CN-122021797-A - Online incremental learning neural network optimization method for laser direct-writing alignment model

CN122021797ACN 122021797 ACN122021797 ACN 122021797ACN-122021797-A

Abstract

The invention provides an online increment learning neural network optimization method of a laser direct-writing alignment model, and relates to the technical field of data processing, wherein the method is used for acquiring alignment mark images, laser scanning paths and processing result data, extracting space coordinates and constructing a space grid mapping relation; the method comprises the steps of establishing an initial alignment model based on a historical sample, dividing model parameters into parameter subareas with adjacent relations according to space grids, inputting current data into the model to obtain an alignment compensation result in the processing process, calculating alignment residual errors by combining the processing result to determine a target space grid area and corresponding parameter subareas, performing incremental adjustment on the target parameter subareas, updating the target parameter subareas according to limited propagation of attenuation rules between the adjacent subareas to realize online optimization of the model, and finally superposing the updated alignment compensation result to a laser scanning path to realize scan track subarea correction and improve alignment precision and processing stability.

Inventors

LIU YIQIONG
Weng Wuchen
HUANG GUANLIAN
YU XIAOGUANG
QIU GUOHUA

Assignees

福建福强精密印制线路板有限公司
福建技术师范学院

Dates

Publication Date: 20260512
Application Date: 20260416

Claims (10)

