to the order of the field data block numbers, each row of the process control data table is read sequentially, and the structured intermediate data of each row is translated into machine instructions that can be recognized by the direct writing equipment. The machine instructions include movement instructions, exposure start instructions, dose parameter setting instructions and exposure stop instructions. All machine instructions obtained from translation are arranged in the order of field data block numbers, and synchronous calibration instructions are inserted between machine instructions of adjacent field data blocks, and finally connected in series to form the complete set of dose-path joint control instructions. The dose-path joint control instruction set is used to guide the direct writing device to complete seamless direct writing of large-aperture diffraction structures.

[0011] Furthermore, the insertion of synchronization calibration instructions between machine instructions of adjacent field data blocks includes: Immediately after each stop exposure command indicating the end of a field data block, a raster position reset command is inserted to reset the raster system of the direct-write device to a known reference position. After the grating position reset command, a dose monitor reading command is inserted. The dose monitor reading command is used to read the current actual deposited dose value and compare the actual deposited dose value with the theoretical calculated value to obtain dose deviation data. If the dose deviation data exceeds the allowable range, a dose compensation coefficient adjustment instruction is inserted after the dose monitor reading instruction. The dose compensation coefficient adjustment instruction is used to correct the dose parameters used in subsequent field data blocks based on the dose deviation data, so as to ensure the dose accuracy of the entire dose-path joint control instruction set during continuous execution.

[0012] Furthermore, the geometric analysis of the design layout of the large-aperture diffraction structure to identify non-periodic or highly complex regions, and the execution of seamless field segmentation based on the dimensions of the design layout of the large-aperture diffraction structure, further includes: After identifying aperiodic or highly complex regions, the regions are locally magnified to analyze their local feature frequencies. Based on these local feature frequencies, the partitioning density of the field data blocks containing the aperiodic or highly complex regions is dynamically adjusted. This allows regions with dense features to be covered by finer-grained field data blocks, while regions with sparse features are covered by coarser-grained field data blocks. This reduces the total number of field data blocks while maintaining accuracy, thereby lowering the computational burden on the trained shape prediction model.

[0013] Furthermore, the step of using the trained morphology prediction model to perform a traversal deduction of the scanning path sequence within each sub-field data block also includes: During the simulation, time series constraints between fields are introduced. Based on the kinematic model of the direct-write device, the total time required to complete the previous field data block is estimated. Combined with the estimated time consumption of the scanning path sequence within the field data block, the scanning path sequence of the next field data block is screened. Scanning path sequences that have a predicted morphology deviation value less than a preset morphology threshold but cause frequent start-stop of the device are eliminated. From the perspective of motion stability, the range of values ​​of the joint optimization variables is further constrained so that the final optimal scanning path sequence takes into account both morphology quality and device motion efficiency.

[0014] Furthermore, the step of fusing the optimal scanning path order and optimal overlap region dose distribution corresponding to each field data block to generate a complete dose-path joint control instruction set for the large-aperture diffraction structure also includes: After generating the complete set of dose-path joint control instructions, a logical integrity check is performed on the complete set of dose-path joint control instructions to check for missing field data blocks, coordinate offset conflicts, or dose parameters exceeding the limits. If an anomaly is detected, the process is traced back to the spatial layout data generation step of the field and overlapping area, the field division parameters are readjusted, and the subsequent optimization and fusion process is re-executed until a logically complete and error-free complete set of dose-path joint control instructions is generated.

[0015] Compared with the prior art, the beneficial effects of the present invention are: Using the spatial layout data of the field division and overlapping areas obtained through geometric analysis of the design layout of a large-aperture diffraction structure, along with key feature points related to topography control extracted from the design layout, as input, a machine learning model is trained with the smoothness of the splicing area topography as the optimization objective. This forms a topography prediction model specifically for splicing large-aperture diffraction structures. Compared to conventional modeling methods that only use a single parameter input, this model can more comprehensively and accurately capture the spatial correlation of the field division and overlapping areas and the topography influencing factors, reducing deviations in the topography prediction process, making the prediction results more consistent with the actual manufacturing scenario, effectively avoiding splicing defects caused by inaccurate predictions, and reducing the difficulty of adjustments during the splicing process.

