A method for printing on large-size panels

Abstract

This invention belongs to the field of printed display technology, specifically relating to a method for printing planning of large-size panels. The method includes: dividing a large-size panel at a certain mounting angle to be printed into multiple regions, wherein the division satisfies the following condition: the Y-axis deviation between the first and last pixels of the same column arranged along the X-axis in the pixel array of each region is less than a threshold; representing the large-size panel printing planning problem after region division as an integer programming model through mathematical modeling, where the printing planning of each partition is considered as the printing planning of a single row of pixels within that partition; converting the integer programming model into a ternary graph form; inputting the ternary graph into a decision node prediction model to obtain the values ​​of each decision node, i.e., the number of prints per nozzle at each stop position of the printhead, thus completing the printing planning method. This invention employs a machine learning prediction model, which can be used to improve the efficiency of patterning planning calculations in the manufacturing of large-size printed display panels, thereby increasing production efficiency.

Description

Technical Field

[0001] This invention belongs to the field of printed display technology, and more specifically, relates to a method for planning the printing of large-size panels for novel displays. Background Technology

[0002] Manufacturing organic light-emitting diodes (OLEDs) using inkjet printing offers advantages such as low environmental requirements, no need for photomasks, rapid development, and material savings, showing broad application prospects in smart products like mobile phones and laptops, and becoming one of the core technologies for the transformation of new displays. However, many problems still need to be overcome from technological realization to manufacturing. In addition to printing materials, achieving high efficiency, consistency, uniformity, and absence of Mura defects in large-size OLED panels remains a significant challenge for inkjet OLEDs.

[0003] Large-size panels require planning a massive number of nozzles and pixels; the G10 panel alone has billions of pixels, and the positioning accuracy varies, severely impacting the efficiency of pixel pit printing planning during a single printing cycle. While solvers can be used to address this massive amount of data, their speed is limited. In other words, a complete and efficient solution for large-size inkjet printing display patterns is currently lacking. Firstly, the printing platform requires high positioning accuracy; different substrate angles significantly affect the efficiency of printing planning and execution. Secondly, as the printing area of ​​the substrate increases, the scale of patterning increases, and the number of pixels and nozzles also increases, leading to a dramatic increase in the computational scale and significantly limiting the solution efficiency in terms of time.

[0004] Therefore, there is a need in the field for an efficient algorithm for patterning planning of large-area printed displays in order to effectively realize large-scale pixel inkjet printing of new displays. Summary of the Invention

[0005] In view of the shortcomings of existing technologies and the need for improvement, this invention provides a large-size panel printing planning method for novel displays, the purpose of which is to provide a large-size panel printing planning method that combines high precision and high speed.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for printing planning on large-size panels for novel displays is provided, comprising:

[0007] For a large panel with a certain installation angle δ for printing planning, it is divided into multiple regions. The division method satisfies the following: the Y-direction deviation between the first and last pixels of the same column arranged along the X direction in the pixel array of each region is less than a threshold; the angle δ represents the angle between the line connecting the first and last pixels of the same column arranged along the X direction in the large panel and the X direction, where the X direction is the direction of printhead movement during printing.

[0008] The printing planning problem of a large-size panel after dividing the area is represented by an integer programming model through mathematical modeling. The printing planning of each partition is regarded as the printing planning of a single row of pixels within that partition. The integer programming model is transformed into a ternary graph. The decision nodes of the ternary graph represent the decision variables in the integer programming model, representing the number of times each nozzle is printed at each stop position of the printhead. The constraint nodes of the ternary graph represent the constraint conditions in the integer programming model, representing the volume constraint of each pixel. The target node of the ternary graph represents the target variable in the integer programming model, representing the maximum number of times all nozzles are printed at each stop position of the printhead.

[0009] By inputting the three-part graph into the decision node prediction model, the values ​​of each decision node are obtained, which represent the number of times each nozzle is printed at each stop position of the printhead, thus completing the large-size panel printing planning method.

[0010] Furthermore, the specific implementation method for dividing the large-size panel into multiple regions is as follows:

[0011] Determine the angle δ between the line connecting the first and last pixels of the same column along the X-axis in the large-size panel to be divided and the X-axis, and the printable Y-axis length r within each pixel as the threshold. y Based on geometric relationships, calculate the X-axis length l of the region to be divided. s for:

[0012] Based on the X-axis spacing g of adjacent pixels in a large-size panel px Calculate the number of rows n in the X direction within each partitioned region. sx :

[0013] Furthermore, the printable Y-axis length r within each pixel is... y The threshold is determined as follows:

[0014] S1. Take the median of the landing point deviations of all nozzles as the printing deviation of the entire printhead:

[0015]

[0016] Among them, v jet,k and θ k These represent the injection velocity and injection angle of nozzle k, respectively, where m is the total number of nozzles, and v print The speed at which the large panel moves during printing is indicated; h represents the height of the printhead from the large panel during printing.