1. The online incremental learning neural network optimization method for the laser direct-writing alignment model is characterized by comprising the following steps of: The method comprises the steps of obtaining an alignment mark image, laser scanning path data and processing result data in the laser direct writing processing process, extracting space coordinate information according to the alignment mark image, carrying out grid division on a processing area according to the space coordinate information, establishing a space grid mapping relation between each grid position and a local position of the laser scanning path, and generating a space grid mapping result; Carrying out parameter partition configuration on the initial alignment model according to a space grid mapping result to divide parameters in the initial alignment model into a plurality of parameter subareas which respectively correspond to different space grid positions, and establishing a space adjacent relation for each parameter subarea to generate an alignment model with a space area corresponding relation; The alignment compensation result is compared with the corresponding processing result data to obtain an alignment residual error distribution result, and a target space grid region is determined according to the alignment residual error distribution result; Performing incremental adjustment on model parameters in the target parameter subregions, setting parameter propagation boundaries between the parameter subregions, and restricting propagation paths of adjustment amounts so that the adjustment amounts are diffused along the spatial adjacent relation between the parameter subregions according to a preset attenuation rule to generate an incremental updated alignment model; and superposing the alignment compensation result output by the alignment model after incremental updating to the corresponding local position in the laser scanning path, and carrying out partition correction on the laser scanning track to obtain an updated laser scanning control result.
2. The method for optimizing an online incremental learning neural network of a laser direct-write alignment model according to claim 1, wherein the method for optimizing the online incremental learning neural network of the laser direct-write alignment model is characterized in that parameter partitioning configuration is performed on an initial alignment model according to a space grid mapping result, parameters in the initial alignment model are divided into a plurality of parameter sub-areas respectively corresponding to different space grid positions, a space adjacent relation is established for each parameter sub-area, and an alignment model with a space area corresponding relation is generated, and the method comprises the following steps: According to the space grid mapping result, carrying out partition calibration on an input area of the initial alignment model, determining the corresponding relation between each grid position in the processing area and each local data processing range in the input area, and generating a corresponding result of the grid position and the local data processing range; Dividing model parameters participating in data processing of each local data processing range in the initial alignment model according to the corresponding results of the grid positions and the local data processing ranges, and generating parameter sets corresponding to each grid position respectively; according to the parameter sets, carrying out position identification processing on each parameter set, mapping the spatial adjacent relation between grid positions into the spatial adjacent relation between each parameter set, and generating a parameter subarea with spatial position identification and spatial adjacent relation; And constructing an alignment model with a spatial region corresponding relation according to the parameter subregion with the spatial position identifier and the spatial adjacent relation.
3. The method for optimizing an online incremental learning neural network of a laser direct-write alignment model according to claim 1, wherein determining a target spatial grid region according to an alignment residual distribution result comprises: extracting residual information corresponding to each grid position according to the alignment residual distribution result, judging the residual information corresponding to each grid position, determining abnormal grid positions of which the residual information meets preset residual judgment conditions, and generating an abnormal grid position set; According to the abnormal grid position set, connectivity combination is carried out on the adjacent abnormal grid positions of the space, and a candidate space grid region is generated; Extracting residual variation information corresponding to each candidate space grid region in a plurality of continuous processing cycles, and continuously judging the residual variation information to generate a continuous offset region judgment result; And according to the continuous offset region judgment result, determining the candidate space grid region meeting the preset continuous offset judgment condition as a target space grid region.
4. The online incremental learning neural network optimization method of the laser direct-writing alignment model according to claim 1, wherein the incremental adjustment is performed on model parameters in target parameter subregions, parameter propagation boundaries are set between the parameter subregions, propagation paths of adjustment amounts are constrained, the adjustment amounts are diffused along a spatial adjacent relation between the parameter subregions according to a preset attenuation rule, and the generation of the incrementally updated alignment model comprises: Determining a target parameter subarea corresponding to the target space grid region according to the position identification of the target space grid region in the alignment model, determining a parameter subarea set communicated with the target parameter subarea through the spatial adjacency according to the spatial adjacency of the target parameter subarea, and generating a parameter adjustment range; Extracting corresponding alignment residual errors according to the target space grid region, and associating the alignment residual errors with the target parameter subregion; Setting parameter propagation boundaries among all the parameter subregions in the parameter adjustment range according to the parameter adjustment range and the parameter adjustment result of the target parameter subregion, limiting the propagation path of the parameter adjustment result, and generating a limited propagation result; According to the limited transmission result, performing step-by-step decreasing processing on the adjustment quantity transmitted to each parameter subarea according to the sequence of the space distance between each parameter subarea and the target parameter subarea in the parameter adjustment range, and generating an increment adjustment result corresponding to each parameter subarea; And updating the alignment model according to the increment adjustment result corresponding to each parameter subarea, and generating an incrementally updated alignment model.
5. The optimization method for online incremental learning neural network of laser direct-write alignment model according to claim 2, wherein dividing model parameters participating in data processing of each local data processing range in the initial alignment model according to the corresponding result of the grid position and the local data processing range, generating parameter sets respectively corresponding to each grid position, comprises: According to the corresponding result of the grid positions and the local data processing ranges, determining fixed data processing paths of the local data processing ranges in the initial alignment model, and generating data processing paths corresponding to the grid positions one by one; Extracting model parameters which participate in operation in the corresponding data processing paths in the initial alignment model according to the data processing paths, classifying according to the data processing paths, and generating parameter classification results corresponding to the data processing paths; according to the parameter classification results, independently packaging the parameter classification results, so that each parameter classification result corresponds to only one grid position, and generating parameter sets corresponding to the grid positions respectively; and binding and marking each parameter set, so that the corresponding relation between each parameter set and the corresponding grid position is kept unchanged in the model updating process.