[0016] The scanning path sequence within each sub-field data block is traversally deduced, and the exposure dose distribution in the overlapping scanning area is simultaneously parameterized. The deduced scanning path sequence within the sub-field and the parameterized exposure dose distribution in the overlapping area are used as joint optimization variables and substituted into the trained morphology prediction model for iterative calculation until the morphology smoothness index converges, thus obtaining the optimal scanning path sequence and the optimal dose distribution in the overlapping area. Compared with conventional methods of optimizing the scanning path or adjusting the exposure dose separately, this approach achieves coordinated adaptation of path and dose, effectively mitigating morphological abrupt changes in the overlapping area caused by path disorder and dose unevenness, improving the overall smoothness of the stitching area, shortening the iterative optimization cycle, and improving the stitching efficiency and quality consistency of large-aperture diffraction structures. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of the large-aperture diffraction structure dose-path joint optimization and splicing method described in this invention; Figure 2 A flowchart generated for geometric analysis and field layout of the design layout; Figure 3 The flowchart shows the training process for the key feature point extraction and morphology prediction model. Figure 4 This is a simulation analysis diagram of the dose compensation coefficient; Figure 5 The image shows the training results of the shape prediction model. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1Geometric analysis was performed on the design layout of the large-aperture diffraction structure to generate spatial layout data for the subfields and overlapping regions. Key feature points related to morphology control were extracted from this design layout. Using the spatial layout data of the subfields and overlapping regions, along with these key feature points, as input, a machine learning model was trained with the smoothness of the stitched region morphology as the optimization objective, resulting in a pre-trained morphology prediction model. Using this pre-trained morphology prediction model, the scanning path sequence within each subfield data block was iteratively deduced, while the exposure dose distribution within the overlapping scanning region was simultaneously parameterized. The deduced scanning path sequence within each subfield and the parameterized overlapping region exposure dose distribution were used as joint optimization variables and substituted into the pre-trained morphology prediction model for iterative calculation until the morphology smoothness index output by the model converged, thus obtaining the optimal scanning path sequence and optimal overlapping region dose distribution for each subfield data block. The optimal scanning path sequence and optimal overlapping region dose distribution for each subfield data block were fused to generate a complete dose-path joint control command set for this large-aperture diffraction structure.

[0020] In one embodiment of the present invention, the spatial layout data of the field division and overlap region is generated using a design layout of a large-aperture blazed diffraction grating for a spectrometer as an example. The overall design layout of the large-aperture blazed diffraction grating is rectangular, and its working area contains periodic parallel scribe lines. However, in the positioning and alignment mark area near the edge, the scribe line structure transforms into a non-periodic complex curve pattern. (See also...) Figure 2 Geometric analysis was performed on the design layout of a large-aperture blazed diffraction grating. By analyzing the grayscale distribution or vector contours of the design layout data, non-periodic or highly complex regions containing positioning alignment marks were identified. Edge contour extraction was performed on the design layout of the large-aperture blazed diffraction grating to obtain the outer boundary data of a rectangle defined by the coordinates of four vertices. The area of ​​this rectangular region was calculated to be 120 mm x 100 mm, or 12,000 square millimeters, with an aspect ratio of 1.2. The maximum single-pass processing field of view of the direct-write device is 25 mm x 25 mm. Based on the area and aspect ratio of the design layout, and ensuring that the size of each sub-field data block does not exceed the maximum single-pass processing field of view, the entire design layout was divided into 35 sub-field data blocks, with the reference side length of each sub-field data block set at 20 mm. Based on the rectangular outer boundary data, and with a reference side length of 20 mm, the 120 mm by 100 mm rectangular area is divided into a uniform grid of 6 columns and 5 rows, thereby generating 30 initial field data blocks that are completely located within the boundary, and 5 initial field data blocks that partially cover the boundary, with physical gaps between these initial field data blocks.

[0021] In some embodiments, the initial field data block set undergoes edge expansion processing, extending each edge of each initial field data block outward by a preset overlap distance, which is set to 2 mm. Through edge expansion, adjacent initial field data blocks overlap within the expanded area, and the overlapping portion is planned as the overlapping scan area. Simultaneously, a boundary trimming operation is performed on all expanded field data blocks, removing all portions exceeding the outer boundary data range of the design layout of the large-aperture blazed diffraction grating, ultimately forming the final spatial layout data containing 35 field data blocks and their overlapping scan areas. After identifying aperiodic or highly complex regions containing positioning alignment marks, these regions undergo local magnification processing. Local magnification processing refers to extracting the design data of this region separately and performing high-resolution mesh analysis to analyze its local characteristic frequencies. The local characteristic frequency is defined as the number of times the structural edge direction changes per unit length within this region. In specific implementations, the calculated local characteristic frequency values ​​for aperiodic or highly complex regions are higher than those for periodic scribed regions. The density of the field data blocks is dynamically adjusted based on the calculated local characteristic frequencies. For non-periodic or highly complex regions with high calculated local characteristic frequencies, a smaller grid granularity is used for coverage, i.e., field data blocks with smaller reference side lengths are used. For periodic scribed regions with low calculated local characteristic frequencies, field data blocks with a reference side length of 20 mm are maintained. Optionally, dynamic adjustment is guided by the following formula: in: This indicates the number of field data blocks required to cover a specific area 'a' after adjustment. This represents the length of region a in the dividing direction. Indicates the reference side length. This represents the calculated local characteristic frequency value of region a. and These represent the maximum and minimum calculated values ​​of local characteristic frequencies for all analyzed regions within the entire design layout, respectively. It is a density adjustment coefficient between 0 and 1, used to control the sensitivity of the partition density to the calculated values ​​of local feature frequencies. (Signature) This indicates rounding up. It can be understood that, in this way, feature-dense aperiodic or highly complex regions are covered by finer-grained field data blocks, while feature-sparse periodic regions are covered by coarser-grained field data blocks. The total number of field data blocks is optimized from the pre-defined 35 evenly divided blocks to 31.

