Thin film packaging ink-jet printing process optimization method

Abstract

The invention belongs to the related technical field of thin film packaging ink-jet printing, and particularly relates to a thin film packaging ink-jet printing process optimization method which comprises the following steps: initializing process parameters, and dividing a thin film packaging ink-jet printing process into four core stages of liquid drop jetting, liquid drop flying, liquid drop deposition and UV curing, through multi-source data fusion and data-driven model training, coupling characteristics of four stages of modeling are realized, data-driven models of all the stages are connected in series, through an end-to-end process link, liquid drop movement and a film forming process are accurately predicted, and collaborative optimization of process parameters is effectively completed in combination with a multi-target optimization model. According to the method, the optimization efficiency can be remarkably improved, and the method can be widely applied to ink-jet printing manufacturing process optimization and quality control in the fields of flexible electronics, organic photoelectricity and the like.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of thin film packaging inkjet printing, and more particularly relates to a thin film packaging inkjet printing process optimization method. BACKGROUND

[0002] As a non-contact printing technology, inkjet printing technology has the advantages of high material utilization, no need for mask, low cost, and being more suitable for large-size and high-precision OLED display device preparation compared with traditional evaporation process. Inkjet printing technology is the main development direction of future new display manufacturing technology.

[0003] The inkjet printing process for thin film manufacturing involves multiple stages, including droplet ejection, droplet flight, droplet deposition and solidification, and has characteristics such as complex process mechanism, difficult manufacturing process detection, and many factors affecting thin film quality. The main process parameters involved in the thin film manufacturing process include voltage amplitude, pulse time, voltage pulse rate, droplet volume, flow field flow rate, temperature, substrate contact angle, droplet radius, droplet spacing, UV power, and light exposure time. These process parameters have a significant impact on thin film manufacturing performance. The main parameters for evaluating thin film performance include film thickness uniformity and edge straightness. In the thin film manufacturing process, there is an interaction between process parameters at different stages. How to achieve coordinated optimization of process parameters between different stages is crucial for preparing high-quality thin films.

[0004] With the wide application of inkjet printing thin film manufacturing technology, the demand for large-size thin film manufacturing is increasing. In current industrial production, process parameters are selected based on personal experience and experiments, and process parameters are manually adjusted until the requirements are met. This method is time-consuming and labor-intensive, wastes printing materials, and it is difficult to obtain the best thin film manufacturing process parameters. Therefore, it is necessary to comprehensively consider the process parameters in the inkjet printing process of thin film manufacturing, and to propose a thin film packaging inkjet printing process coordinated optimization method and system. SUMMARY

[0005] In view of the above defects or improvement needs of the prior art, the present application provides a thin film packaging inkjet printing process optimization method, which aims to efficiently and coordinately adjust the process parameters between different printing stages in the thin film manufacturing process, and realize the high-quality preparation of thin films.

[0006] To achieve the above-mentioned purpose, according to one aspect of the present application, a thin film packaging inkjet printing process optimization method is provided, comprising: S1, initializing process parameters; S2, input the voltage amplitude, pulse time, voltage trapezoidal wave interval time and voltage pulse rate in the process parameters into the droplet ejection stage data-driven model constructed by pre-training, simulate the droplet ejection process by the model, and predict the output droplet volume; input the predicted output droplet volume and the flow field flow rate, flow field temperature and flow field pressure in the process parameters into the droplet flight stage data-driven model constructed by pre-training, simulate the flight trajectory of the droplet in the flow field by the model, and predict the output droplet landing point accuracy; input the predicted output droplet landing point accuracy, droplet volume and the substrate contact angle and target droplet spacing in the process parameters into the droplet deposition stage data-driven model constructed by pre-training, simulate the droplet deposition morphology by the model, and predict the output wet film morphology grayscale image and wet film thickness; input the predicted output wet film morphology grayscale image and the UV power, illumination time and radiation intensity in the process parameters into the UV curing stage data-driven model constructed by pre-training, simulate the energy transfer and chemical reaction in the curing process by the model, and predict the output cured film morphology; wherein during training, multi-source data fusion is performed between the training sample data required by each data-driven model, including time sequence alignment, space matching and cross-stage process parameter fusion; S3, using a multi-objective Bayesian optimization algorithm, based on the cured film morphology, taking the film thickness distribution error and edge straightness error as the minimization target, adjusting the process parameters of each stage, and re-executing S2.

[0007] Further, the droplet ejection stage data-driven model includes a data processing unit and a Bi-LSTM model; wherein the data processing unit is used to extract high-frequency features in the voltage amplitude and voltage pulse rate respectively, and the Bi-LSTM model is used to simulate the droplet ejection process based on the pulse time, voltage trapezoidal wave interval time and the extracted high-frequency features in the voltage amplitude and voltage pulse rate, and output the droplet volume and initial flight speed; In the loss function used for training the droplet ejection stage data-driven model, a regularization term is configured, and the regularization term adopts a simplified form of the Navier-Stokes equation to constrain the droplet volume conservation.