6. The method for optimizing an online incremental learning neural network of a laser direct-write alignment model according to claim 2, wherein performing a position identification process on each parameter set according to the parameter set, mapping a spatial adjacency relationship between grid positions to a spatial adjacency relationship between parameter sets, and generating a parameter sub-region having a spatial position identification and a spatial adjacency relationship, comprises: According to the parameter sets, assigning a space coordinate identifier consistent with the corresponding grid position to each parameter set, and generating a parameter set with the space coordinate identifier; Constructing a grid topological structure according to the connection relation between grid positions in the processing area, determining the direct adjacent relation between the grid positions according to the grid topological structure, and generating a grid adjacent relation result; According to parameter sets with space coordinate marks and grid adjacency relation results, mapping direct adjacency relations among grid positions into fixed adjacency relations among the parameter sets, and generating parameter set adjacency relation results; And organizing each parameter set according to the parameter set with the spatial coordinate mark and the parameter set adjacency relation result to generate a parameter subarea with the spatial position mark and the fixed topological adjacency relation.
7. The method for optimizing an online incremental learning neural network of a laser direct-write alignment model according to claim 3, wherein connectivity merging is performed on spatially adjacent abnormal grid positions according to an abnormal grid position set to generate a candidate spatial grid region, comprising: Establishing a grid topological structure based on the space grid mapping result; Determining the node position of each abnormal grid position in the grid topological structure according to the abnormal grid position set, and generating an abnormal node position result; according to the abnormal node position result and the connection relation between nodes in the grid topological structure, identifying abnormal grid position pairs with continuous connection paths, and generating a path connection set; combining abnormal grid positions which are mutually communicated through continuous connection paths according to the path connection set to generate a plurality of communication node sets; and determining the region boundaries corresponding to the connected node sets according to the connected node sets, and generating candidate space grid regions.
8. The optimization method of online incremental learning neural network of laser direct-write alignment model according to claim 3, wherein extracting residual variation information corresponding to each candidate space grid region in a plurality of continuous processing cycles, and performing continuity judgment on the residual variation information, and generating a continuous offset region judgment result, comprises: Determining the corresponding region positions of the candidate space grid region in a plurality of continuous processing cycles according to the candidate space grid region, and generating a region position tracking result; Extracting residual records of the candidate space grid region in a plurality of continuous processing periods according to the region position tracking result, and generating a residual change sequence arranged in time sequence; comparing the residual variation directions among adjacent processing periods one by one according to the residual variation sequence, determining a processing period sequence meeting a continuous consistent variation relation, and generating a continuity judging result; And marking the candidate space grid areas meeting the continuous consistent change relation according to the continuity judging result to generate a continuous offset area judging result.
9. The method for optimizing an online incremental learning neural network of a laser direct-write alignment model according to claim 4, wherein setting parameter propagation boundaries between the parameter subregions in the parameter adjustment range according to the parameter adjustment results of the parameter adjustment range and the target parameter subregions, limiting propagation paths of the parameter adjustment results, and generating limited propagation results, comprises: Constructing a parameter sub-region topological structure according to a fixed adjacency relation between the parameter sub-regions determined by the space grid mapping relation; Determining the node position of a target parameter subregion in a parameter subregion topological structure according to the parameter adjustment range, determining a parameter subregion set communicated with the target parameter subregion in the parameter adjustment range through a space adjacency relation based on the parameter subregion topological structure, and generating a connection relation result; Determining the parameter boundary positions among all the parameter subregions in the parameter adjustment range according to the connection relation result and the parameter subregion topological structure, and generating a parameter propagation boundary; According to the parameter propagation boundary and the connection relation result, a parameter propagation path set which starts from a target parameter sub-region and extends along a space adjacent relation is established, and a parameter propagation path result is generated; And carrying out path constraint on the parameter adjustment result of the target parameter subregion according to the parameter propagation path result, enabling the parameter adjustment result to be transmitted only along the path in the parameter propagation path result, and carrying out truncation processing on the adjustment amount exceeding the parameter propagation boundary to generate a limited propagation result.
10. The optimization method for online incremental learning neural network of laser direct-write alignment model according to claim 4, wherein the step-by-step decreasing processing is performed on the adjustment amounts transferred to each parameter sub-region according to the spatial distance sequence between each parameter sub-region and the target parameter sub-region in the parameter adjustment range and according to the limited propagation result, and the generation of the incremental adjustment result corresponding to each parameter sub-region comprises: Determining a spatial distance relation between each parameter subarea and a target parameter subarea in a parameter adjustment range according to the limited transmission result, and carrying out hierarchical division on each parameter subarea in the parameter adjustment range according to the spatial distance relation to generate a parameter subarea hierarchical result; According to the parameter sub-region level result, gradually distributing the adjustment quantity transmitted outwards by the target parameter sub-region, so that the adjustment quantity received by the parameter sub-region with a farther space distance is smaller than the adjustment quantity received by the parameter sub-region with a nearer space distance, and generating a descending distribution result; Judging whether the adjustment quantity received by each level parameter subarea meets a preset termination condition according to the decremental allocation result, and stopping transmitting the adjustment quantity to the subsequent level and the subsequent level when the adjustment quantity corresponding to the subsequent level meets the preset termination condition, so as to generate the decremental allocation result after termination control; And updating the model parameters of the sub-regions of each parameter in the parameter adjustment range according to the decremental distribution result after termination control, and generating an increment adjustment result corresponding to each parameter sub-region.