[0022] Optionally, when analyzing local characteristic frequencies using magnified processing, the calculated local characteristic frequency value for the positioning alignment mark area containing curved patterns is significantly higher than that for areas containing only parallel straight lines. It can be understood that the operation of dynamically adjusting the partition density based on local characteristic frequencies is performed after identifying high-complexity areas, resulting in a non-uniform distribution of field data blocks across the entire design layout.

[0023] In one embodiment of the present invention, a key feature point extraction and morphology prediction model training are implemented using a design layout of a large-aperture blazed diffraction grating for a spectrometer as an example. Key feature points related to morphology control are extracted from the design layout of the large-aperture blazed diffraction grating. (See reference...) Figure 3 The gradient calculation of the design layout of the large-aperture blazed diffraction grating was performed using the Sobel operator. Edge regions in the layout image whose gray-level value change rate exceeded a set threshold were identified. The set threshold was set to 15% of the gray-level gradient amplitude. Uniform sampling was performed in the edge regions at fixed intervals of five pixels to obtain one thousand key feature points. The spatial layout data of the field and overlapping areas, the key feature points, and the pre-defined standard topography reference surface were correlated. The standard topography reference surface is an ideal plane. A multi-dimensional feature vector was constructed, which includes the field position coordinates, the key feature point coordinates, and the normal vector of the standard topography reference surface. The field position coordinates are the three-dimensional coordinates of the center point of the field data block, the key feature point coordinates are the two-dimensional image coordinates of each sampling point, and the normal vector of the standard topography reference surface is the unit normal vector of the reference surface at that feature point. In some embodiments, actual morphological measurement data accumulated during historical direct writing processes are utilized. This data is acquired using a white light interferometer. The deviation between the actual morphology and the standard morphological reference surface under different scanning path sequences and different combinations of dose distributions in the overlapping areas is labeled. The deviation is measured in nanometers and serves as the supervised learning label for the machine learning model. Support vector regression (SVR) is selected as the main structure of the machine learning model. SVR uses a radial basis function as its kernel function, takes a multidimensional feature vector as input, and the supervised learning label as output to train the SVR algorithm. Optionally, the multidimensional feature vector has six dimensions, including the X and Y coordinates of the field center point, the U and V coordinates of key feature points, and the Nx and Ny components of the normal vector. The training process includes data normalization, cross-validation, and hyperparameter tuning. The hyperparameters include a penalty parameter C and a kernel function parameter γ, resulting in a trained morphological prediction model that can output predicted values ​​of morphological deviation. In a specific implementation, the prediction function of the SVR algorithm is expressed as: in: This represents the predicted value of the morphological deviation. This represents the i-th multidimensional feature vector in the training samples. This represents the multidimensional feature vector to be predicted. This indicates the number of support vectors. and It is a Lagrange multiplier. It is a bias term. It is the radial basis kernel function, defined as , This is the kernel function parameter, which controls the width of the kernel function. In some embodiments, the training dataset contains 10,000 sets of historical processing data. Each set of data includes a multi-dimensional feature vector and the corresponding actual shape deviation. The dataset is divided into a training set and a test set in an 8:2 ratio. The support vector regression algorithm is trained using the training set, and the model's prediction accuracy is evaluated using the test set.

