Online incremental learning neural network optimization method for laser direct writing alignment model

Abstract

The application provides an online incremental learning neural network optimization method for a laser direct writing alignment model, relates to the technical field of data processing, and comprises the following steps: acquiring an alignment mark image, a laser scanning path and processing result data, extracting spatial coordinates and constructing a spatial grid mapping relationship; establishing an initial alignment model based on historical samples, and dividing model parameters into parameter sub-regions with an adjacency relationship according to the spatial grid; in the processing process, inputting current data into the model to obtain an alignment compensation result, combining the processing result to calculate an alignment residual error, determining a target spatial grid region and a corresponding parameter sub-region; incrementally adjusting the target parameter sub-region, and updating the target parameter sub-region according to a decay rule and limited propagation between adjacent sub-regions to realize online optimization of the model; and finally superimposing the updated alignment compensation result on the laser scanning path to realize partition correction of a scanning track and improve alignment accuracy and processing stability.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an online incremental learning neural network optimization method for laser direct writing alignment models. Background Technology

[0002] In existing technologies, laser direct writing alignment and processing optimization are typically achieved by combining photolithography control models with neural network methods. Specifically, a training dataset is constructed by collecting exposure dose distribution and actual processing morphology data. A convolutional neural network (such as U-net) is then used to establish a nonlinear mapping model between exposure parameters and forming results. After offline training, this model is used to predict optimal exposure control parameters, thereby optimizing alignment and processing accuracy in exposure, development, and other steps. Simultaneously, the system relies on computer-controlled laser beam scanning and CCD alignment detection to achieve mapping and alignment correction from graphic data to the actual structure.

[0003] However, in advanced packaging applications such as large-size PCB laser direct writing, these methods typically use fixed training models for inference. When the exposure response changes due to batch material variations or equipment thermal drift during processing, the model cannot be updated in real time. For example, in continuous roll-to-roll exposure, changes in substrate tension can cause misalignment of alignment marks. If a pre-trained model is still used for compensation, it will result in accumulated alignment errors, increasing the deviation in subsequent interlayer stacking and leading to localized circuit misalignment or open circuit defects. This lack of online incremental learning capability directly affects the consistency and yield of multilayer structure processing. Summary of the Invention

[0004] The purpose of this invention is to provide an online incremental learning neural network optimization method for laser direct writing alignment models, aiming to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] An online incremental learning neural network optimization method for laser direct-write alignment models, the method comprising:

[0007] Acquire alignment mark images, laser scanning path data, and processing result data during the laser direct writing process; extract spatial coordinate information from the alignment mark images; divide the processing area into grids based on the spatial coordinate information, establish a spatial grid mapping relationship between each grid position and the local position of the laser scanning path, and generate spatial grid mapping results.

[0008] An initial alignment model is established based on historical processing samples; the parameters of the initial alignment model are partitioned according to the spatial grid mapping results, so that the parameters in the initial alignment model are divided into multiple parameter sub-regions corresponding to different spatial grid positions, and spatial adjacency relationships are established for each parameter sub-region to generate an alignment model with spatial region correspondence.

[0009] The alignment mark image and laser scanning path data in the current processing are input into the alignment model to obtain the alignment compensation results corresponding to each spatial grid position; the alignment compensation results are compared with the corresponding processing result data to obtain the alignment residual distribution results, and the target spatial grid area is determined based on the alignment residual distribution results.

[0010] The target parameter sub-region is determined based on the target spatial grid region; the model parameters within the target parameter sub-region are incrementally adjusted, and parameter propagation boundaries are set between the parameter sub-regions to constrain the propagation path of the adjustment amount, so that the adjustment amount diffuses along the spatial adjacency relationship between the parameter sub-regions according to the preset decay rule, and the incrementally updated alignment model is generated.

[0011] The alignment compensation results output by the incrementally updated alignment model are superimposed onto the corresponding local positions in the laser scanning path, and the laser scanning trajectory is partitioned and corrected to obtain the updated laser scanning control results.

[0012] The above-described solution of the present invention has at least the following beneficial effects:

[0013] First, by acquiring alignment mark images, laser scanning paths, and processing result data, and constructing a spatial grid mapping relationship, this invention achieves a spatial correspondence between the processing area and model parameters, enabling the alignment model to transform from a unified overall modeling approach to a regionalized modeling approach with spatial resolution capabilities, thereby improving the ability to express local errors.

[0014] Secondly, based on this, the parameters of the initial alignment model are divided into multiple parameter sub-regions corresponding to the spatial grid positions, and spatial adjacency relationships are established, so that the model has spatial structural characteristics consistent with the actual processing area. Compared with the traditional overall model, it can independently optimize local areas and reduce the problem of mutual interference between errors in different areas.

[0015] Furthermore, by comparing the alignment compensation results with the actual processing results, the present invention obtains the alignment residual distribution. Combined with abnormal grid identification and continuous offset determination, it can accurately locate the target spatial grid region where a systematic offset occurs, thereby avoiding ineffective adjustments to non-critical regions and improving the targeting of updates.

[0016] Based on this, incremental adjustments are made only to the target parameter sub-regions. By setting parameter propagation boundaries and decay rules, the adjustment amount is allowed to diffuse in a limited manner under spatial adjacency, achieving a smooth transition update from local to neighborhood. This not only ensures the stability of model updates but also effectively prevents global performance degradation caused by over-adjustment.

[0017] Finally, by superimposing the incrementally updated alignment compensation results onto the corresponding local positions of the laser scanning path, the scanning trajectory is corrected in a partitioned manner. This enables the system to continuously adapt to changing factors (such as material differences and equipment drift) during the processing, thereby significantly reducing the cumulative alignment error, improving the consistency and yield of multi-layer structure processing, and overcoming the problem that the model cannot be updated online in the prior art. Attached Figure Description

[0018] Figure 1 This is a flowchart of an online incremental learning neural network optimization method for a laser direct-write alignment model provided in an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0020] like Figure 1 As shown, embodiments of the present invention propose an online incremental learning neural network optimization method for laser direct writing alignment models, the method comprising:

[0021] Acquire alignment mark images, laser scanning path data, and processing result data during the laser direct writing process; extract spatial coordinate information from the alignment mark images; divide the processing area into grids based on the spatial coordinate information, establish a spatial grid mapping relationship between each grid position and the local position of the laser scanning path, and generate spatial grid mapping results; the local position is the position segment / coordinate segment in the physical scanning path.

[0022] An initial alignment model is established based on historical processing samples; the parameters of the initial alignment model are partitioned according to the spatial grid mapping results, so that the parameters in the initial alignment model are divided into multiple parameter sub-regions corresponding to different spatial grid positions, and spatial adjacency relationships are established for each parameter sub-region to generate an alignment model with spatial region correspondence.

[0023] The alignment mark image and laser scanning path data in the current processing are input into the alignment model to obtain the alignment compensation results corresponding to each spatial grid position; the alignment compensation results are compared with the corresponding processing result data to obtain the alignment residual distribution results, and the target spatial grid area is determined based on the alignment residual distribution results.

[0024] The target parameter sub-region is determined based on the target spatial grid region; the model parameters within the target parameter sub-region are incrementally adjusted, and parameter propagation boundaries are set between the parameter sub-regions to constrain the propagation path of the adjustment amount, so that the adjustment amount diffuses along the spatial adjacency relationship between the parameter sub-regions according to the preset decay rule, and the incrementally updated alignment model is generated.

[0025] The alignment compensation results output by the incrementally updated alignment model are superimposed onto the corresponding local positions in the laser scanning path, and the laser scanning trajectory is partitioned and corrected to obtain the updated laser scanning control results.

[0026] In this embodiment of the invention, by acquiring alignment mark images and extracting spatial coordinate information, and combining them with laser scanning path data to establish a spatial grid mapping relationship for the processing area, a correspondence is formed between each position in the processing area and the local position of the scanning path. This transforms the alignment problem from an overall error problem into a local error problem with spatial positioning capabilities, providing a clear data foundation for subsequent adjustment of model parameters by region.

[0027] After establishing the initial alignment model, the model parameters are partitioned and configured according to the spatial grid mapping results, and spatial adjacency relationships are established between each parameter sub-region. This ensures that the model parameters maintain a consistent mapping relationship with the spatial positions in the actual processing area, so that the model can output corresponding results for different spatial positions when performing alignment compensation, avoiding mutual interference between errors in different regions.

[0028] After comparing the alignment compensation results with the processing results data, the alignment residual distribution is obtained, and the target space grid region is determined accordingly. This allows the model parameter adjustment to be performed only on the local areas where there is offset, thereby avoiding the generation of new deviations in the already stable areas due to the uniform adjustment of the entire model, and improving the stability of the alignment correction process.

[0029] When making incremental adjustments to the target parameter sub-regions, parameter propagation boundaries are set between parameter sub-regions, and the propagation path of the adjustment amount is constrained, so that the adjustment amount is diffused in a controlled manner along the spatial adjacency relationship. In this way, while correcting the parameters of the target region, the adjustment effect is gradually transmitted to the related regions. The propagation amplitude is controlled by the attenuation rule, so that the error correction between different spatial regions remains continuous and the processing instability caused by local abrupt changes is avoided.

[0030] After obtaining the incrementally updated alignment model, the alignment compensation results are superimposed on the corresponding local positions in the laser scanning path. The scanning trajectory is then partitioned and corrected so that the alignment errors at each spatial grid position gradually converge. This suppresses the cumulative expansion of alignment errors during continuous processing and keeps the processing accuracy within a stable range.