Description

Online incremental learning neural network optimization method for laser direct-writing alignment model Technical Field The invention relates to the technical field of data processing, in particular to an online incremental learning neural network optimization method for a laser direct-writing alignment model. Background In the prior art, laser direct writing alignment and processing optimization are generally realized by combining a photoetching control model and a neural network method. Specifically, a training data set is constructed by collecting exposure dose distribution and actual processing morphology data, a nonlinear mapping model between exposure parameters and forming results is established by utilizing a convolutional neural network (such as U-net), and the nonlinear mapping model is used for predicting optimal exposure control parameters after offline training, so that alignment and processing precision optimization is realized in the steps of exposure, development and the like. Meanwhile, the system depends on a computer to control laser beam scanning and CCD alignment detection, so that mapping and alignment correction of graphic data to an actual structure are realized. However, in the large-size PCB laser direct writing application scenario in advanced packaging, the method generally adopts a fixed training model for reasoning, and when exposure response changes due to material batch differences or equipment thermal drift in the processing process, the model cannot be updated in real time. For example, in the continuous roll-to-roll exposure process, the tension change of the substrate can cause the position deviation of the alignment mark, and if the pre-training model is still adopted for compensation, the accumulated alignment error can be caused, so that the subsequent interlayer superposition deviation is increased, and the local line dislocation or open circuit defect occurs. The lack of online incremental learning capability directly affects the consistency and yield of multi-layer structure processing. Disclosure of Invention The invention aims to provide an online incremental learning neural network optimization method for a laser direct-writing alignment model, and aims to solve the problems in the background art. In order to solve the technical problems, the technical scheme of the invention is as follows: An online incremental learning neural network optimization method of a laser direct-writing alignment model, the method comprising the following steps: The method comprises the steps of obtaining an alignment mark image, laser scanning path data and processing result data in the laser direct writing processing process, extracting space coordinate information according to the alignment mark image, carrying out grid division on a processing area according to the space coordinate information, establishing a space grid mapping relation between each grid position and a local position of the laser scanning path, and generating a space grid mapping result; Carrying out parameter partition configuration on the initial alignment model according to a space grid mapping result to divide parameters in the initial alignment model into a plurality of parameter subareas which respectively correspond to different space grid positions, and establishing a space adjacent relation for each parameter subarea to generate an alignment model with a space area corresponding relation; The alignment compensation result is compared with the corresponding processing result data to obtain an alignment residual error distribution result, and a target space grid region is determined according to the alignment residual error distribution result; Performing incremental adjustment on model parameters in the target parameter subregions, setting parameter propagation boundaries between the parameter subregions, and restricting propagation paths of adjustment amounts so that the adjustment amounts are diffused along the spatial adjacent relation between the parameter subregions according to a preset attenuation rule to generate an incremental updated alignment model; and superposing the alignment compensation result output by the alignment model after incremental updating to the corresponding local position in the laser scanning path, and carrying out partition correction on the laser scanning track to obtain an updated laser scanning control result. The scheme of the invention at least comprises the following beneficial effects: Firstly, the invention realizes the spatial correspondence between the processing area and the model parameters by acquiring the alignment mark image, the laser scanning path and the processing result data and constructing the spatial grid mapping relation, so that the alignment model can be converted from integral unified modeling to regional modeling with spatial resolution capability, thereby improving the expression capability of local errors. Secondly, on the basis, the parameters of