[0024] In one embodiment of the present invention, dose-path joint optimization is implemented using the fifth field data block of a large-aperture blazed diffraction grating as an example. A pre-trained morphology prediction model is used to ergonomically deduce the scanning path sequence within the fifth field data block, and the exposure dose distribution within the overlapping scanning area corresponding to the fifth field data block is simultaneously parameterized. Three candidate scanning path sequences are defined for the fifth field data block: unidirectional scanning, spiral scanning, and helical scanning. The unidirectional scanning path sequence corresponds to a path parameter, namely the scan line spacing; the spiral scanning path sequence corresponds to the path parameter, the number of spiral turns; and the helical scanning path sequence corresponds to the path parameter, the helical starting radius. A dose gradient function is defined for the overlapping scanning area corresponding to the fifth field data block. The dose gradient function includes two control parameters, which are used to control the dose peak in the central region of the overlapping area and the dose attenuation slope in the edge region, respectively. By adjusting these two control parameters, the exposure dose distribution within the overlapping scanning area is parameterized. In practical implementation, a one-dimensional coordinate system is established with the central axis of the overlapping scan area as the axis of symmetry. The origin of the one-dimensional coordinate system is located at the center of the overlapping scan area, and the positive direction points towards the interior of the fifth field data block. The dose gradient function is defined as a quadratic function about the one-dimensional coordinate system. The quadratic function takes the form of a parabola, with its vertex located at the origin of the one-dimensional coordinate system. The function value at the vertex is the peak dose, and the absolute values ​​of the slopes of the quadratic function on both sides of the vertex are the dose attenuation slopes. By adjusting the value of the peak dose, the cumulative exposure at the center of the overlapping area is controlled. By adjusting the value of the dose attenuation slope, the rate of change of exposure during the transition from the center of the overlapping area to the interior of the field is controlled. The mathematical expression of the dose gradient function is: in: Represents one-dimensional coordinates The relative exposure dose value at that location, This indicates the peak dose, i.e., the dose located at the origin ( The dose value at ) A coefficient representing the dose decay slope correlation, with a value greater than zero. This represents the position in a one-dimensional coordinate system, with values ​​ranging from one end of the overlap region to the other. In some embodiments, the path parameters of the candidate scan path order corresponding to the fifth field data block and the dose peak value of the dose gradient function in the overlap region are used. and dose attenuation slope coefficient These two control parameters, along with the spatial layout data of the field and overlapping area of ​​the fifth field data block and key feature points, together form the current joint optimization variable set. The joint optimization variable set is a vector containing seven elements. The joint optimization variable set is input into the trained topography prediction model to calculate the predicted value of the topography deviation under the current joint optimization variable set. The initial calculation result of the predicted value of the topography deviation is 15.2 nanometers. Comparing the predicted topographic deviation value of 15.2 nm with the predicted topographic deviation value from the previous iteration (since this is the first iteration, the predicted topographic deviation value from the previous iteration is set to infinity), the absolute value of the difference between the two is greater than the preset convergence threshold of 1.0 nm. Therefore, the seven parameters in the joint optimization variable set are adjusted using the gradient descent method, with the learning rate set to 0.01. A new joint optimization variable set is generated, and the process of substituting these parameters into the trained topographic prediction model for calculation and comparison is repeated. After the seventh iteration, the change in the predicted topographic deviation value is less than the convergence threshold of 1.0 nm for three consecutive iterations. The scan path sequence and overlap region dose distribution corresponding to the seventh iteration are used as the optimal scan path sequence and optimal overlap region dose distribution for the fifth field data block. The optimal scan path sequence is a spiral scan with a spiral starting radius of 0.1 mm. The optimal overlap region dose distribution has a peak dose value of... The dose attenuation slope coefficient is 1.25 (normalized units). It is 0.08 per square millimeter.

[0025] Optionally, time-series constraints between fields are introduced during the simulation. Based on the kinematic model of the direct-write device, which considers the acceleration, constant speed, and deceleration phases of the platform, the total time required to complete the previous fourth field data block is estimated to be 120 seconds. Combined with the estimated time consumption of different candidate scan path sequences within the fifth field data block, the scan path sequence for the fifth field data block is filtered. The estimated time consumption for unidirectional scanning is 95 seconds, for spiral scanning is 110 seconds, and for helical scanning is 105 seconds. From the perspective of motion stability, the range of values ​​for the joint optimization variables is further constrained, eliminating scan path sequences whose predicted morphology deviation is less than a preset morphology threshold of 20 nanometers but would cause frequent start-stop operations of the device. For example, a specific spiral scanning variant requiring multiple sudden stops and starts at the field boundaries is eliminated. It can be understood that by introducing time-series constraints, the final optimal scan path sequence simultaneously considers morphology quality and device motion efficiency. In some embodiments, when adjusting the joint optimization variable set, the gradient descent method prioritizes adjusting the peak dose. and dose attenuation slope coefficient These two parameters are sensitive to the smoothness of the morphology. Optionally, the preset morphology threshold of 20 nm is determined according to the process specifications. It can be understood that the ergonomic derivation and iterative calculation process is performed independently for each field data block, ultimately generating the optimal scan path sequence and optimal overlap region dose distribution for each field data block.

[0026] In one embodiment of the present invention, a dose-path joint control instruction set is generated using a large-aperture blazed diffraction grating containing five field data blocks as an example. The optimal scan path sequence and optimal overlap region dose distribution corresponding to the five field data blocks are fused. A global process control data table is created, indexed by the field data block number, with each row recording all process parameters corresponding to one field data block. The name of the optimal scan path sequence corresponding to each field data block, the path parameters, the two control parameters of the optimal overlap region dose distribution (peak dose and dose attenuation slope), and the coordinate offset of the field data block in the overall design layout are sequentially filled into the corresponding rows of the process control data table. See Table 1 for an example of the process control data table content.