[0031] In a preferred embodiment of the present invention, acquiring alignment mark images, laser scanning path data, and processing result data during the laser direct writing process includes:

[0032] The imaging acquisition unit installed on the laser direct writing device performs real-time image acquisition of the alignment marks in the processing area. The imaging acquisition unit triggers the acquisition operation according to the preset time interval or scanning node during the laser scanning process, generating an alignment mark image containing the position and shape information of the alignment marks.

[0033] The laser scanning control unit records the motion trajectory information of the laser beam during the processing. The laser scanning path is recorded in time sequence or spatial position to generate corresponding laser scanning path data. The laser scanning path data includes the scanning start position, scanning direction and position change sequence during the scanning process.

[0034] After processing is completed, the processing area is inspected by the detection unit to obtain actual processing result data. The processing result data includes the actual position, shape contour and offset information of the processed structure relative to the design position.

[0035] The alignment mark image, laser scanning path data, and processing result data are stored together according to the same processing cycle to form a data set corresponding to the same processing batch.

[0036] In a preferred embodiment of the present invention, extracting spatial coordinate information from the alignment marker image includes:

[0037] Preprocessing operations are performed on the alignment marker image, including grayscale conversion and noise suppression, to improve the recognition clarity of the alignment marker region and generate a preprocessed image.

[0038] Identify the boundary regions of alignment markers in the preprocessed image, determine the contour range of the alignment markers through boundary detection, and generate alignment marker regions;

[0039] The geometric center position is calculated based on the alignment mark region, and this geometric center position is used as the spatial position representation of the alignment mark. The geometric center position is obtained by statistically averaging the positions of each pixel in the region.

[0040] Based on the geometric center position and the calibration relationship of the imaging system, the image coordinates are converted into actual spatial coordinates in the coordinate system of the processing equipment, and the corresponding spatial coordinate information is generated.

[0041] In a preferred embodiment of the present invention, the processing area is divided into grids based on spatial coordinate information, a spatial grid mapping relationship is established between each grid position and a local position of the laser scanning path, and a spatial grid mapping result is generated, including:

[0042] The boundary range of the processing area is determined based on the spatial coordinate information, and the processing area is divided into multiple grid units according to the preset grid division rules. The grid division rules are to divide the processing area into multiple equally spaced intervals in the horizontal and vertical directions, thereby forming a regularly arranged grid structure.

[0043] Each grid cell is numbered so that each grid cell corresponds to a unique grid location identifier, thereby generating a set of grid locations;

[0044] Based on the laser scanning path data, the spatial distribution of the laser scanning path in the processing area is determined, and the laser scanning path is divided into multiple local path segments, so that each path segment corresponds to a part of the spatial range in the processing area.

[0045] Based on the spatial location of each grid cell and the coverage relationship of each local path segment, the correspondence between each grid cell and the corresponding local path segment is determined, and the mapping relationship between grid location and local path segment is generated.

[0046] Based on the mapping relationship, each grid position is associated with the corresponding local position of the laser scanning path to generate a spatial grid mapping result.

[0047] In a preferred embodiment of the present invention, establishing an initial alignment model based on historical processing samples includes:

[0048] Acquire alignment mark images, laser scanning path data, and corresponding processing result data from multiple historical processing cycles, and organize the data from each historical processing cycle to generate a historical processing sample set;

[0049] Based on the historical processing sample set, the spatial coordinate information, scanning path information and corresponding morphological deviation features in each sample are correlated to form a one-to-one correspondence between input data and output data, and to generate model training samples.

[0050] The model training samples are input into a preset neural network structure, and the model parameters are iteratively adjusted so that the model output gradually approximates the corresponding processing result data, thus generating an initial alignment model.

[0051] During model training, the difference between the model output and the actual processing result is evaluated, and the model parameters are adjusted according to the difference until the model output meets the preset error range requirements, thereby completing the establishment of the initial alignment model.

[0052] In a preferred embodiment of the present invention, determining the corresponding target parameter sub-region based on the target spatial grid region includes:

[0053] Based on the location identifiers of each grid in the target spatial grid region, obtain the location range of the target spatial grid region in the spatial grid mapping result, and generate a target grid location set;

[0054] Based on the target grid location set, find the parameter sub-region identifiers that correspond one-to-one with each grid location in the alignment model. The parameter sub-region identifiers are spatial location identifiers established during the model building phase, and generate a target parameter sub-region identifier set.

[0055] Based on the target parameter sub-region identifier set, extract the model parameter set corresponding to each identifier from the alignment model to generate the target parameter sub-region;

[0056] Based on the spatial adjacency relationship between the target parameter sub-region and its adjacent parameter sub-regions, the position of the target parameter sub-region in the parameter sub-region topology is determined, and its connection relationship with its adjacent parameter sub-regions is marked, generating parameter update positioning results.

[0057] In a preferred embodiment of the present invention, the alignment compensation result output by the incrementally updated alignment model is superimposed onto the corresponding local position in the laser scanning path, and the laser scanning trajectory is partitioned and corrected to obtain the updated laser scanning control result, including:

[0058] Based on the incrementally updated alignment model, input the alignment mark image and laser scanning path data in the current processing process to generate the alignment compensation result corresponding to each grid position. The alignment compensation result represents the position correction amount required for the corresponding grid position.

[0059] Based on the spatial grid mapping results, the alignment compensation results corresponding to each grid position are mapped to the corresponding local path segments in the laser scanning path to generate path compensation mapping results;

[0060] Based on the path compensation mapping results, the scanning position of each local path segment is corrected so that the position of each scanning point in the local path segment is superimposed with the corresponding alignment compensation result on the original scanning position, thereby generating the corrected local scanning path.

[0061] Based on each corrected local scanning path, the complete laser scanning path is stitched together to generate the overall corrected laser scanning trajectory.

[0062] Based on the overall corrected laser scanning trajectory, an updated laser scanning control command is output, causing the laser scanning control unit to perform the scanning operation according to the corrected trajectory and generate an updated laser scanning control result.

[0063] In a preferred embodiment of the present invention, the initial alignment model is configured with parameter partitions based on the spatial grid mapping results, so that the parameters in the initial alignment model are divided into multiple parameter sub-regions corresponding to different spatial grid positions, and spatial adjacency relationships are established for each parameter sub-region to generate an alignment model with spatial region correspondence, including:

[0064] Based on the spatial grid mapping results, the input region of the initial alignment model is partitioned and calibrated to determine the correspondence between each grid position in the processing region and each local data processing range in the input region, generating the correspondence results between grid positions and local data processing ranges; where the input region is the model input data space region after calibration and mapping with the processing region; and the local data processing range is the data processing block in the model input corresponding to the physical position.

[0065] Based on the correspondence between grid positions and local data processing ranges, the model parameters involved in the data processing of each local data processing range in the initial alignment model are divided to generate parameter sets corresponding to each grid position.

[0066] Based on the parameter set, the position identification of each parameter set is processed, and the spatial adjacency relationship between each grid position is mapped to the spatial adjacency relationship between each parameter set, generating parameter sub-regions with spatial position identification and spatial adjacency relationship;

[0067] Based on parameter sub-regions with spatial location identifiers and spatial adjacency relationships, construct a positioning model with spatial region correspondence relationships.

[0068] In this embodiment of the invention, by partitioning and labeling the input region of the initial alignment model and establishing a correspondence between the grid position of the processing region and the local data processing range, the model input data has a clear spatial regional affiliation, thereby ensuring that data from different spatial locations participate in the calculation only within their corresponding data processing range. Based on this, the model parameters participating in the data processing of each local data processing range are divided, and a one-to-one correspondence is established between the model parameters and the actual processing region through location identifiers and spatial adjacency mapping, ensuring that the internal parameter structure of the model is consistent with the spatial structure of the processing region. In this way, when performing alignment compensation, the model can call the corresponding parameter sub-region for a specific spatial region, avoiding parameter coupling interference between different regions. Simultaneously, it provides a structural basis for parameter propagation driven by spatial adjacency, enabling subsequent incremental adjustment processes to have a clear propagation path and constraint range.

[0069] In a preferred embodiment of the present invention, based on the spatial grid mapping results, the input region of the initial alignment model is partitioned and calibrated to determine the correspondence between the grid positions in the processing region and the local data processing ranges in the input region, generating the correspondence results between grid positions and local data processing ranges, including:

[0070] Based on the spatial grid mapping results, the spatial distribution range of each grid position in the processing area is obtained, and the position interval of each grid position in the actual processing coordinate system is extracted to generate the spatial range result of the grid position.

[0071] Based on the spatial range of the grid position, the input region of the initial alignment model is divided into multiple input sub-regions according to the same spatial division method as the processing region, so that each input sub-region corresponds to a grid position, and the input region partitioning result is generated.

[0072] Based on the input region partitioning results, the scope of each input sub-region in the model data processing process is determined. The scope of the scope is the data region that participates in the model input calculation, and the local data processing scope result is generated.

[0073] Based on the results of the grid location spatial range and the local data processing range, a one-to-one correspondence between each grid location and the corresponding local data processing range is established, and the corresponding results between the grid location and the local data processing range are generated.

[0074] Based on the corresponding results, the data stream in the input region is calibrated so that data from different grid locations enter the corresponding local data processing range in the model, forming the model input structure after partition calibration.

[0075] In a preferred embodiment of the present invention, a positioning model with spatial region correspondence is constructed based on parameter sub-regions having spatial location identifiers and spatial adjacency relationships, including:

[0076] Based on the parameter sub-regions with spatial location identifiers, determine the distribution position of each parameter sub-region in the model structure, and bind each parameter sub-region to the corresponding grid position to generate the binding relationship between the parameter sub-region and the spatial grid position;

[0077] Based on the parameter sub-regions with spatial adjacency, the connection relationships between each parameter sub-region are extracted, and the connection relationships are mapped to the connection structure between parameters inside the model, so that the connection method between model parameters is consistent with the mesh topology, and the parameter connection structure result is generated.