[0017] S2, Obtain the pixel Y-axis dimension R y Calculate the printable Y-axis length r within each pixel. yr y =R y -2×d error .

[0018] Furthermore, the decision node prediction model is constructed in the following manner:

[0019] The large-size panel at different included angles δ is divided into multiple regions according to the aforementioned division method;

[0020] The large-size panel printing planning problem after each division of the region is represented by an integer programming model through mathematical modeling and converted into a tripartite graph form to obtain a tripartite graph set. An integer programming solver is used to solve the integer programming model corresponding to each included angle δ to obtain the values ​​of all decision variables in the corresponding model, which constitute the decision node labels of the corresponding tripartite graph. Each tripartite graph and its corresponding decision node label are used as a training sample to obtain the first training sample set.

[0021] The decision node prediction network is trained using the first training sample set to obtain the decision node prediction model.

[0022] Furthermore, the decision node prediction network is a graph sampling aggregation-based decision node prediction network. Specifically, it first performs partial aggregation and extraction of node information in the three-part graph: information aggregation is performed on each target node and each constraint node to obtain the first embedding; information aggregation is performed on each decision node based on each first embedding to obtain the second embedding; all second embeddings are merged to obtain the graph embedding; then, the value of each decision variable is predicted by an MLP multilayer perceptron based on the graph embedding to generate the classification probability of the decision node.

[0023] Furthermore, after predicting the values ​​of each decision node, the method also includes:

[0024] The values ​​of each decision node are input into the decision variable selection model, and the decision variable selection model determines the accurate and inaccurate decision node values ​​among the decision node values.

[0025] Each inaccurate decision node value is modified to a range of values ​​smaller than its current value; all accurate decision node values ​​and the modified decision node value ranges are combined to form a new decision node value vector.

[0026] The new decision node value vector is input into the graph sampling aggregation branch and bound model. The graph sampling aggregation branch and bound model solves for the decision node values ​​corresponding to the value range of each decision node in the new decision node value vector, determines an accurate value within the corresponding value range, and obtains the final decision node value vector, thus completing the large-size panel inkjet printing planning method.

[0027] Furthermore, in constructing the decision node prediction model, an open-source integer programming solver is used to solve the integer programming model corresponding to each included angle δ; the graph sampling aggregation branch and bound model is then constructed in the following manner:

[0028] In the process of solving the integer programming model corresponding to each included angle δ, the decision variable selection strategy and branching strategy in the solution process are also obtained as the decision labels of the corresponding tripartite graph; a training sample is determined by the tripartite graph corresponding to the included angle δ and its corresponding decision labels to form the second training sample set;

[0029] The graph sampling aggregation branch bound network is trained using the second training sample set to obtain the graph sampling aggregation branch bound model.

[0030] Furthermore, the decision variable selection model is constructed in the following manner:

[0031] Using a trained decision node prediction model, based on each ternary graph in the first training sample set, the values ​​of each decision node corresponding to the predicted output are predicted. The values ​​of each decision node corresponding to each ternary graph in the predicted output, as well as the labels of whether each decision node is selected or not, are used as training samples to construct a third training sample set. The labels of whether each decision node is selected or not are obtained in the following way: based on the values ​​of each decision node corresponding to the predicted output of the ternary graph and the labels of the decision nodes in the first training sample set, the decision nodes with accurate values ​​and the decision nodes with inaccurate values ​​are selected to form a binary classification vector of whether each decision node is selected or not, which is used as the label of whether each decision node is selected or not in the predicted output of the ternary graph.

[0032] The decision variable selection network is trained using the third training sample set to obtain the decision variable selection model.

[0033] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects:

[0034] (1) Regarding the pixel planning problem for large-size panels, this invention first proposes a parallel planning strategy for large-size panels. Based on different substrate correction errors, the row number boundary, which can be considered as the same row for printing, can be obtained, greatly simplifying the number of pixel pits involved in the printing planning and reducing the exponential increase in planning difficulty caused by the rapid increase in the number of pixels. Simultaneously, regional planning can be parallelized, shortening planning time and improving efficiency. Secondly, this invention models the patterning planning problem as an integer programming problem, using the uniformity of volume within each pixel pit as a constraint to limit pixel volume differences, and minimizing the number of printing operations as the optimization objective, thus shortening the time spent printing a single substrate. Furthermore, this invention describes the integer programming problem in the form of a graph, treating the decision variables, constraint coefficients, and optimization objective as three types of nodes, forming a ternary graph. When there are non-zero coefficients between variables and constraints, there is an edge connection between two nodes. Representing the integer programming problem in graph form allows it to be used as input for training a graph neural network, satisfying permutation invariance and ensuring that the training prediction results are not changed by the node order. Solving the printing planning problem using deep learning significantly reduces computation time and effectively improves the production efficiency of large-scale substrate production lines. In other words, the patterned integer planning problem is transformed into a graph topology, which can be used as input to a graph neural network. Through node classification in the graph neural network, the planning time for large-area patterning is shortened, effectively improving the printing planning efficiency of large-area, high-precision inkjet display production lines and achieving efficient manufacturing.