[0027] Table 1: Process Control Data Sheet In practice, after data is entered, structured intermediate data is formed. Following the order of the field data block numbers, each row of the process control data table is read sequentially, and the structured intermediate data of each row is translated into machine instructions that the direct-write equipment can recognize. Machine instructions include movement instructions, exposure start instructions, dose parameter setting instructions, and exposure stop instructions. Movement instructions contain the X and Y coordinate offsets of the target position; the exposure start instruction triggers the beam to open; the dose parameter setting instruction is used to write the dose peak and dose attenuation slope coefficients; and the exposure stop instruction is used to close the beam. All translated machine instructions are arranged according to the field data block numbers, and a synchronization calibration instruction is inserted between the machine instructions of adjacent field data blocks. In some embodiments, immediately after each exposure stop instruction indicating the end of a field data block, a grating position reset instruction is inserted. This grating position reset instruction is used to reset the grating system of the direct-write equipment to a known reference position, which is the mechanical zero point of the grating. Following the grating position reset command, a dose monitor reading command is inserted. This command reads the actual deposited dose value and compares it with the theoretically calculated value to obtain dose deviation data. Optionally, if the dose deviation data exceeds the allowable range, a dose compensation coefficient adjustment command is inserted after the dose monitor reading command. This command corrects the dose parameters used in subsequent field-specific data blocks based on the dose deviation data. The formula used for the dose compensation coefficient adjustment command is: in: This represents the new dose compensation coefficient calculated from subsequent field-specific data blocks. This indicates the dose compensation factor currently in use. This represents the theoretically calculated dose value for the current field data block. This represents the actual deposited dose value read from the dose monitor. It can be understood that real-time comparison and adjustment ensure the dose accuracy of a complete dose-path joint control instruction set during continuous execution. These instructions are ultimately concatenated into a complete dose-path joint control instruction set, which is a list of hundreds of ordered machine instructions. In some embodiments, the insertion logic of synchronous calibration instructions is mandatory; that is, regardless of whether the process parameters of adjacent field data blocks are the same, grating position reset instructions and dose monitor reading instructions are inserted between their instruction blocks. The insertion of dose compensation coefficient adjustment instructions is conditional; they are only generated and inserted when a dose deviation exceeding the allowable range is detected. Optionally, the allowable range is set to ±3% of the theoretically calculated dose value. It can be understood that the final generated complete dose-path joint control instruction set is used to guide the direct-write equipment to complete seamless direct writing of large-aperture diffraction structures.

[0028] See Figure 4 This is a simulation analysis chart of dose compensation coefficients, showing the changes in dose compensation coefficients for five data blocks in different fields, with an allowable deviation range of ±3%. The compensation coefficients for all fields fall within the allowable deviation range of 0.97 to 1.03, indicating that the dose monitoring and compensation strategy is effective and no additional dose compensation coefficient adjustment commands were triggered. Field 1 has the highest compensation coefficient, close to the +3% upper limit, reflecting that the actual deposited dose is slightly lower than the theoretical value. Field 4 has the lowest compensation coefficient, reflecting that the actual deposited dose is slightly higher than the theoretical value. Overall, the fluctuations are stable, with no cases exceeding the threshold, ensuring the dose accuracy of continuous processing. This result verifies the effectiveness of the synchronous calibration command; the grating position reset and dose monitoring process can stably control the dose deviation within the allowable range, avoiding abnormal seam morphology.

[0029] In one embodiment of the present invention, a logical integrity check is performed using a complete set of dose-path joint control instructions for a large-aperture blazed diffraction grating as an example. After generating the complete set of dose-path joint control instructions, a logical integrity check is performed on the complete set of dose-path joint control instructions. The logical integrity check checks for missing field data blocks, coordinate offset conflicts, or dose parameters exceeding limits. In specific implementation, the check process first parses all movement instructions in the instruction set, extracts the unique number and coordinate offset corresponding to each field data block, and forms a field data block coordinate mapping table. Then, the extracted field data block numbers are compared with the total number of field data blocks recorded in the spatial layout data of the field and overlapping areas to check for missing numbers, i.e., missing field data blocks. Next, the field data block coordinate mapping table is traversed, and the coordinate offsets of any two field data blocks are compared to determine whether two different field data blocks are assigned the exact same coordinate offset, i.e., coordinate offset conflicts. At the same time, it parses all the dose parameter setting instructions in the instruction set, reads the values ​​of the dose peak and dose attenuation slope parameters, compares them with the parameter safety range defined in the direct writing equipment process manual, and checks whether any parameter values ​​exceed the maximum or minimum allowable values, i.e., dose parameters are out of bounds.