[0078] Based on the binding relationship between the parameter sub-regions and the spatial grid positions, as well as the parameter connection structure results, the parameter organization of the initial alignment model is reconstructed so that the model can perform partitioned calculations according to spatial regions when performing data processing, generating a model structure with spatial region constraints.

[0079] Based on the model structure constrained by spatial regions, the data transmission path of the model is configured so that the input data is processed in the model according to the corresponding parameter sub-regions, and information is transmitted between adjacent parameter sub-regions according to spatial adjacency, thereby generating a positioning model with spatial region correspondence.

[0080] In a preferred embodiment of the present invention, determining the target spatial grid region based on the alignment residual distribution results includes:

[0081] Based on the alignment residual distribution results, the residual information corresponding to each grid position is extracted, and the residual information corresponding to each grid position is judged to determine the abnormal grid positions that meet the preset residual judgment conditions, and an abnormal grid position set is generated.

[0082] Based on the set of anomalous grid locations, connectivity merging is performed on spatially adjacent anomalous grid locations to generate candidate spatial grid regions;

[0083] For each candidate spatial grid region, the residual change information corresponding to the candidate spatial grid region in multiple consecutive processing cycles is extracted, and the continuity of the residual change information is determined to generate the continuous offset region determination result.

[0084] Based on the results of the continuous offset region determination, the candidate spatial grid regions that meet the preset continuous offset determination conditions are determined as the target spatial grid regions.

[0085] In this embodiment of the invention, by performing grid-by-grid analysis on the alignment residual distribution results and filtering out abnormal grid locations based on preset residual judgment conditions, areas with offsets can be accurately identified. Furthermore, the spatial connectivity between grids is used to merge the abnormal grid locations, integrating discrete abnormal points into continuous regions, thereby avoiding local misjudgments caused by single-point errors. Further, by analyzing the residual change information of candidate spatial grid regions over multiple consecutive processing cycles and performing continuity judgment, regions that maintain a consistent offset trend across multiple processing cycles are identified as target spatial grid regions. This filters out interference from occasional errors or random fluctuations, allowing subsequent parameter adjustments to focus on regions with stable offset characteristics, improving the targeting and stability of parameter updates.

[0086] In a preferred embodiment of the present invention, the method for determining the preset residual judgment condition includes:

[0087] Obtain the alignment residual distribution results corresponding to multiple historical processing cycles, and extract the residual records of each grid position under normal processing conditions to generate a historical residual sample set;

[0088] Based on the historical residual sample set, the residual fluctuation range of each grid position under normal processing conditions is statistically analyzed, the residual reference range that can characterize the normal error variation range is determined, and the residual benchmark result is generated.

[0089] Based on the residual baseline results and the alignment accuracy requirements of the processing equipment, the upper limit of the residual is determined. The upper limit of the residual is the residual boundary that simultaneously meets the historical normal fluctuation range and the actual processing accuracy requirements, and the residual threshold results are generated.

[0090] Based on the residual threshold results, when the current residual at any grid position exceeds the upper limit of the residual, the grid position is determined to meet the preset residual judgment condition. Thus, the preset residual judgment condition is determined as an abnormal judgment condition that is jointly limited by the historical residual fluctuation range and the processing accuracy requirements.

[0091] In a preferred embodiment of the present invention, the method for determining the preset continuous offset determination condition includes:

[0092] Obtain residual change records corresponding to candidate spatial grid regions in multiple historical processing cycles, and extract the residual change direction and duration information of each candidate spatial grid region in continuous processing cycles to generate a historical continuous offset sample set.

[0093] Based on the historical continuous offset sample set, we distinguish between short-term residual changes caused by random fluctuations and continuous residual changes caused by changes in operating conditions, determine the minimum continuous processing cycle requirement that can characterize the continuous offset state, and generate cycle determination results.

[0094] Based on the historical continuous migration sample set, the occurrence of the same residual change direction in the same candidate spatial grid region within multiple consecutive processing cycles is statistically analyzed to determine the direction consistency requirement that can characterize the continuous migration state and generate direction determination results.

[0095] Based on the period determination result and the direction determination result, when the same candidate spatial grid region maintains a consistent residual change direction within a period of no less than a preset continuous processing cycle, the candidate spatial grid region is determined to meet the preset continuous offset determination condition, thereby determining the preset continuous offset determination condition as a continuous offset determination condition that simultaneously meets the continuous period requirement and the direction consistency requirement.

[0096] In a preferred embodiment of the present invention, incremental adjustments are made to the model parameters within the target parameter sub-region, and parameter propagation boundaries are set between the parameter sub-regions to constrain the propagation path of the adjustment amount, so that the adjustment amount diffuses along the spatial adjacency relationship between the parameter sub-regions according to a preset attenuation rule, generating an incrementally updated alignment model, including:

[0097] Based on the position identifier of the target spatial grid region in the alignment model, the target parameter sub-region corresponding to the target spatial grid region is determined, and based on the spatial adjacency relationship of the target parameter sub-region, the set of parameter sub-regions connected to the target parameter sub-region through spatial adjacency relationship is determined, and the parameter adjustment range is generated.

[0098] Extract the corresponding alignment residuals from the target spatial grid region and associate the alignment residuals with the target parameter sub-region; perform directional correction on the model parameters within the target parameter sub-region based on the associated alignment residuals to generate the parameter adjustment results for the target parameter sub-region;

[0099] Based on the parameter adjustment range and the parameter adjustment results of the target parameter sub-region, parameter propagation boundaries are set between each parameter sub-region within the parameter adjustment range to restrict the propagation path of the parameter adjustment results and generate restricted propagation results;

[0100] Based on the limited propagation results, the adjustment amount transmitted to each parameter sub-region is progressively reduced according to the spatial distance between each parameter sub-region within the parameter adjustment range and the target parameter sub-region, thereby generating the incremental adjustment result corresponding to each parameter sub-region.

[0101] Based on the incremental adjustment results corresponding to each parameter sub-region, the alignment model is updated to generate the incrementally updated alignment model.

[0102] In this embodiment of the invention, by mapping the target spatial grid region to the corresponding parameter sub-region and determining the parameter adjustment range based on spatial adjacency, the model parameter adjustment is extended from a single region to related regions connected to it, thereby forming an adjustment range with spatial continuity. Based on this, the alignment residual is associated with the target parameter sub-region and directionally corrected, giving the model parameter adjustment a clear direction for the source of error. Furthermore, a parameter propagation boundary is set within the parameter adjustment range, and the propagation path is constrained, ensuring that parameter adjustment is only transmitted along predetermined spatial adjacency paths, avoiding the influence of irrelevant regions. Simultaneously, by progressively decreasing the allocation according to spatial distance, parameter sub-regions far from the target region receive smaller adjustments, thus achieving a balance between local correction and overall smoothing. Finally, by incrementally updating each parameter sub-region, the model can gradually eliminate local deviations and maintain overall stability during continuous processing.

[0103] In a preferred embodiment of the present invention, determining the target parameter sub-region corresponding to the target spatial grid region based on the position identifier of the target spatial grid region in the alignment model includes:

[0104] Based on the identification information of each grid location in the target spatial grid region, extract the set of all grid locations contained in the target spatial grid region to generate the target grid location set;

[0105] Based on the spatial location identifier mapping relationship established during the model building phase, the parameter sub-region identifiers that correspond one-to-one with the target grid location set are found in the alignment model to generate the target parameter identifier set;

[0106] Based on the target parameter identifier set, extract the corresponding model parameter set from the parameter storage structure of the alignment model to generate the target parameter sub-region;

[0107] Based on the structural location of the target parameter sub-region in the model, it is marked so that subsequent parameter adjustment operations can locate the target parameter sub-region and generate parameter location results.

[0108] In a preferred embodiment of the present invention, based on the spatial adjacency relationship of the target parameter sub-region, a set of parameter sub-regions connected to the target parameter sub-region through spatial adjacency is determined, and a parameter adjustment range is generated, including:

[0109] Based on the spatial location identifier of the target parameter sub-region, determine its node position in the topology of the parameter sub-region and generate the target node position result;

[0110] Based on the target node location results, find the adjacent nodes that have a direct connection relationship with it in the parameter sub-region topology and generate a set of adjacent parameter sub-regions;

[0111] Based on the set of adjacent parameter sub-regions, an extended search is performed along the connection paths in the topology to identify parameter sub-regions that are connected to the target parameter sub-region through continuous connection paths, and a set of connected parameter sub-regions is generated.

[0112] Based on the set of connected parameter sub-regions, determine the range of all parameter sub-regions involved in parameter adjustment, and generate the parameter adjustment range.

[0113] In a preferred embodiment of the present invention, the model parameters within the target parameter sub-region are directionally corrected based on the correlated alignment residuals to generate parameter adjustment results for the target parameter sub-region, including:

[0114] Based on the alignment residuals corresponding to the target spatial grid region, the residual information is mapped to each parameter unit in the target parameter sub-region according to its spatial location, generating parameter residual distribution results;

[0115] Based on the parameter residual distribution results, determine the direction of influence of each parameter unit on the alignment error, and determine the parameter adjustment direction based on the direction of influence, so that the model output tends to reduce the alignment residual;

[0116] Based on the parameter adjustment direction, the model parameters within the target parameter sub-region are adjusted. The adjustment is achieved by increasing or decreasing the current value of the parameter, generating parameter correction results.

[0117] Based on the parameter correction results, all parameters within the target parameter sub-region are updated uniformly to generate the parameter adjustment results for the target parameter sub-region.