[0035] (2) The decision node prediction model adopts a graph sampling aggregation framework to meet the training process of large-scale graphs, aggregates random neighbors of each node, reduces the aggregation scale, learns the aggregation pattern between nodes, realizes the processing of large-scale planning problems in the printing planning process, and improves training efficiency.

[0036] (3) The predicted values ​​obtained by convolution are filtered a second time to extract variables with relatively certain probability values ​​and narrow the range of values ​​for variables with relatively dispersed probability values ​​to form a high-precision initial solution. Then, a graph sampling aggregation branch and bound model that has learned the solver branch and bound strategy (ensuring that the model reaches or even exceeds the solver's accuracy) is adopted to solve the specific values ​​of the range of values ​​in the initial solution. That is, the large problem is divided into small problems for solving, which improves the efficiency of processing the same type of planning problem. Attached Figure Description

[0037] Figure 1 A flowchart illustrating a method for planning printing on a large-size panel for a novel display, provided by an embodiment of the present invention;

[0038] Figure 2 A schematic diagram of substrate pixel partitioning for a large-size panel with partitioning provided in an embodiment of the present invention;

[0039] Figure 3 This is a schematic diagram of a large-size panel pixel partitioning scheme provided in an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram illustrating the conversion of inkjet printing integer planning into a ternary graph structure in an inkjet printing planning method for a large-size panel of a novel display provided by an embodiment of the present invention.

[0041] Figure 5 This is a schematic diagram of the parameters affecting the droplet landing point error provided in an embodiment of the present invention;

[0042] Figure 6 A flowchart illustrating a method for planning the printing of large-size panels for novel displays, provided in an embodiment of the present invention.

[0043] Figure 7 A network framework diagram of a large-size panel inkjet printing planning method for a novel display provided by an embodiment of the present invention;

[0044] Figure 8 This is a schematic diagram of the convolution method based on the graph neural network solving algorithm provided in an embodiment of the present invention;

[0045] Figure 9 The following is an architecture diagram of a decision node prediction network and a graph sampling aggregation branch delimitation network based on graph sampling aggregation provided in an embodiment of the present invention;

[0046] Figure 10 This is a schematic diagram of a decision variable selection network provided in an embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0048] Example 1

[0049] A method for printing large-size panels for novel displays, such as Figure 1 As shown, it includes:

[0050] S1. Divide the large-size panel under a certain installation angle δ for printing planning into multiple regions. The division method satisfies the following: the Y-direction deviation between the first and last pixels of the same column arranged along the X direction in the pixel array of each region is less than a threshold; the angle δ represents the angle between the line connecting the first and last pixels of the same column arranged along the X direction in the large-size panel and the X direction, which represents the relative position coordinates between the nozzle array and the pixels. The X direction is the direction of print head movement during printing.

[0051] S2. The large-size panel printing planning problem after dividing the area is represented by an integer programming model through mathematical modeling. The printing planning of each partition is regarded as the printing planning of a single row of pixels within that partition. The integer programming model is converted into a ternary graph. The decision nodes of the ternary graph represent the decision variables in the integer programming model, representing the number of times each nozzle is printed at each stop position of the printhead. The constraint nodes of the ternary graph represent the constraint conditions in the integer programming model, representing the volume constraint of each pixel. The target node of the ternary graph represents the target variable in the integer programming model, representing the maximum number of times all nozzles are printed at each stop position of the printhead.

[0052] S3. Input the three-part graph into the decision node prediction model to obtain the value of each decision node, that is, the number of times each nozzle is printed at each stop position of the nozzle, thus completing the large-size panel printing planning method.

[0053] It should be noted that the method for partitioning large-size panels is tailored to the specific working conditions of large-size panels. When the large-size substrate being printed consists of multiple area panels, it is determined whether the angle correction accuracy covers the Y-direction deviation of a single area panel. If it does, the area panel is treated as a single row of pixels for planning, and the next area panel is then determined. Otherwise, the single area panel is further divided into multiple sub-regions. If the large-size substrate is a single pixel panel without its own partitions, the pixel area is divided according to the above partitioning method based on the angle deviation limit to form multi-pixel areas. Figure 2 and Figure 3 To address pixel partitioning methods for different types of substrates and reduce printing point defects caused by substrate correction errors, the nozzle array is arranged along the Y direction, with Z and Y degree of freedom for movement. The pixel substrate is laid out along the X and Y directions, and the substrate moves along the X direction with the platform during printing.