[0030] In some embodiments, the logical integrity verification algorithm detects a coordinate offset conflict: the X-coordinate offset of field number 3 and field number 5 is 41.0 mm, but their Y-coordinate offsets are different. The verification algorithm also finds that the peak dose parameter of field number 4 is 1.22, exceeding the upper limit of 1.20 specified in the direct-write equipment process manual, which constitutes a dose parameter out-of-bounds error. Due to the detection of the coordinate offset conflict and dose parameter out-of-bounds anomalies, the algorithm backtracks to the spatial layout data generation step of the field and overlapping area. In the backtracking adjustment, the field division parameters are readjusted, the reference side length of the field data block is changed from 20 mm to 18 mm, and the number and layout of the field data blocks are recalculated. The subsequent geometric analysis, morphology prediction model training, dose-path joint optimization, and data fusion processes are then re-executed. Optionally, the severity of the coordinate offset conflict can be quantified and evaluated using the following formula: in: This represents the Euclidean distance between the coordinate offsets of two data blocks in different fields. and Represents field data blocks coordinate offset, and Represents field data blocks The coordinate offset. When When the value is less than the device positioning resolution, it is determined to be a coordinate offset conflict. It can be understood that by readjusting the field division parameters and re-executing the entire process, the newly generated dose-path joint control instruction set no longer contains anomalies such as missing field data blocks, coordinate offset conflicts, or dose parameter out-of-bounds errors. In some embodiments, after retrospective adjustment, the total number of field data blocks increases from 31 to 35. The coordinate offsets of field number 3 and field number 5 are replanned to ensure they are different from each other and conform to the device's movement range. The peak dose parameter of field number 4 is adjusted to 1.19 after re-optimization, falling within the device's allowable range.

[0031] See Figure 5This is a graph showing the training results of a shape prediction model. Both curves rapidly decrease from their initial high values, eventually stabilizing around round 100, indicating that the model is continuously learning and converging. The training loss eventually converges to near 0, while the validation loss converges to approximately 0.1-0.15, indicating a good overall fit. The fluctuation range of the training loss gradually decreases with increasing rounds, stabilizing between 0 and 0.1 in the later stages, suggesting a sufficient fit to the training data. The validation loss is consistently higher than the training loss and fluctuates more dramatically, especially showing significant peaks around rounds 20, 60, and 80, reflecting fluctuations in the model's generalization ability on unknown data. During the rapid decline period from rounds 0 to 20, the loss value drops sharply from its high level, indicating that the model is rapidly learning the basic pattern. During the fluctuating decline period from rounds 20 to 60, the validation loss shows multiple peaks, indicating that the model is adjusting its parameters to adapt to the validation data. During the stable convergence period from rounds 60 to 100, both loss curves stabilize, and the model's performance is essentially finalized.

[0032] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for dose-path joint optimization and stitching of large-aperture diffraction structures, characterized in that, The method includes: Geometric analysis was performed on the design layout of the large-aperture diffraction structure to generate spatial layout data of the field division and overlap regions. Key feature points related to morphology control are extracted from the design layout of the large-aperture diffraction structure. The spatial layout data of the field division and overlap area, as well as the key feature points, are used as input to train a machine learning model with the morphology smoothness of the splicing area as the optimization objective, and a trained morphology prediction model is obtained. Using the pre-trained morphology prediction model, the scanning path sequence within each sub-field data block is traversally deduced, and the exposure dose distribution within the overlapping scanning area is simultaneously parameterized. The deduced scanning path sequence within the sub-field and the parameterized overlapping area exposure dose distribution are used as joint optimization variables and substituted into the pre-trained morphology prediction model for iterative calculation until the morphology smoothness index output by the pre-trained morphology prediction model converges, thus obtaining the optimal scanning path sequence and optimal overlapping area dose distribution corresponding to each sub-field data block. The optimal scanning path sequence and optimal overlap region dose distribution corresponding to each field data block are fused to generate a complete set of dose-path joint control instructions for the large-aperture diffraction structure.

2. The method for dose-path joint optimization and stitching of large-aperture diffraction structures according to claim 1, characterized in that, The geometric analysis of the design layout of the large-aperture diffraction structure, forming spatial layout data of the field division and overlap regions, includes: Geometric analysis is performed on the design layout of the large-aperture diffraction structure to identify non-periodic or highly complex regions. Based on the dimensions of the design layout, a seamless field segmentation operation is performed, dividing the entire design layout into several field data blocks with fixed spatial relationships. Simultaneously, an overlapping scanning area is planned for each field data block to eliminate splicing errors, forming spatial layout data for the fields and overlapping areas. Specifically, this includes: The edge contour of the design layout of the large-aperture diffraction structure is extracted to obtain the outer boundary data of the design layout, and the area and aspect ratio of the region enclosed by the outer boundary data are calculated. Based on the area and aspect ratio, and combined with the maximum single processing field of view of the direct writing device, the number of sub-field data blocks and the reference side length of each sub-field data block are determined, wherein the reference side length is less than the maximum single processing field of view. Using the external boundary data of the design layout as a reference, the entire design layout is divided into rectangular grids according to the determined number of field data blocks and the reference side length, generating an initial set of field data blocks. There are uncovered gaps between the field data blocks in the initial set of field data blocks. The initial set of field data blocks is subjected to edge expansion processing, extending the edge of each field data block outward by a preset overlap distance, so that adjacent field data blocks overlap. The overlapping part is the overlapping scanning area. At the same time, the part that exceeds the outer boundary data range of the design layout is trimmed to form the final spatial layout data of the field and the overlapping area.