[0118] In a preferred embodiment of the present invention, the alignment model is updated based on the incremental adjustment results corresponding to each parameter sub-region to generate an incrementally updated alignment model, including:

[0119] Based on the incremental adjustment results corresponding to each parameter sub-region, the update amount of the model parameters in each parameter sub-region is obtained, and a set of parameter update information is generated.

[0120] Based on the parameter update information set, the corresponding parameter sub-regions in the alignment model are replaced and updated so that the model parameters in each parameter sub-region are updated to the adjusted parameter values, thus generating a locally updated model.

[0121] Based on the local update model, the parameters of the parameter sub-regions that are not involved in the update are kept unchanged, thereby completing the overall parameter update of the model and generating a complete updated model;

[0122] Based on the complete updated model, the data processing path of the model is checked for consistency to ensure that the updated model structure is consistent with the original topology, and the incrementally updated alignment model is generated.

[0123] In a preferred embodiment of the present invention, based on the correspondence between grid positions and local data processing ranges, the model parameters involved in the data processing of each local data processing range in the initial alignment model are divided to generate parameter sets corresponding to each grid position, including:

[0124] Based on the correspondence between grid positions and local data processing ranges, the fixed data processing paths for each local data processing range in the initial alignment model are determined, and data processing paths corresponding one-to-one with each grid position are generated.

[0125] Based on the data processing path, extract the model parameters that participate in the calculation in the corresponding data processing path in the initial alignment model, and classify them according to the data processing path to generate parameter classification results corresponding to each data processing path;

[0126] Based on the parameter classification results, each parameter classification result is independently encapsulated so that each parameter classification result corresponds to only one grid position, generating parameter sets corresponding to each grid position respectively;

[0127] Each parameter set is bound and labeled so that the correspondence between each parameter set and the corresponding grid position remains unchanged during the model update process.

[0128] In this embodiment of the invention, by determining a fixed data processing path for each local data processing range within the initial alignment model, data from different spatial grid locations have a stable and consistent data flow direction within the model, thereby avoiding path differences for data at the same location in different calculation processes. Model parameters are extracted and categorized based on the data processing path, so that parameters participating in the same path operation are centrally divided into corresponding parameter classification results, and further, through independent encapsulation, parameter sets corresponding one-to-one with grid locations are formed, thus giving model parameters a clear spatial affiliation. By binding and marking each parameter set, the correspondence between the parameter set and the grid location remains unchanged during model updates, ensuring that in subsequent incremental adjustments, model parameter adjustments always act on the corresponding spatial region, avoiding parameter drift across regions and improving the stability and controllability of model updates.

[0129] In a preferred embodiment of the present invention, based on the correspondence between grid positions and local data processing ranges, a fixed data processing path is determined for each local data processing range in the initial alignment model, and a data processing path corresponding one-to-one with each grid position is generated, including:

[0130] Based on the correspondence between grid positions and local data processing ranges, the corresponding data entry positions of each grid position in the model input area are read, and grid position input mapping results are generated.

[0131] Based on the grid location input mapping results, along the data transfer process connected to the corresponding data entry location in the initial alignment model, identify the processing units at each level involved in the data processing of the grid location, and generate a local data processing unit sequence.

[0132] Based on the local data processing unit sequence, the complete data transfer process from the input region to the output result is determined, and this complete data transfer process is defined as a fixed data processing path at the corresponding grid position, generating the data processing path result;

[0133] Based on the data processing path results, the data processing path corresponding to each grid location is recorded, so that the same data processing path is always called in the subsequent model update process for the same grid location, generating a data processing path that corresponds one-to-one with each grid location.

[0134] In a preferred embodiment of the present invention, model parameters participating in the calculation in the corresponding data processing path of the initial alignment model are extracted according to the data processing path, and categorized according to the data processing path to generate parameter classification results corresponding to each data processing path, including:

[0135] Based on the data processing path, determine the processing layer and processing node that participate in the corresponding path data operation in the initial alignment model, and generate the path participation unit result;

[0136] Based on the path participation unit results, model parameters connected to each processing layer and processing node are extracted. The model parameters include weight parameters and bias parameters used to complete the corresponding path data processing, and path parameter extraction results are generated.

[0137] Based on the path parameter extraction results, the data processing paths are classified and organized according to their grid positions, so that model parameters in the same path are grouped into the same parameter category, generating parameter classification results.

[0138] Based on the parameter classification results, establish the association between each parameter category and the corresponding data processing path, and generate parameter classification results corresponding to each data processing path.

[0139] In a preferred embodiment of the present invention, based on the parameter classification results, each parameter classification result is independently encapsulated so that each parameter classification result corresponds to only one grid position, generating parameter sets corresponding to each grid position, including:

[0140] Based on the parameter classification results, the model parameter sets belonging to the same data processing path are divided into an independent parameter unit, generating parameter unit division results;

[0141] Based on the parameter unit partitioning results, each independent parameter unit is encapsulated to give each independent parameter unit an independent storage identifier and a call identifier, thereby generating the encapsulated parameter unit results.

[0142] Based on the encapsulated parameter cell results, check whether each independent parameter cell contains model parameters from data processing paths corresponding to other grid locations, and remove data processing path parameters that do not belong to the current grid location to generate parameter cell results for a single grid.

[0143] Based on the parameter element results corresponding to a single grid, each independent parameter element is determined as a parameter set corresponding to each grid position, generating parameter sets corresponding to each grid position.

[0144] In a preferred embodiment of the present invention, each parameter set is bound and marked so that the correspondence between each parameter set and the corresponding grid position remains unchanged during the model update process, including:

[0145] Based on the parameter sets corresponding to each grid position, write a grid position identifier that is consistent with the corresponding grid position for each parameter set, and generate a parameter set identifier result;

[0146] Based on the parameter set identification results, a binding relationship table between the parameter set and the mesh position is established in the model parameter management structure, and the parameter binding relationship results are generated;

[0147] Based on the parameter binding relationship results, when calling the parameter set during the model update process, the corresponding grid position identifier is read first, and then the scope of the parameter set is limited according to the parameter binding relationship table to prevent the parameter set from being called cross-referenced between different grid positions, and to generate binding constraint results.

[0148] Based on the binding constraint results, after each model parameter update, the consistency between the updated parameter set identifier and the grid position identifier is verified to ensure that the correspondence between each parameter set and the corresponding grid position remains unchanged.

[0149] In a preferred embodiment of the present invention, based on the parameter set, position identification processing is performed on each parameter set, and the spatial adjacency relationship between each grid position is mapped to the spatial adjacency relationship between each parameter set, generating a parameter sub-region with spatial position identification and spatial adjacency relationship, including:

[0150] Based on the parameter set, assign a spatial coordinate identifier consistent with the corresponding grid position to each parameter set, and generate a parameter set with spatial coordinate identifier;

[0151] Based on the connection relationships between grid positions in the processing area, a grid topology is constructed, and the direct adjacency relationships between each grid position are determined according to the grid topology, generating grid adjacency relationship results;

[0152] Based on the parameter set with spatial coordinate identifiers and the grid adjacency relationship results, the direct adjacency relationship between each grid position is mapped to the fixed adjacency relationship between each parameter set, generating the parameter set adjacency relationship result;

[0153] Based on the parameter sets with spatial coordinate identifiers and the parameter set adjacency results, each parameter set is organized to generate parameter sub-regions with spatial location identifiers and fixed topological adjacency relationships.

[0154] In this embodiment of the invention, by assigning each parameter set a spatial coordinate identifier consistent with the grid position, the model parameters structurally possess clear spatial positioning information, thereby achieving a direct mapping relationship between the parameters and the processing area. Based on this, by constructing a grid topology and determining the connection relationships between each grid position, the spatial structure of the processing area is transformed into a computable topological relationship, and this topological relationship is further mapped to the parameter sets, forming fixed parameter set adjacency relationships. By organizing the parameter sets with spatial coordinate identifiers and their adjacency relationships, the model parameters structurally form parameter sub-regions with spatial position identifiers and topological connection relationships, thereby providing a clear path basis for subsequent parameter propagation based on adjacency relationships, and enabling the parameter adjustment process to have traceable propagation direction and range constraints.

[0155] In a preferred embodiment of the present invention, each parameter set is assigned a spatial coordinate identifier consistent with the corresponding grid position, based on the parameter set, thereby generating a parameter set with spatial coordinate identifiers, including:

[0156] Based on the grid position identifiers corresponding to the parameter sets, read the corresponding grid positions of each parameter set in the processing area and generate the parameter set position correspondence results;

[0157] Based on the position correspondence results of the parameter set, extract the coordinate information of each grid position in the spatial grid mapping result. The coordinate information includes the horizontal and vertical positions of the grid position in the processing area, and generate the grid coordinate result.

[0158] Based on the grid coordinate results, the corresponding spatial coordinate information is written into the identifier field of each parameter set, so that each parameter set has a unique spatial location identifier, and a parameter set with spatial coordinate identifier is generated.

[0159] Based on the parameter set with spatial coordinate identifiers, a consistency check is performed on the identifier content of the parameter set to ensure that the spatial coordinate identifier of each parameter set is consistent with the corresponding grid position.

[0160] In a preferred embodiment of the present invention, a mesh topology is constructed based on the connection relationships between mesh positions in the processing area, and the direct adjacency relationships between each mesh position are determined based on the mesh topology to generate a mesh adjacency relationship result, including:

[0161] Based on the grid division results of the processing area, extract all grid positions and the spatial distribution relationship between each grid position to generate grid distribution results;

[0162] Based on the grid distribution results, identify directly adjacent grid positions in the horizontal or vertical direction, establish connection relationships between directly adjacent grid positions, and generate grid connection results;

[0163] Based on the mesh connection results, all mesh locations and their connection relationships are organized to construct a mesh topology structure that reflects the interconnection status of each mesh location, and generate mesh topology results;

[0164] Based on the grid topology results, the directly connected adjacent grid locations of each grid location are extracted to generate the direct adjacency relationships between each grid location, and a grid adjacency relationship result is formed.