[0054] The large-scale pixel panel is divided into multiple partitioned pixel arrangements. Each pixel partition is repositioned according to the angular deviation of the substrate. Modeling and parallel planning are performed for each partition separately. The planning results are then merged in the printing direction to generate the final printing data for the large-scale panel. This method partitions the panel to reduce the influence of angle, but partitioning increases the number of pixels involved in the planning. Therefore, parallel construction of the positional relationships between all nozzles and pixel pits in each region shortens the model construction time. The solutions are then merged to ensure high accuracy. Finally, the printing data is generated in parallel to accelerate the process and expedite printing.

[0055] Large-size panels, through a partitioned parallel planning strategy, can obtain the row number boundary that can be regarded as the same row of printing based on different substrate correction errors. This greatly simplifies the number of pixel pits involved in printing planning and reduces the exponential increase in planning difficulty caused by the rapid increase in the number of pixels. At the same time, partitioned planning can achieve parallelization, shorten planning time and improve efficiency.

[0056] Regarding the transformation of integer programming models into ternary graph form (i.e., ternary graph topology), such as Figure 4 As shown, a large-size panel to be printed has been installed with a fixed installation angle δ. An integer programming model is constructed under this installation angle δ, where the number of prints per pixel is a constraint and the optimization objective is to minimize the total number of prints.

[0057]

[0058] st

[0059]

[0060] Among them, a i,k×j,δ For different substrate offset angles δ, the positional relationship between the nozzle k and pixel i at the nozzle stopping point j is given. When the nozzle is within the sprayable length range of the pixel, then a... i,k×j,δ The value is the volume of nozzle k, otherwise it is 0; x k×j,1 V represents the number of times nozzle k sprays at the nozzle's resting point j. ub and V lb The upper and lower limits of the volume for each pixel, y j V represents the maximum number of sprays from all nozzles at docking point j. dn and V p Let k, i, and j represent the median of the droplet volume ejected in a single stroke across all nozzle volumes and the volume required to print a pixel, respectively. k, i, and j represent the nozzle index, pixel index, and docking point index, respectively, where k ∈ [1, m], i ∈ [1, n], and j ∈ [1, p].

[0061] Transform the above integer programming model into a tripartite graph representation:

[0062] G = (X, C, Y, E)

[0063] X is the set of decision variable nodes, defined by x k×j,1 Composition; C represents the set of constraint nodes consisting of V ub V lb and V p Composition; Y represents the target node set, consisting of y j Composition; E represents the connecting edge between all nodes, consisting of the first node and the last node, and the solution to each integer programming problem is used as the label of the decision node.

[0064] First, this method models the patterning planning problem as an integer programming problem, using the uniformity of volume within each pixel pit as a constraint to limit pixel volume differences, and minimizing the number of printing passes as the optimization objective to reduce the time spent printing a single substrate. This method increases the system's fitness and is applicable to patterning planning systems of any precision. Furthermore, this method describes the integer programming problem in graph form, treating the decision variables, constraint coefficients, and optimization objective as three types of nodes, forming a ternary graph. When there are non-zero coefficients between variables and constraints, there is an edge connection between two nodes. Representing the integer programming problem in graph form allows it to be used as input for training graph neural networks, and it satisfies permutation invariance, meaning the training prediction results are not affected by the node order.

[0065] As a preferred embodiment, the specific implementation method for dividing a large-size panel into multiple regions is as follows:

[0066] Determine the angle δ between the line connecting the first and last pixels of the same column along the X-axis in the large-size panel to be divided and the X-axis, and the printable Y-axis length r within each pixel as the threshold. y Based on geometric relationships, calculate the X-axis length l of the region to be divided. s for:

[0067] Based on the X-axis spacing g of adjacent pixels in a large-size panel px Calculate the number of rows n in the X direction within each partitioned region. sx :

[0068] As a preferred embodiment, the printable Y-axis length r within each pixel is... y The threshold is determined as follows:

[0069] S1. The printable range of a pixel is determined by the pixel size and the printing error of the nozzle. The median of the landing point deviations of all nozzles is taken as the printing deviation of the entire printhead.

[0070]

[0071] Among them, v jet,k and θ k These represent the injection velocity and injection angle of nozzle k, respectively, where m is the total number of nozzles, and v print This indicates the movement speed of the large panel during printing; h represents the height of the printhead from the large panel during printing, such as... Figure 5 As shown.

[0072] S2, Obtain the pixel Y-axis dimension R y Calculate the printable Y-axis length r within each pixel.y r y =R y -2×d error .