3. The method for dose-path joint optimization and stitching of large-aperture diffraction structures according to claim 2, characterized in that, Key feature points related to morphology control are extracted from the design layout of the large-aperture diffraction structure. Using the spatial layout data of the field division and overlap regions, along with these key feature points, as input, a machine learning model is trained with the smoothness of the splicing region's morphology as the optimization objective. This yields a trained morphology prediction model, including: Gradient calculation is performed on the design layout of the large-aperture diffraction structure to identify edge regions in the design layout whose gray value change rate exceeds a set threshold, and the key feature points are obtained by uniformly sampling in the edge regions. The spatial layout data of the field and the overlapping area, the key feature points, and the pre-set standard topography reference surface are correlated to construct a multi-dimensional feature vector. The multi-dimensional feature vector includes the field position coordinates, the key feature point coordinates, and the normal vector of the standard topography reference surface. Using actual morphology measurement data accumulated during historical direct writing, the deviation between the actual morphology and the standard morphology reference surface under different scanning path sequences and different combinations of dose distribution in the overlapping area is marked, serving as the supervised learning label for the machine learning model. The Support Vector Regression algorithm is selected as the main structure of the machine learning model. The multidimensional feature vector is used as input and the supervised learning label is used as output. The Support Vector Regression algorithm is trained to obtain the trained shape prediction model that can output the predicted value of shape deviation. The smaller the absolute value of the predicted value of shape deviation, the higher the shape smoothness.

4. The method for dose-path joint optimization and stitching of large-aperture diffraction structures according to claim 3, characterized in that, Using the pre-trained morphology prediction model, the scanning path sequence within each sub-field data block is traversally deduced, and the exposure dose distribution within the overlapping scanning area is simultaneously parameterized. The deduced scanning path sequence within the sub-field and the parameterized overlapping area exposure dose distribution are used as joint optimization variables and substituted into the pre-trained morphology prediction model for iterative calculation until the morphology smoothness index output by the pre-trained morphology prediction model converges, thus obtaining the optimal scanning path sequence and optimal overlapping area dose distribution for each sub-field data block, including: Multiple candidate scanning path sequences are defined for the current field data block to be optimized. The candidate scanning path sequences include unidirectional scanning, loop scanning and spiral scanning, and each scanning path sequence corresponds to a path parameter. A dose gradient function is defined for the overlapping scanning area. The dose gradient function includes two control parameters, which are used to control the dose peak in the center region and the dose attenuation slope in the edge region of the overlapping area, respectively. By adjusting the two control parameters, the parameterized adjustment of the exposure dose distribution in the overlapping area is achieved. The path parameters of the candidate scanning path sequence corresponding to the current field data block, the two control parameters of the dose gradient function in the overlapping area, and the spatial layout data and key feature points of the field and the overlapping area are combined to form the current joint optimization variable group. The current set of joint optimization variables is input into the trained shape prediction model to calculate the predicted value of shape deviation under the current set of joint optimization variables. The predicted value of the topographic deviation is compared with the predicted value of the topographic deviation in the previous iteration. If the difference is greater than the preset convergence threshold, the parameters in the joint optimization variable group are adjusted by gradient descent to generate a new joint optimization variable group. The process of substituting and comparing is repeated until the change in the predicted value of the topographic deviation is less than the convergence threshold multiple times in a row. The corresponding scanning path sequence and overlapping area dose distribution at this time are taken as the optimal scanning path sequence and optimal overlapping area dose distribution.

5. The method for dose-path joint optimization and stitching of large-aperture diffraction structures according to claim 4, characterized in that, A dose gradient function is defined for the overlapping scan area. This function includes two control parameters, which control the dose peak in the central region and the dose attenuation slope in the edge region, respectively. By adjusting these two control parameters, parameterized adjustment of the exposure dose distribution in the overlapping area is achieved, including: A one-dimensional coordinate system is established with the central axis of the overlapping scanning area as the axis of symmetry. The origin of the one-dimensional coordinate system is located at the center of the overlapping scanning area, and the positive direction points to the interior of the field data block. The dose gradient function is defined as a quadratic function about the one-dimensional coordinate. The quadratic function is in the form of a parabola, with its vertex located at the origin of the one-dimensional coordinate system. The function value at the vertex is the peak dose, and the absolute value of the slope of the quadratic function on both sides of the vertex is the dose attenuation slope. By adjusting the value of the peak dose, the cumulative exposure at the center of the overlapping area is controlled. By adjusting the value of the dose attenuation slope, the rate of change of exposure during the transition from the center of the overlapping area to the interior of the subfield is controlled, thereby generating the overlapping area exposure dose distribution that meets the input requirements of the trained morphology prediction model.