[0165] In a preferred embodiment of the present invention, based on the parameter set with spatial coordinate identifiers and the grid adjacency relationship results, the direct adjacency relationship between each grid position is mapped to the fixed adjacency relationship between each parameter set, generating the parameter set adjacency relationship result, including:

[0166] Based on the parameter set with spatial coordinate identifiers, extract the spatial coordinate identifiers corresponding to each parameter set and generate the parameter set coordinate results;

[0167] Based on the parameter set coordinate results and the grid adjacency relationship results, find the parameter set pairs that correspond to the direct adjacency relationships of each grid position, and generate the parameter set correspondence results;

[0168] Based on the parameter set correspondence results, the direct adjacency relationship between each grid position is written into the connection identifier between the corresponding parameter sets, so that the corresponding parameter sets form a fixed adjacency relationship consistent with the grid adjacency relationship, and the parameter set adjacency relationship result is generated;

[0169] Based on the adjacency results of the parameter sets, the adjacency directions and connection boundaries between the parameter sets are recorded, so that the fixed adjacency relationships between the parameter sets remain unchanged during the model update process.

[0170] In a preferred embodiment of the present invention, based on the parameter sets with spatial coordinate identifiers and the parameter set adjacency relationship results, each parameter set is organized to generate a parameter sub-region with spatial location identifiers and fixed topological adjacency relationships, including:

[0171] Based on the parameter sets with spatial coordinate identifiers, sort each parameter set according to the spatial coordinate identifiers to generate a spatial arrangement result of the parameter sets;

[0172] Based on the spatial arrangement result and the adjacency relationship result of the parameter set, the parameter sets with fixed adjacency relationships are organized according to the spatial connection order to generate the parameter set topology organization result;

[0173] Based on the topology organization results of the parameter sets, a corresponding parameter sub-region identifier is assigned to each parameter set, and parameter sets with the same spatial location attribute and fixed connection attribute are identified as the corresponding parameter sub-regions, generating parameter sub-region partitioning results;

[0174] Based on the parameter sub-region division results, establish spatial location identifiers and fixed topological adjacency descriptions for the parameter sub-regions, and generate parameter sub-regions with spatial location identifiers and fixed topological adjacency relationships.

[0175] In a preferred embodiment of the present invention, based on the set of abnormal grid locations, connectivity merging is performed on spatially adjacent abnormal grid locations to generate candidate spatial grid regions, including:

[0176] Establish a grid topology based on the spatial grid mapping results;

[0177] Based on the set of abnormal grid locations, determine the node position of each abnormal grid location in the grid topology and generate abnormal node position results;

[0178] Based on the results of abnormal node locations, and according to the connection relationship between nodes in the grid topology, abnormal grid location pairs with continuous connection paths are identified, and a path connection set is generated.

[0179] Based on the path connection set, the abnormal grid positions that are interconnected by continuous connection paths are merged to generate multiple connected node sets;

[0180] Based on the set of connected nodes, determine the region boundary corresponding to each set of connected nodes and generate candidate spatial grid regions.

[0181] In this embodiment of the invention, a grid topology is established based on the spatial grid mapping results, clearly expressing the connection relationships between grid locations, thus providing a structural basis for anomaly region identification. Based on this, anomaly grid locations are mapped to node locations in the topology, and pairs of anomaly grid locations with continuous connection paths are identified through the connection relationships between nodes, revealing the spatial correlation between discrete anomaly points. Furthermore, by merging anomaly grid locations with connection paths, multiple connected node sets are formed, thereby integrating scattered anomalies into a spatially continuous regional structure. Finally, by determining the regional boundaries of each connected node set, the candidate spatial grid region has a clear scope definition, providing a stable regional basis for subsequent offset region determination and avoiding inaccurate region division due to discrete anomalies.

[0182] In a preferred embodiment of the present invention, establishing a grid topology based on spatial grid mapping results includes:

[0183] Based on the spatial grid mapping results, extract the spatial distribution information of all grid positions in the processing area to generate grid position distribution results;

[0184] Based on the grid location distribution results, the arrangement order and relative positional relationship of each grid location in the processing area are read to generate the basic results of grid location connection;

[0185] Based on the grid position connection results, node connection relationships will be established between adjacent grid positions in the horizontal or vertical directions to generate grid connection results;

[0186] Based on the mesh connection results, all mesh locations and their connection relationships are uniformly organized to establish a mesh topology structure that characterizes the connection status between each mesh location.

[0187] In a preferred embodiment of the present invention, the node position of each abnormal grid location in the grid topology is determined based on the abnormal grid location set, and abnormal node position results are generated, including:

[0188] Based on the set of abnormal grid locations, extract the grid location identifiers corresponding to each abnormal grid location to generate abnormal grid identifier results;

[0189] Based on the abnormal grid identification results, find the nodes corresponding to the location identifiers of each abnormal grid in the grid topology and generate node matching results;

[0190] Based on the node matching results, determine the specific node position of each abnormal grid location in the grid topology, record the distribution relationship between each node position, and generate abnormal node position results;

[0191] Based on the results of the abnormal node locations, the distribution status of the abnormal nodes in the grid topology is marked to provide basic node information for subsequent path identification.

[0192] In a preferred embodiment of the present invention, based on the abnormal node location results and the connection relationships between nodes in the mesh topology, abnormal mesh location pairs with continuous connection paths are identified, and a path connection set is generated, including:

[0193] Based on the abnormal node location results, read the node connection information of each abnormal node in the grid topology and generate the abnormal node connection results.

[0194] Based on the abnormal node connection results, check whether any two abnormal nodes can be connected through a continuous node connection relationship, and generate a node connectivity determination result.

[0195] Based on the node connectivity determination results, abnormal nodes with continuous connection paths are recorded in pairs to generate abnormal grid location pairs.

[0196] Based on the results of abnormal grid location pairs, the abnormal grid location pairs with continuous connection paths are summarized to generate a path connection set.

[0197] In a preferred embodiment of the present invention, based on the path connection set, abnormal grid positions that are interconnected by continuous connection paths are merged to generate multiple connected node sets, including:

[0198] Based on the path connection set, extract the connection relationships between each abnormal grid location pair and generate the abnormal location connection results;

[0199] Based on the connection results of abnormal locations, abnormal grid locations with common connection paths are grouped into the same connected group to generate connected group results;

[0200] Based on the connectivity grouping results, the abnormal grid positions in each connectivity group are uniformly organized so that each connectivity group contains only abnormal grid positions that are connected to each other through continuous connection paths, generating multiple sets of connected nodes.

[0201] Based on multiple sets of connected nodes, the independence between different sets of connected nodes is verified, so that each set of connected nodes is separated from the others and there is no cross connection relationship.

[0202] In a preferred embodiment of the present invention, determining the region boundaries corresponding to each set of connected nodes and generating candidate spatial grid regions based on the set of connected nodes includes:

[0203] Based on the set of connected nodes, extract the spatial distribution range of all abnormal grid locations in each set of connected nodes to generate the connected region distribution result;

[0204] Based on the distribution results of connected regions, determine the boundary positions of each set of connected nodes in the horizontal and vertical directions, and generate the region boundary results;

[0205] Based on the region boundary results, the grid range covered by each set of connected nodes is determined as the corresponding region boundary range, and the region range results are generated.

[0206] Based on the regional range results, the regional boundary range corresponding to each set of connected nodes is determined as the candidate spatial grid region, and the candidate spatial grid region is generated.

[0207] In a preferred embodiment of the present invention, for each candidate spatial grid region, residual change information corresponding to the candidate spatial grid region within multiple consecutive processing cycles is extracted, and the continuity of the residual change information is determined to generate a continuous offset region determination result, including:

[0208] Based on the candidate spatial grid region, determine the corresponding region position of the candidate spatial grid region in multiple consecutive processing cycles, and generate region position tracking results;

[0209] Based on the regional location tracking results, the residual records of the candidate spatial grid region in multiple consecutive processing cycles are extracted, and a residual change sequence arranged in chronological order is generated.

[0210] Based on the residual change sequence, the direction of residual change between adjacent processing cycles is compared one by one to determine the processing cycle sequence that satisfies the continuous and consistent change relationship, and a continuity determination result is generated.

[0211] Based on the continuity determination results, candidate spatial grid regions that satisfy the continuous and consistent change relationship are marked, and continuous offset region determination results are generated.

[0212] In this embodiment of the invention, by tracking the position of candidate spatial grid regions across multiple consecutive processing cycles, the position of the same spatial region remains consistently identified across different processing stages, thus ensuring the temporal comparability of residual data. Based on this, the residual records in each processing cycle are organized chronologically into a residual change sequence, allowing for a continuous representation of residual change trends. By comparing the residual change directions between adjacent processing cycles one by one, processing cycle sequences that satisfy a continuous and consistent change relationship are identified, thereby distinguishing regions with stable change trends from randomly fluctuating regions. Finally, candidate spatial grid regions that satisfy a continuous and consistent change relationship are marked, ensuring that the determination of target spatial grid regions is based on continuous change characteristics, thereby improving the stability of offset region identification and avoiding misjudgments due to single processing anomalies.

[0213] In a preferred embodiment of the present invention, the definition of the processing cycle includes:

[0214] A processing cycle is defined as the process by which a laser direct writing device performs a complete alignment, exposure scan, and processing result detection on the same processing area. The start time of the processing cycle is the moment when the alignment mark image of the current processing area begins to be acquired, and the end time is the moment when the processing result data corresponding to the current processing area is acquired and stored.