[0073] Substituting into the above formula, the simplified number of pixel rows when the substrate offset angle is δ can be expressed by the following formula:

[0074]

[0075] Based on this formula, the simplified number of partition rows can be obtained, thus dividing the entire substrate into regions. Let N be the total number of rows in the X direction of the pixel. px The total number of columns in the Y direction is N. py The total number of pixels n to be planned after partitioning p for:

[0076] As a preferred implementation method, the above-mentioned decision node prediction model is constructed in the following manner:

[0077] The large-size panel at different included angles δ is divided into multiple regions according to the above division method. That is, the division method of the large-size panel at each included angle δ satisfies the following: the Y-direction deviation between the first and last pixels of the same column arranged along the X direction in the pixel array of each region is less than the threshold; the included angle δ represents the angle between the line connecting the first and last pixels of the same column arranged along the X direction in the large-size panel and the X direction.

[0078] The large-size panel printing plan after dividing the area under each included angle δ is represented by an integer programming model through mathematical modeling. The integer programming model corresponding to each included angle δ is converted into a ternary graph form, resulting in a set of ternary graphs. Each ternary graph has decision nodes, constraint nodes, and target nodes. Decision nodes represent decision variables in the integer programming model, representing the number of times each nozzle is printed at each stop position of the printhead. Constraint nodes represent constraint conditions in the integer programming model, representing the volume constraint of each pixel. Target nodes represent target variables in the integer programming model, representing the maximum number of times all nozzles are printed at each stop position of the printhead.

[0079] An integer programming solver is used to solve the integer programming model corresponding to each included angle δ, and the values ​​of all decision variables in the corresponding model are obtained to form the decision node labels of the corresponding tripartite graph. Each tripartite graph and its corresponding decision node labels are used as training samples to construct the first training sample set.

[0080] The decision node prediction network is trained using the first training sample set to obtain the decision node prediction model.

[0081] As a preferred implementation, the aforementioned decision node prediction network is a decision node prediction network based on graph sampling aggregation (GraphSAGE). Specifically, it first performs partial aggregation and extraction of node information in the three-part graph: information aggregation is performed on each target node and each constraint node to obtain a first embedding; information aggregation is performed on each decision node based on each first embedding to obtain a second embedding; all second embeddings are merged to obtain a graph embedding; then, an MLP (Multilayer Perceptron) is used to predict the value of each decision variable based on the graph embedding to generate the classification probability of the decision node.

[0082] A large-scale tripartite graph set is used as the training set for the decision node prediction network, and the weight parameters are learned using the information aggregation strategy between nodes:

[0083] The aggregation function of a node is represented as:

[0084]

[0085] in, and Let V be the convolutional embedding representations of the v-th node in the r-th and (r-1)-th layers, respectively. W is the embedding representation of the neighbors of the v-th node in the (r-1)-th layer. r Let be the weight parameters of the r-th layer. The convolution size is reduced by randomly selecting the representations of each node's neighboring nodes for convolution. The network ultimately outputs the embedding representation v of each node. k×j k×j is the feature dimension.

[0086] Predicting the overall value using a conditionally independent distribution model:

[0087]

[0088] In the formula, x k×j Let x be the value of the k×j-th dimension node x, M represents the parameter set of the current integer programming, i.e. M=(a,V,1), where a is the left constraint coefficient, V is the right constraint coefficient, 1 is the optimization objective coefficient, and L=m×p represents the total dimension of the features.

[0089] The model parameters θ are learned by minimizing the following weighted loss function with respect to θ:

[0090]

[0091] Wherein, weight w k×j Used to reduce any bias when sampling for training instance M.

[0092] At this time x k×jThe value can take integer values ​​beyond 0 and 1. These integer values ​​are converted to one-hot encoding, and the value of each bit represents the probability of a node. The prediction layer consists of a multilayer perceptron, and the prediction results and accuracy are expressed as follows:

[0093] t k×j =MLP(v k×j ,θ);

[0094]

[0095] Among them, v k×j For the convolutional embedding output, all nodes use the same MLP parameters.

[0096] The decision node prediction model of this preferred method can accommodate graph inputs of varying sizes. Since aggregation only focuses on the neighbor information of each node, the graph input is not constrained by the size of each ternary graph, allowing for merged inputs. Furthermore, the graph sampling aggregation framework satisfies the training process for large-scale graphs, aggregating random neighbors of each node to reduce the aggregation scale and learn aggregation patterns between nodes. This enables the handling of large-scale planning problems in the printing planning process, improving training efficiency.

[0097] This can be used as a preferred implementation method, such as Figure 6 and Figure 7 As shown, after predicting the values ​​of each decision node, the method also includes:

[0098] The predicted values ​​of each decision node are input into the decision variable selection model, which then determines the accurate and inaccurate decision node values.

[0099] Each inaccurate decision node value is modified to a range of values ​​smaller than its current value; all accurate decision node values ​​and the modified decision node value ranges are combined to form a new decision node value vector.