6. The method for dose-path joint optimization and stitching of large-aperture diffraction structures according to claim 5, characterized in that, The process involves fusing the optimal scanning path sequence and optimal overlap region dose distribution corresponding to each field data block to generate a complete dose-path joint control instruction set for the large-aperture diffraction structure, including: Create a global process control data table, which is indexed by the number of the sub-field data block, and each row records all process parameters corresponding to a sub-field data block; The name of the optimal scanning path sequence, path parameters, two control parameters of the optimal overlap area dose distribution, and the coordinate offset of the field data block in the overall design layout are sequentially filled into the corresponding rows of the process control data table to form structured intermediate data. According to the order of the field data block numbers, each row of the process control data table is read sequentially, and the structured intermediate data of each row is translated into machine instructions that can be recognized by the direct writing equipment. The machine instructions include movement instructions, exposure start instructions, dose parameter setting instructions and exposure stop instructions. All machine instructions obtained from translation are arranged in the order of field data block numbers, and synchronous calibration instructions are inserted between machine instructions of adjacent field data blocks, and finally connected in series to form the complete set of dose-path joint control instructions. The dose-path joint control instruction set is used to guide the direct writing device to complete seamless direct writing of large-aperture diffraction structures.

7. The method for dose-path joint optimization and stitching of large-aperture diffraction structures according to claim 6, characterized in that, The insertion of synchronization calibration instructions between machine instructions of adjacent field data blocks includes: Immediately after each stop exposure command indicating the end of a field data block, a raster position reset command is inserted to reset the raster system of the direct-write device to a known reference position. After the grating position reset command, a dose monitor reading command is inserted. The dose monitor reading command is used to read the current actual deposited dose value and compare the actual deposited dose value with the theoretical calculated value to obtain dose deviation data. If the dose deviation data exceeds the allowable range, a dose compensation coefficient adjustment instruction is inserted after the dose monitor reading instruction. The dose compensation coefficient adjustment instruction is used to correct the dose parameters used in subsequent field data blocks based on the dose deviation data, so as to ensure the dose accuracy of the entire dose-path joint control instruction set during continuous execution.

8. The method for dose-path joint optimization and stitching of large-aperture diffraction structures according to claim 7, characterized in that, The step of performing geometric analysis on the design layout of the large-aperture diffraction structure to identify non-periodic or highly complex regions, and performing seamless field segmentation based on the dimensions of the design layout of the large-aperture diffraction structure, further includes: After identifying aperiodic or highly complex regions, the regions are locally magnified to analyze their local feature frequencies. Based on these local feature frequencies, the partitioning density of the field data blocks containing the aperiodic or highly complex regions is dynamically adjusted. This allows regions with dense features to be covered by finer-grained field data blocks, while regions with sparse features are covered by coarser-grained field data blocks. This reduces the total number of field data blocks while maintaining accuracy, thereby lowering the computational burden on the trained shape prediction model.

9. The method for dose-path joint optimization and stitching of large-aperture diffraction structures according to claim 8, characterized in that, The step of using the trained shape prediction model to traverse and deduce the scanning path sequence within each sub-field data block also includes: During the simulation, time series constraints between fields are introduced. Based on the kinematic model of the direct-write device, the total time required to complete the previous field data block is estimated. Combined with the estimated time consumption of the scanning path sequence within the field data block, the scanning path sequence of the next field data block is screened. Scanning path sequences that have a predicted morphology deviation value less than a preset morphology threshold but cause frequent start-stop of the device are eliminated. From the perspective of motion stability, the range of values ​​of the joint optimization variables is further constrained so that the final optimal scanning path sequence takes into account both morphology quality and device motion efficiency.

10. The method for dose-path joint optimization and stitching of large-aperture diffraction structures according to claim 9, characterized in that, The step of fusing the optimal scanning path sequence and optimal overlap region dose distribution corresponding to each field data block to generate a complete dose-path joint control instruction set for the large-aperture diffraction structure further includes: After generating the complete set of dose-path joint control instructions, a logical integrity check is performed on the complete set of dose-path joint control instructions to check for missing field data blocks, coordinate offset conflicts, or dose parameters exceeding the limits. If an anomaly is detected, the process is traced back to the spatial layout data generation step of the field and overlapping area, the field division parameters are readjusted, and the subsequent optimization and fusion process is re-executed until a logically complete and error-free complete set of dose-path joint control instructions is generated.

Images