[0215] For continuous processing scenarios, the multiple complete processing processes completed sequentially by the laser direct writing device according to the preset cycle are defined as multiple processing cycles, so that each processing cycle corresponds to a set of independent alignment mark images, laser scanning path data, alignment compensation results and processing result data.

[0216] For scenarios involving regional processing, the process of completing a full alignment, scanning, and result detection for each region on the same substrate is defined as an independent processing cycle, so that different regions correspond to different processing cycles and generate corresponding residual records respectively.

[0217] Each processing cycle is sequentially numbered, and a processing cycle sequence is established according to the time sequence, so that subsequent residual change analysis can be carried out on the basis of a unified time sequence;

[0218] In the continuity determination process, only processing processes with complete data acquisition, consistent data correlation, and corresponding to the same type of processing object are included in the processing cycle sequence to ensure that the residual changes between adjacent processing cycles are comparable.

[0219] In a preferred embodiment of the present invention, based on the residual change sequence, the residual change directions between adjacent processing cycles are compared one by one to determine the processing cycle sequence that satisfies a continuous and consistent change relationship, and a continuity determination result is generated, including:

[0220] Based on the residual records of the candidate spatial grid region in multiple consecutive processing cycles, the residual records are arranged in chronological order of the processing cycles to generate a residual change sequence.

[0221] Based on the residual change sequence, the residual records of the previous processing cycle and the next processing cycle are read sequentially, and the change trends of the two are compared to determine whether the next processing cycle belongs to one of the following directions of change: residual increase, residual decrease, or residual unchanged, and the result of the residual change direction between adjacent processing cycles is generated.

[0222] Based on the residual change direction, adjacent processing cycles with the same residual change direction are continuously recorded, so that the part that maintains the same change direction between multiple consecutive adjacent processing cycles forms a continuous change segment, and the result of continuous change segment of processing cycle is generated.

[0223] Based on the results of the continuous change segment of the processing cycle, the processing cycle sequence range contained in the continuous change segment is extracted, and the processing cycle sequence range is determined as the processing cycle sequence that satisfies the continuous and consistent change relationship, thus generating the processing cycle sequence result;

[0224] Based on the processing cycle sequence results, determine whether the candidate spatial grid region meets the preset continuous offset judgment condition; when the processing cycle sequence results meet the preset continuous offset judgment condition, generate the continuity judgment result of the corresponding candidate spatial grid region.

[0225] In a preferred embodiment of the present invention, based on the parameter adjustment range and the parameter adjustment results of the target parameter sub-region, parameter propagation boundaries are set between the parameter sub-regions within the parameter adjustment range to restrict the propagation path of the parameter adjustment results, generating restricted propagation results, including:

[0226] The parameter sub-region topology is constructed based on the fixed adjacency relationships between parameter sub-regions determined by the spatial grid mapping relationship;

[0227] Based on the parameter adjustment range, determine the node position of the target parameter sub-region in the parameter sub-region topology, and based on the parameter sub-region topology, determine the set of parameter sub-regions connected to the target parameter sub-region through spatial adjacency within the parameter adjustment range, and generate the connection relationship result;

[0228] Based on the connection relationship results and the topology of the parameter sub-regions, determine the parameter boundary positions between each parameter sub-region within the parameter adjustment range, and generate the parameter propagation boundary;

[0229] Based on the parameter propagation boundary and connection relationship results, a set of parameter propagation paths is established starting from the target parameter sub-region and extending along the spatial adjacency relationship, and parameter propagation path results are generated;

[0230] Based on the parameter propagation path results, path constraints are applied to the parameter adjustment results of the target parameter sub-region, ensuring that the parameter adjustment results are only transmitted along the paths in the parameter propagation path results, and the adjustment amounts exceeding the parameter propagation boundary are truncated to generate restricted propagation results.

[0231] In this embodiment of the invention, a parameter sub-region topology is constructed based on the fixed adjacency relationships between parameter sub-regions, giving the spatial connection relationships between model parameters a clear structural expression, thus providing a computable path basis for parameter propagation. Based on this, the parameter adjustment range is mapped onto the topology, and the node positions of the target parameter sub-region and the set of parameter sub-regions connected to it are determined, giving the propagation range a clear structural boundary. The parameter propagation boundary is formed by determining the boundary positions between parameter sub-regions, giving the connection relationships between different parameter sub-regions a clear transmission limit. Combined with the propagation path set, path constraints are applied to the parameter adjustment results, ensuring that parameters are propagated only along predetermined paths. Simultaneously, by truncating adjustments exceeding the propagation boundary, parameter adjustments are prevented from spreading to irrelevant regions, thereby achieving dual control over the propagation range and propagation path, improving the controllability of the parameter update process.

[0232] In a preferred embodiment of the present invention, constructing a parameter sub-region topology based on the fixed adjacency relationship between parameter sub-regions determined by the spatial grid mapping relationship includes:

[0233] Based on the spatial grid mapping relationship, extract the mapping relationship between each grid position and the corresponding parameter sub-region, and generate the result corresponding to the parameter sub-region position;

[0234] Based on the position correspondence of the parameter sub-regions, the spatial arrangement relationship of each parameter sub-region in the processing area is read, and the fixed adjacency relationship between each parameter sub-region that is consistent with the mesh adjacency relationship is extracted to generate the parameter sub-region adjacency result.

[0235] Based on the parameter sub-region adjacency results, each parameter sub-region is defined as a topology node, and node connection relationships are established between parameter sub-regions with fixed adjacency relationships to generate parameter sub-region connection results.

[0236] Based on the connection results of the parameter sub-regions, all parameter sub-regions and their node connection relationships are uniformly organized to generate a parameter sub-region topology structure that represents the fixed connection state between each parameter sub-region.

[0237] In a preferred embodiment of the present invention, the node positions of the target parameter sub-region in the parameter sub-region topology are determined according to the parameter adjustment range, and the set of parameter sub-regions connected to the target parameter sub-region through spatial adjacency within the parameter adjustment range is determined based on the parameter sub-region topology, generating a connection relationship result, including:

[0238] Based on the parameter adjustment range, extract the identifiers of all parameter sub-regions involved in this parameter propagation control, and generate the parameter adjustment range results;

[0239] Based on the adjusted range parameters, the node positions corresponding to the target parameter sub-region are found in the parameter sub-region topology, and the target node positioning results are generated.

[0240] Based on the target node localization result, along the node connection relationship in the parameter sub-region topology, within the parameter adjustment range, identify the parameter sub-regions connected to the target parameter sub-region through a fixed adjacency relationship, and generate the connected parameter sub-region identification result;

[0241] Based on the identification results of the connected parameter sub-regions, a record of the connection relationships between the target parameter sub-region and each connected parameter sub-region is established, and the connection relationship results are generated.

[0242] In a preferred embodiment of the present invention, the parameter boundary positions between each parameter sub-region within the parameter adjustment range are determined based on the connection relationship results and the parameter sub-region topology, and a parameter propagation boundary is generated, including:

[0243] Based on the connection relationship results, extract the connection positions between any two adjacent parameter sub-regions within the parameter adjustment range, and generate the adjacent connection position results;

[0244] Based on the results of adjacent connection positions, and combined with the connection directions between nodes in the topology of the parameter sub-region, the contact area of ​​each adjacent parameter sub-region in the model parameter organization structure is determined, and the parameter contact area results are generated.

[0245] Based on the parameter contact area results, the edge positions where parameter propagation occurs between each adjacent parameter sub-region are extracted, and the edge positions are defined as parameter boundary positions to generate parameter boundary position results.

[0246] Based on the results of the parameter boundary positions, the boundary positions of each parameter are recorded uniformly to form the parameter propagation boundary within the parameter adjustment range.

[0247] In a preferred embodiment of the present invention, based on the parameter propagation boundary and connection relationship results, a set of parameter propagation paths is established starting from the target parameter sub-region and extending along spatial adjacency relationships, generating parameter propagation path results, including:

[0248] Based on the parameter propagation boundary, determine the boundary position of the adjacent parameter sub-regions that the parameter adjustment result can pass through, and generate the boundary passage result;

[0249] Based on the boundary passage results and connection relationship results, starting from the target parameter sub-region, the reachable parameter sub-regions are searched step by step along the fixed adjacency connection direction between parameter sub-regions to generate the propagation reachability results.

[0250] Based on the propagation reachability results, the node connection order that starts from the target parameter sub-region and passes through the boundary positions of adjacent parameter sub-regions is determined as the parameter propagation path, and a single parameter propagation path result is generated.

[0251] Based on the results of a single parameter propagation path, all paths that meet the propagation conditions are summarized to generate a parameter propagation path set, and a parameter propagation path result is formed.

[0252] In a preferred embodiment of the present invention, based on the parameter propagation path results, path constraints are applied to the parameter adjustment results of the target parameter sub-region, ensuring that the parameter adjustment results are transmitted only along the paths in the parameter propagation path results, and truncation is performed on adjustment amounts exceeding the parameter propagation boundaries to generate restricted propagation results, including:

[0253] Based on the parameter propagation path results, read the allowed propagation direction of the parameter adjustment results in the target parameter sub-region within the parameter adjustment range, and generate the allowed propagation direction results;

[0254] Based on the allowed propagation direction results, the parameter adjustment results of the target parameter sub-region are allocated, so that the parameter adjustment results are only transmitted to the adjacent parameter sub-regions along the path direction recorded in the parameter propagation path results, and the path constraint propagation results are generated.