[0100] The new decision node value vector is input into the graph sampling aggregation branch and bound model. The graph sampling aggregation branch and bound model solves for the decision node values ​​corresponding to the value range of each decision node in the new decision node value vector. Within the corresponding value range, an accurate value is determined, and the final decision node value vector is obtained, thus completing the large-size panel inkjet printing planning method.

[0101] In other words, this preferred method first solves the problem by using a decision node prediction model based on graph sampling aggregation, a decision variable selection and scaling model, and a graph sampling aggregation branch and bound model. The decision node prediction model provides pre-assignments for all variables in the planning problem and learns the node classification task of the graph through convolution of graph sampling aggregation to classify the decision variable nodes. The decision variable selection model and scaling operation determine which variable assignments to use based on the predicted results, and narrows the range of other unselected variables. The large problem is divided into smaller problems for solving, with a high-precision initial solution as the output. Finally, the graph sampling aggregation branch and bound model, which learns the branching strategy of the solver, achieves an efficient solution to the final problem.

[0102] Furthermore, the model's training dataset consists of a graph dataset of printing planning problems. Each integer programming problem can obtain an optimal solution through a solver. The integer programming problems are then categorized according to... Figure 4 The method is transformed into a graph data structure, and the corresponding optimal solution is used as the label of the graph node. At the same time, the branch variable selection path of the solver's solution process is recorded as the training data for the decision node prediction model and the graph sampling aggregation branch and bound model, respectively. The decision variable selection and scaling model is trained by the prediction results generated by the decision node prediction model, and different nodes are selected to obtain the loss in order to obtain appropriate weight parameters.

[0103] In other words, from the predicted results, the values ​​of variables with high feasibility are selected as the final values ​​of those variables. The range of values ​​for variables with uncertain probabilities is narrowed to a value smaller than the current value. A small-scale graph is output after determining the values ​​of some decision nodes. Furthermore, a graph sampling aggregation branch-and-bound model is employed to improve the prediction efficiency of this model. Specifically, as an optimization, this invention uses a graph sampling aggregation network algorithm to predict the values ​​of some variables in an integer programming problem. This serves as a "hot start" solution for large-scale integer programming problems, dividing the large programming problem into smaller subproblems for solving, significantly improving computational efficiency and shortening the production line process window time.

[0104] In summary, this preferred method, when running online, divides the pixel planning area of ​​the entire large-size panel according to the current substrate positioning, constructs an integer programming model for printing planning, represents it with a graph data structure, inputs it into a graph sampling aggregation prediction network to obtain prediction results, and then uses a trained decision variable selection network to select inaccurate decision node values ​​from the prediction results and scale them to obtain an initial solution, transforming the original large-scale planning problem into a subproblem. Finally, a graph sampling aggregation branch and bound network is used to further determine only the value range of the scaled decision nodes, outputting the final optimized decision node value results.

[0105] As a preferred implementation, in the process of constructing the above decision node prediction model, an open-source integer programming solver is used to solve the integer programming model corresponding to each included angle δ; then the above graph sampling aggregation branch and bound model is constructed in the following way:

[0106] In the process of solving the integer programming model corresponding to each included angle δ, the decision variable selection strategy and branching strategy in the solution process are also obtained as the decision labels of the corresponding tripartite graph; a training sample is determined by the tripartite graph corresponding to the included angle δ and its corresponding decision labels to form the second training sample set;

[0107] The graph sampling aggregation branch bound network is trained using the second training sample set to obtain the graph sampling aggregation branch bound model.

[0108] The network framework of the graph sampling aggregation branch-bound network is similar to that of the graph convolutional network, and the output is the classification distribution on the candidate set C.

[0109] The prediction probability is expressed by the following formula:

[0110] t′ k×j =MLP(v′) k×j ;φ);

[0111]

[0112] In the formula, v′ k×j The output is the node embedding, and φ is the learning parameter of the policy. This distribution is independent of other dimensions conditioned on M.

[0113] A graph sampling aggregation branch-and-bound method is adopted to mimic the solver's branch-and-bound strategy, improving the efficiency for handling the same type of planning problems. By using the solver's node selection methods and branching strategies as samples, the model ensures that it reaches or even exceeds the solver's accuracy. In other words, the graph sampling aggregation branch-and-bound model learns the solver's neural branch probabilities, avoiding the use of the solver itself, and only performs further prediction output on a subset of selected decision nodes, thus improving solution speed while maintaining accuracy.

[0114] As a preferred implementation method, the above decision variable selection model is constructed in the following manner:

[0115] Using a trained decision node prediction model, based on each ternary graph in the first training sample set, the values ​​of each decision node corresponding to the predicted output are predicted. The values ​​of each decision node corresponding to each ternary graph in the predicted output, as well as the labels of whether each decision node is selected or not, are used as training samples to construct a third training sample set. The labels of whether each decision node is selected or not are obtained in the following way: based on the values ​​of each decision node corresponding to the predicted output of the ternary graph and the labels of the decision nodes in the first training sample set, the decision nodes with accurate values ​​and the decision nodes with inaccurate values ​​are selected to form a binary classification vector of whether each decision node is selected or not, which is used as the label of whether each decision node is selected or not in the predicted output of the ternary graph.

[0116] The decision variable selection network is trained using the third training sample set to obtain the decision variable selection model.

[0117] A binary classifier can be trained to determine which variables to predict and which to avoid predicting. For each variable x... k×j Introducing additional output z k×j ∈{0,1}, used to determine whether to assign x k×j Conditionally independent models can be trained using the following loss function:

[0118]

[0119]

[0120] In the formula, C represents the coverage threshold of the expected relative frequency of the assigned variable, ψ is a quadratic penalty term, and λ is the importance hyperparameter for controlling the coverage of the set threshold.

[0121] A conditional independent model is used to perform a secondary screening of the predicted values ​​obtained from convolution. Variables with relatively certain probability values ​​are extracted, while variables with relatively dispersed probability values ​​are selected. The range of values ​​for these variables is then narrowed to facilitate subsequent solution calculations.

[0122] Figure 8 This is a convolutional method for decision node prediction models and graph sampling aggregation branch-bound models based on graph sampling aggregation.

[0123] Step 1: Convolve the constraint node C and the target node Y to aggregate the feature information of the decision variable node X.

[0124] Step 2: Convolve the decision variable node X to aggregate the updated features C′ of the constraint node and the updated features Y′ of the target node.

[0125] Step 3: Connect the updated decision variable feature X′ and constraint node feature C′ to form feature h1, and connect the decision variable feature X′ and target node feature Y′ to form feature h2.

[0126] Step 4: Aggregate features h1 and h2 to obtain h g , as the output of the embedded features.

[0127] Figure 9 The convolutional framework for the decision node prediction model and the branch-bound model based on graph sampling aggregation consists of two graph sampling aggregation convolutional layers, feature merging, two fully connected layers, and finally Softmax regression.

[0128] Figure 10 The multilayer perceptron network for the decision variable selection model consists of three fully connected layers: the first layer is the input for variable selection, the second layer is the hidden layer, and the third layer is the output layer, which outputs the range of the variable selection results.

[0129] It should be noted that, as an optimization of any solution in this embodiment, in the event of abnormal blockage of some nozzles during the printing process, the decision node prediction model modifies the volume feature of the nozzle to 0 to update the nozzle status in real time, improve the algorithm's adaptability, and ensure the system's stability and accuracy.

[0130] In summary, this invention provides a novel algorithm for solving large-scale pixel inkjet printing planning, comprising: 1) a large-scale panel partitioning planning method, which plans the large-scale pixel panel in parallel according to the pixel partitions on the panel, and finally merges the planning data to complete the efficient planning of the large area panel; 2) an integer planning model dataset for inkjet path planning, which constructs the positional relationship between the printhead and the large-size substrate under different substrate placement angles, creates an integer planning model set, and converts it into graph data format as an offline training dataset, with each graph represented as G = (X, C, Y, E); 3) a decision node prediction network based on graph sampling aggregation, which aggregates the feature information of the decision variable point set of the graph with the neighbor constraint point set through a two-layer aggregation network to generate the graph embedding of each decision point, and predicts the value of the variable through a multilayer perceptron; 4) a decision variable selection model, which filters the predicted values ​​through a neural network to select the actual predicted values, and then performs subsequent tightening to reduce the range of unselected variables; 5) a graph sampling aggregation branch and bound network, which learns the solver branch strategy to complete the final solution and generate the inkjet planning result. This invention can be used to improve the efficiency of pattern planning calculation in the manufacturing of large-size inkjet display panels, accelerate the solution of new batches based on information from previous batches, and improve production efficiency.

[0131] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for planning printing on large-size panels, characterized in that, include: A certain installation angle for inkjet printing planning The large panel below is divided into multiple regions, and the division method satisfies the following conditions: the Y-axis deviation between the first and last pixels of the same column arranged along the X-axis in the pixel array of each region is less than a threshold; the included angle This indicates the angle between the line connecting the first and last pixels of the same column arranged along the X direction in a large-size panel and the X direction, which is the direction in which the printhead moves during printing. The printing planning problem of a large-size panel after dividing the area is represented by an integer programming model through mathematical modeling. The printing planning of each partition is regarded as the printing planning of a single row of pixels within that partition. The integer programming model is converted into a ternary graph. The decision nodes of the ternary graph represent the decision variables in the integer programming model, representing the number of times each nozzle is printed at each stop position of the printhead. The constraint nodes of the ternary graph represent the constraint conditions in the integer programming model, representing the volume constraint of each pixel. The target node in the ternary graph represents the target variable in the integer programming model, which represents the maximum number of times all nozzles print at each stop position of the nozzle; Input the three-part graph into the decision node prediction model to obtain the value of each decision node, that is, the number of times each nozzle prints at each stop position of the nozzle, thus completing the large-size panel printing planning method. Specifically, a decision node prediction network is trained using a first training sample set to obtain a decision node prediction model. The decision node prediction network is a graph sampling aggregation-based decision node prediction network. Specifically, it first performs partial aggregation and extraction of node information in the three-part graph: information aggregation is performed on each target node and each constraint node to obtain a first embedding; information aggregation is performed on each decision node based on each first embedding to obtain a second embedding; all second embeddings are merged to obtain a graph embedding; then, an MLP (Multilayer Perceptron) is used to predict the value of each decision variable based on the graph embedding to generate the classification probability of the decision node.