[0255] Based on the path constraint propagation results, check whether the parameter adjustment results exceed the parameter propagation boundary during the transmission process. For the adjustment part that exceeds the parameter propagation boundary, perform a stop transmission process and generate a boundary truncation result.

[0256] Based on the boundary truncation results, the parameter adjustments retained within the parameter propagation path range are summarized to generate restricted propagation results.

[0257] In a preferred embodiment of the present invention, based on the limited propagation results, the adjustment amount transmitted to each parameter sub-region is progressively reduced according to the spatial distance between each parameter sub-region within the parameter adjustment range and the target parameter sub-region, generating the incremental adjustment result corresponding to each parameter sub-region, including:

[0258] Based on the limited propagation results, the spatial distance relationship between each parameter sub-region within the parameter adjustment range and the target parameter sub-region is determined, and the parameter sub-region within the parameter adjustment range is hierarchically divided according to the spatial distance relationship to generate parameter sub-region hierarchical results;

[0259] Based on the parameter sub-region hierarchy results, the adjustment amount transmitted outward from the target parameter sub-region is distributed level by level, so that the adjustment amount received by the parameter sub-region at a higher spatial distance is less than the adjustment amount received by the parameter sub-region at a lower spatial distance, thus generating a decreasing distribution result.

[0260] Based on the decreasing allocation result, it is determined whether the adjustment amount received by each level parameter sub-region meets the preset termination condition. When the adjustment amount corresponding to the subsequent level meets the preset termination condition, the transmission of adjustment amount to the subsequent level and the level after that is stopped, and the decreasing allocation result after termination control is generated.

[0261] Based on the decreasing allocation results after termination control, the model parameters of each parameter sub-region within the parameter adjustment range are updated, generating the incremental adjustment results corresponding to each parameter sub-region.

[0262] In this embodiment of the invention, by hierarchically dividing the parameter propagation process according to the spatial distance relationship between each parameter sub-region within the parameter adjustment range and the target parameter sub-region, a hierarchical structure from near to far is created, providing an orderly basis for the allocation of adjustment amounts. Based on this, by allocating adjustment amounts level by level, parameter sub-regions closer to the target parameter sub-region receive larger adjustment amounts, while those farther away receive smaller adjustment amounts, thus forming an adjustment distribution that matches the spatial distance. Furthermore, by setting a termination condition and stopping propagation when the adjustment amount at subsequent levels meets the termination condition, the adjustment process automatically ends after reaching a predetermined range, preventing the propagation range from continuously expanding. Finally, by updating each parameter sub-region, the parameter adjustment can cover the relevant area while avoiding impact on parts far from the target area, thereby achieving coordination between local correction and overall stability.

[0263] In a preferred embodiment of the present invention, based on the restricted propagation results, the spatial distance relationship between each parameter sub-region within the parameter adjustment range and the target parameter sub-region is determined, and the parameter sub-regions within the parameter adjustment range are hierarchically divided according to the spatial distance relationship to generate a parameter sub-region hierarchy result, including:

[0264] Based on the limited propagation results, extract the identification information of all parameter sub-regions within the parameter adjustment range and the identification information of the target parameter sub-region to generate parameter sub-region identification results;

[0265] Based on the parameter sub-region identification results, the node positions of each parameter sub-region in the parameter sub-region topology and their connection paths with the target parameter sub-region are read to generate node position relationship results.

[0266] Based on the node position relationship results, the number of adjacent connections traversed between each parameter sub-region and the target parameter sub-region is counted, and the number of adjacent connections traversed is determined as the spatial distance relationship between the corresponding parameter sub-region and the target parameter sub-region, generating the spatial distance relationship result;

[0267] Based on the spatial distance relationship, the parameter sub-regions directly adjacent to the target parameter sub-region are divided into the first level, the parameter sub-regions connected to the target parameter sub-region through an intermediate parameter sub-region are divided into the second level, and the remaining parameter sub-regions are divided in the same way to generate the parameter sub-region hierarchy result.

[0268] In a preferred embodiment of the present invention, based on the parameter sub-region hierarchy results, the adjustment amount transmitted outward from the target parameter sub-region is allocated level by level, such that the adjustment amount received by the parameter sub-region at a more distant spatial level is less than the adjustment amount received by the parameter sub-region at a closer spatial level, generating a decreasing allocation result, including:

[0269] Based on the parameter sub-region hierarchy results, read the hierarchy order between each level parameter sub-region and the target parameter sub-region, and generate the hierarchy order results;

[0270] Based on the hierarchical order results, the parameter adjustment results of the target parameter sub-region are used as the initial adjustment amount, and the initial adjustment amount is allocated to the first-level parameter sub-region adjacent to the target parameter sub-region to generate the first-level allocation result;

[0271] Based on the first-level allocation result, the adjustment amount received by the subsequent level parameter sub-regions is reduced in order from near to far, so that the adjustment amount received by each subsequent level parameter sub-region is less than the adjustment amount received by the previous level parameter sub-region, thus generating the adjustment amount allocation result for each level.

[0272] Based on the adjustment allocation results at each level, the corresponding adjustment information is written to each parameter sub-region within the same level to generate a decreasing allocation result.

[0273] In a preferred embodiment of the present invention, the method for determining the preset termination condition includes:

[0274] Obtain the adjustment records received by each level of parameter sub-region during parameter propagation in multiple historical processing cycles, as well as the corresponding model update results, and generate a historical propagation sample set;

[0275] Based on the historical propagation sample set, the impact of different level parameter sub-regions on the model output changes when receiving different adjustment amounts is statistically analyzed, and the adjustment amount impact results are generated.

[0276] Based on the results of the adjustment amount, the minimum adjustment amount range that can cause an effective change in the model output is determined, and the adjustment amount boundary that is below the minimum adjustment amount range and is insufficient to cause an effective change in the output is determined as the termination decision boundary, and the termination threshold result is generated.

[0277] Based on the termination threshold results and combined with the parameter propagation level requirements, it is determined that when the adjustment amount received by a parameter sub-region of a subsequent level reaches the termination judgment boundary, the subsequent level is judged to meet the preset termination condition. Thus, the preset termination condition is determined as a termination judgment condition jointly limited by the validity of the adjustment amount and the propagation level.

[0278] In a preferred embodiment of the present invention, based on the decreasing allocation result, it is determined whether the adjustment amount received by each level parameter sub-region meets the preset termination condition, and when the adjustment amount corresponding to the subsequent level meets the preset termination condition, the transmission of adjustment amount to the subsequent level and subsequent levels is stopped, generating a decreasing allocation result after termination control, including:

[0279] Based on the decreasing allocation results, extract the adjustment amount information received by each level parameter sub-region and generate the level adjustment amount results;

[0280] Based on the hierarchical adjustment results, the adjustment amounts received by each hierarchical parameter sub-region are compared sequentially with the preset termination conditions to generate a termination determination result.

[0281] Based on the termination determination result, when the adjustment amount received by a certain subsequent level parameter sub-region meets the preset termination condition, the adjustment amount is stopped from being transmitted to that subsequent level and subsequent levels, and the adjustment amount already received by each level parameter sub-region before termination is retained, and a termination control result is generated.

[0282] Based on the termination control results, the distribution of adjustment amounts for each level of parameter sub-regions after termination is organized to generate a decreasing allocation result after termination control.

[0283] In a preferred embodiment of the present invention, based on the decreasing allocation result after termination control, the model parameters of each parameter sub-region within the parameter adjustment range are updated to generate the incremental adjustment result corresponding to each parameter sub-region, including:

[0284] Based on the decreasing allocation result after termination control, extract the final adjustment amount corresponding to each parameter sub-region within the parameter adjustment range and generate the parameter update amount result;

[0285] Based on the parameter update results, the final adjustment received by each parameter sub-region is written into the corresponding model parameter update record to generate the parameter update record results;

[0286] Based on the parameter update record results, the model parameters of each parameter sub-region within the parameter adjustment range are updated so that the model parameters of each parameter sub-region are adjusted according to the corresponding final adjustment amount, and local parameter update results are generated.

[0287] Based on the local parameter update results, the update status of all parameter sub-regions within the parameter adjustment range is summarized to generate the incremental adjustment results corresponding to each parameter sub-region.

[0288] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An online incremental learning neural network optimization method for laser direct-write alignment models, characterized in that, The method includes: Acquire alignment mark images, laser scanning path data, and processing result data during the laser direct writing process; extract spatial coordinate information from the alignment mark images; divide the processing area into grids based on the spatial coordinate information, establish a spatial grid mapping relationship between each grid position and the local position of the laser scanning path, and generate spatial grid mapping results. An initial alignment model is established based on historical processing samples; the parameters of the initial alignment model are partitioned according to the spatial grid mapping results, so that the parameters in the initial alignment model are divided into multiple parameter sub-regions corresponding to different spatial grid positions, and spatial adjacency relationships are established for each parameter sub-region to generate an alignment model with spatial region correspondence. The alignment mark image and laser scanning path data in the current processing are input into the alignment model to obtain the alignment compensation results corresponding to each spatial grid position; the alignment compensation results are compared with the corresponding processing result data to obtain the alignment residual distribution results, and the target spatial grid area is determined based on the alignment residual distribution results. The target parameter sub-region is determined based on the target spatial grid region; the model parameters within the target parameter sub-region are incrementally adjusted, and parameter propagation boundaries are set between the parameter sub-regions to constrain the propagation path of the adjustment amount, so that the adjustment amount diffuses along the spatial adjacency relationship between the parameter sub-regions according to the preset decay rule, and the incrementally updated alignment model is generated. The alignment compensation results output by the incrementally updated alignment model are superimposed on the corresponding local positions in the laser scanning path, and the laser scanning trajectory is partitioned and corrected to obtain the updated laser scanning control results. Generate the incrementally updated alignment model, including: Based on the position identifier of the target spatial grid region in the alignment model, the target parameter sub-region corresponding to the target spatial grid region is determined, and based on the spatial adjacency relationship of the target parameter sub-region, the set of parameter sub-regions connected to the target parameter sub-region through spatial adjacency relationship is determined, and the parameter adjustment range is generated. Extract the corresponding alignment residuals from the target spatial grid region and associate the alignment residuals with the target parameter sub-region; perform directional correction on the model parameters within the target parameter sub-region based on the associated alignment residuals to generate the parameter adjustment results for the target parameter sub-region; Based on the parameter adjustment range and the parameter adjustment results of the target parameter sub-region, parameter propagation boundaries are set between each parameter sub-region within the parameter adjustment range to restrict the propagation path of the parameter adjustment results and generate restricted propagation results; Based on the limited propagation results, the adjustment amount transmitted to each parameter sub-region is progressively reduced according to the spatial distance between each parameter sub-region within the parameter adjustment range and the target parameter sub-region, thereby generating the incremental adjustment result corresponding to each parameter sub-region. Based on the incremental adjustment results corresponding to each parameter sub-region, the alignment model is updated to generate the incrementally updated alignment model.