2. The large-size panel inkjet printing planning method according to claim 1, characterized in that, The specific implementation method for dividing a large-size panel into multiple regions is as follows: Determine the angle between the line connecting the first and last pixels of the same column along the X-axis in the large panel to be divided and the X-axis. And the printable Y-axis length within each pixel, which serves as the threshold. Calculate the X-axis length of the region to be divided based on geometric relationships. for: ; Based on the X-axis spacing of adjacent pixels in a large-size panel Calculate the number of rows in the X direction within each partitioned region. : .

3. The large-size panel inkjet printing planning method according to claim 1 or 2, characterized in that, The printable Y-axis length within each pixel The threshold is determined as follows: S1. Take the median of the landing point deviations of all nozzles as the printing deviation of the entire printhead: ； in, and The first The jet speed and angle of nozzle number 1 This represents the total number of nozzles. This indicates the speed at which the large panel moves during printing; This indicates the height of the printhead from the large panel during printing; S2, Obtain pixel Y-axis dimension Calculate the printable Y-axis length within each pixel. for: .

4. The large-size panel inkjet printing planning method according to claim 1, characterized in that, The decision node prediction model is constructed in the following manner: For different included angles The large panel below is divided into multiple areas according to the aforementioned division method; The large-size panel printing planning problem after each region division is represented by an integer programming model through mathematical modeling and converted into a ternary graph form, resulting in a set of ternary graphs. An integer programming solver is then used to solve for each included angle. The corresponding integer programming model yields the values ​​of all decision variables in the model, forming the decision node labels of the corresponding tripartite graph; each tripartite graph and its corresponding decision node labels are used as training samples to obtain the first training sample set. The decision node prediction network is trained using the first training sample set to obtain the decision node prediction model.

5. The large-size panel inkjet printing planning method according to claim 4, characterized in that, After predicting the values ​​of each decision node, the method also includes: The values ​​of each decision node are input into the decision variable selection model, and the decision variable selection model determines the accurate and inaccurate decision node values ​​among the decision node values. Modify the values ​​of each inaccurate decision node to a range smaller than its current value; construct a new decision node value vector from all accurate decision node values ​​and the modified decision node value ranges. The new decision node value vector is input into the graph sampling aggregation branch and bound model. The graph sampling aggregation branch and bound model solves for the decision node values ​​corresponding to the value range of each decision node in the new decision node value vector, determines an accurate value within the corresponding value range, and obtains the final decision node value vector, thus completing the large-size panel inkjet printing planning method.

6. The large-size panel inkjet printing planning method according to claim 5, characterized in that, In constructing the decision node prediction model, an open-source integer programming solver is used to solve for each included angle. The corresponding integer programming model; then the graph sampling aggregation branch and bound model is constructed in the following way: At each angle In the process of solving the corresponding integer programming model, the decision variable selection strategy and branching strategy in the solution process are also obtained as decision labels for the corresponding tripartite graph; With that included angle The corresponding tripartite graph and its corresponding decision label determine a training sample, which constitutes the second training sample set; The graph sampling aggregation branch bound network is trained using the second training sample set to obtain the graph sampling aggregation branch bound model.

7. The large-size panel inkjet printing planning method according to claim 5, characterized in that, The decision variable selection model is constructed in the following manner: Using a trained decision node prediction model, based on each ternary graph in the first training sample set, the values ​​of each decision node corresponding to the predicted output are predicted. The values ​​of each decision node corresponding to each ternary graph in the predicted output, as well as the labels of whether each decision node is selected or not, are used as training samples to construct a third training sample set. The labels of whether each decision node is selected or not are obtained in the following way: based on the values ​​of each decision node corresponding to the predicted output of the ternary graph and the labels of the decision nodes in the first training sample set, the decision nodes with accurate values ​​and the decision nodes with inaccurate values ​​are selected to form a binary classification vector of whether each decision node is selected or not, which is used as the label of whether each decision node is selected or not in the predicted output of the ternary graph. The decision variable selection network is trained using the third training sample set to obtain the decision variable selection model.

Images