2. The online incremental learning neural network optimization method for the laser direct-write alignment model according to claim 1, characterized in that, Based on the spatial grid mapping results, the initial alignment model is partitioned into multiple parameter sub-regions, each corresponding to a different spatial grid location. Spatial adjacency relationships are established between these sub-regions, generating an alignment model with spatial region correspondences, including: Based on the spatial grid mapping results, the input region of the initial alignment model is partitioned and calibrated to determine the correspondence between the grid positions in the processing region and the local data processing ranges in the input region, and to generate the correspondence results between the grid positions and the local data processing ranges. Based on the correspondence between grid positions and local data processing ranges, the model parameters involved in the data processing of each local data processing range in the initial alignment model are divided to generate parameter sets corresponding to each grid position. Based on the parameter set, the position identification of each parameter set is processed, and the spatial adjacency relationship between each grid position is mapped to the spatial adjacency relationship between each parameter set, generating parameter sub-regions with spatial position identification and spatial adjacency relationship; Based on parameter sub-regions with spatial location identifiers and spatial adjacency relationships, construct a positioning model with spatial region correspondence relationships.

3. The online incremental learning neural network optimization method for the laser direct-write alignment model according to claim 1, characterized in that, The target spatial grid region is determined based on the alignment residual distribution results, including: Based on the alignment residual distribution results, the residual information corresponding to each grid position is extracted, and the residual information corresponding to each grid position is judged to determine the abnormal grid positions that meet the preset residual judgment conditions, and an abnormal grid position set is generated. Based on the set of anomalous grid locations, connectivity is merged for spatially adjacent anomalous grid locations to generate candidate spatial grid regions; For each candidate spatial grid region, the residual change information corresponding to the candidate spatial grid region in multiple consecutive processing cycles is extracted, and the continuity of the residual change information is determined to generate the continuous offset region determination result. Based on the results of the continuous offset region determination, the candidate spatial grid regions that meet the preset continuous offset determination conditions are determined as the target spatial grid regions.

4. The online incremental learning neural network optimization method for the laser direct-write alignment model according to claim 2, characterized in that, Based on the correspondence between grid positions and local data processing ranges, the model parameters involved in data processing within each local data processing range of the initial alignment model are divided, generating parameter sets corresponding to each grid position, including: Based on the correspondence between grid positions and local data processing ranges, the fixed data processing paths for each local data processing range in the initial alignment model are determined, and data processing paths corresponding one-to-one with each grid position are generated. Based on the data processing path, extract the model parameters that participate in the calculation in the corresponding data processing path in the initial alignment model, and classify them according to the data processing path to generate parameter classification results corresponding to each data processing path; Based on the parameter classification results, each parameter classification result is independently encapsulated so that each parameter classification result corresponds to only one grid position, generating parameter sets corresponding to each grid position respectively; Each parameter set is bound and labeled so that the correspondence between each parameter set and the corresponding grid position remains unchanged during the model update process.

5. The online incremental learning neural network optimization method for the laser direct-write alignment model according to claim 2, characterized in that, Based on the parameter sets, each parameter set is marked with a location identifier, and the spatial adjacency relationships between each grid location are mapped to the spatial adjacency relationships between each parameter set, generating parameter sub-regions with spatial location identifiers and spatial adjacency relationships, including: Based on the parameter set, assign a spatial coordinate identifier consistent with the corresponding grid position to each parameter set, and generate a parameter set with spatial coordinate identifier; Based on the connection relationships between grid positions in the processing area, a grid topology is constructed, and the direct adjacency relationships between each grid position are determined according to the grid topology, generating grid adjacency relationship results; Based on the parameter set with spatial coordinate identifiers and the grid adjacency relationship results, the direct adjacency relationship between each grid position is mapped to the fixed adjacency relationship between each parameter set, generating the parameter set adjacency relationship result; Based on the parameter sets with spatial coordinate identifiers and the parameter set adjacency results, each parameter set is organized to generate parameter sub-regions with spatial location identifiers and fixed topological adjacency relationships.

6. The online incremental learning neural network optimization method for the laser direct-write alignment model according to claim 3, characterized in that, Based on the set of anomalous grid locations, connectivity merging is performed on spatially adjacent anomalous grid locations to generate candidate spatial grid regions, including: Establish a grid topology based on the spatial grid mapping results; Based on the set of abnormal grid locations, determine the node position of each abnormal grid location in the grid topology and generate abnormal node position results; Based on the results of abnormal node locations, and according to the connection relationship between nodes in the grid topology, abnormal grid location pairs with continuous connection paths are identified, and a path connection set is generated. Based on the path connection set, the abnormal grid positions that are interconnected by continuous connection paths are merged to generate multiple connected node sets; Based on the set of connected nodes, determine the region boundary corresponding to each set of connected nodes and generate candidate spatial grid regions.

7. The online incremental learning neural network optimization method for the laser direct-write alignment model according to claim 3, characterized in that, For each candidate spatial grid region, the residual change information corresponding to that candidate spatial grid region within multiple consecutive processing cycles is extracted, and the continuity of this residual change information is determined to generate a continuous offset region determination result, including: Based on the candidate spatial grid region, determine the corresponding region position of the candidate spatial grid region in multiple consecutive processing cycles, and generate region position tracking results; Based on the regional location tracking results, the residual records of the candidate spatial grid region in multiple consecutive processing cycles are extracted, and a residual change sequence arranged in chronological order is generated. Based on the residual change sequence, the direction of residual change between adjacent processing cycles is compared one by one to determine the processing cycle sequence that satisfies the continuous and consistent change relationship, and a continuity determination result is generated. Based on the continuity determination results, candidate spatial grid regions that satisfy the continuous and consistent change relationship are marked, and continuous offset region determination results are generated.

8. The online incremental learning neural network optimization method for the laser direct-write alignment model according to claim 1, characterized in that, Based on the parameter adjustment range and the parameter adjustment results of the target parameter sub-region, parameter propagation boundaries are set between the parameter sub-regions within the parameter adjustment range to restrict the propagation path of the parameter adjustment results, generating restricted propagation results, including: The parameter sub-region topology is constructed based on the fixed adjacency relationships between parameter sub-regions determined by the spatial grid mapping relationship; Based on the parameter adjustment range, determine the node position of the target parameter sub-region in the parameter sub-region topology, and based on the parameter sub-region topology, determine the set of parameter sub-regions connected to the target parameter sub-region through spatial adjacency within the parameter adjustment range, and generate the connection relationship result; Based on the connection relationship results and the parameter sub-region topology, determine the parameter boundary positions between each parameter sub-region within the parameter adjustment range, and generate the parameter propagation boundary; Based on the parameter propagation boundary and connection relationship results, a set of parameter propagation paths is established starting from the target parameter sub-region and extending along the spatial adjacency relationship, and parameter propagation path results are generated; Based on the parameter propagation path results, path constraints are applied to the parameter adjustment results of the target parameter sub-region, ensuring that the parameter adjustment results are only transmitted along the paths in the parameter propagation path results, and the adjustment amounts exceeding the parameter propagation boundary are truncated to generate restricted propagation results.

9. The online incremental learning neural network optimization method for the laser direct-write alignment model according to claim 1, characterized in that, Based on the limited propagation results, the adjustment amount transmitted to each parameter sub-region is progressively reduced according to the spatial distance between each parameter sub-region within the parameter adjustment range and the target parameter sub-region, generating the incremental adjustment result corresponding to each parameter sub-region, including: Based on the limited propagation results, the spatial distance relationship between each parameter sub-region within the parameter adjustment range and the target parameter sub-region is determined, and the parameter sub-region within the parameter adjustment range is hierarchically divided according to the spatial distance relationship to generate parameter sub-region hierarchical results; Based on the parameter sub-region hierarchy results, the adjustment amount transmitted outward from the target parameter sub-region is distributed level by level, so that the adjustment amount received by the parameter sub-region at a higher spatial distance is less than the adjustment amount received by the parameter sub-region at a lower spatial distance, thus generating a decreasing distribution result. Based on the decreasing allocation result, it is determined whether the adjustment amount received by each level parameter sub-region meets the preset termination condition. When the adjustment amount corresponding to the subsequent level meets the preset termination condition, the transmission of adjustment amount to the subsequent level and the level after that is stopped, and the decreasing allocation result after termination control is generated. Based on the decreasing allocation results after termination control, the model parameters of each parameter sub-region within the parameter adjustment range are updated, generating the incremental adjustment results corresponding to each parameter sub-region.

Images