[0008] Further, the droplet flight stage data-driven model adopts a graph neural network, wherein an air resistance equation is configured in the message passing function of the graph neural network to simulate the influence of air resistance on the droplet falling process. When training the droplet flight stage data-driven model, the minimum deviation between the actual landing point and the set target landing point is taken as the optimization target to optimize the trajectory simulation accuracy of the graph neural network in the droplet flight process.

[0009] Further, the droplet deposition stage data-driven model adopts a conditional generative adversarial architecture, wherein the generator in the conditional generative adversarial architecture adopts a U-Net network; the skip connection layer of the U-Net network connects the condition vector obtained by encoding the input of the driving model.

[0010] Further, the UV curing stage data-driven model adopts a convolutional neural network, which includes an input layer, a convolutional layer, a self-attention layer and a fully connected layer, wherein the self-attention mechanism is used to capture long-range spatial dependencies; the UV curing stage data-driven model first converts the received wet film topography grayscale image into a voxelized 3D network, and then simulates the energy transfer and chemical reaction in the curing process based on the voxelized 3D network and the UV power, illumination time and radiation intensity through the convolutional neural network.

[0011] Further, the global objective function corresponding to the minimization of the film thickness distribution error and the edge straightness error is:

[0012] In the formula, is the global objective function value; is a dynamic weight coefficient; is the edge curvature, which is used to quantify the edge straightness error; N represents the number of edge segments; is the predicted film thickness vector; is the target film thickness vector; represents the L2 norm difference between and represents the profile height of the m-th edge segment; represents the transverse coordinate position of the m-th edge segment; Therefore, when adjusting the process parameters of each stage, the gradients of the process parameters with respect to the optimization objectives are calculated by automatic differentiation, and backpropagation is performed.

[0013] Further, the dynamic time warping algorithm is used for time series alignment, and the iterative closest point algorithm is used for spatial matching of multi-source spatial data.

[0014] Further, the implementation manner of the cross-stage process parameter fusion is: The process parameters required for the input of each model are taken as a type of modal data, and a total of four types of modal data are divided; the VAE model including an encoder and a decoder is trained using the four types of modal data, and after the training is completed, the encoder in the VAE model is used to map each modal data to the latent space respectively, to obtain a feature vector of a unified dimension, which is used as the input required for the training of each model.

[0015] ​​​​Furthermore, the multi-source data fusion also includes parameter importance analysis, which is implemented as follows: A graph attention network is used to establish a process chain relationship graph, in which a data-driven model of four process stages is used as nodes. The parameter transfer relationships between process stages are represented as edges in the graph, with edge weights. Representation phase The output parameters of the stage The strength of the influence of the input parameters; Through attention mechanisms, computation nodes Its neighboring nodes Attention weights between This indicates the difference in importance of different node features in the association analysis; among them,

[0016] In the formula, For nodes The feature vector of each node represents the information of that node, including the input parameters and prediction output information of the data-driven model of the corresponding process stage. For attention parameters; For nodes The neighbor set, containing only nodes The next stage node corresponding to the process stage; For activation functions; Based on the attention mechanism weights, aggregate neighbor features and update the node feature vector: In the formula, The weight matrix is ​​formed by... All corresponding weights constitute, As the activation function, the node feature vector is updated iteratively through multiple layers of GAT to capture the nonlinear correlation between stages; Node feature similarity is obtained by measuring the similarity between node feature vectors, based on attention weights. Based on the similarity of node features, the parameter transfer strength between each process stage is calculated to form an association strength matrix, which is used to quantify the coupling relationship of multi-stage processes. The correlation strength matrix, actual printing process parameters at each stage, and actual printing result parameters are used as input features of the SHAP analysis model. By calculating the SHAP value of each process parameter, the influence of various process parameters on the global objective is obtained, and key influencing parameters are identified to guide the targeted adjustment of process parameters.

[0017] According to another aspect of the present application, a thin film encapsulation inkjet printing process optimization device is provided, characterized by comprising a memory and a processor, the memory stores a computer program, and the processor implements the steps of the method as described above when executing the computer program.

[0018] Overall, compared with the prior art, the technical solution provided by the present application has the following beneficial effects: The present application provides a thin film encapsulation inkjet printing process optimization method, which realizes the synergistic optimization of process parameters between different stages. Specifically, the thin film encapsulation inkjet printing process is divided into four core stages of droplet ejection, droplet flight, droplet deposition and UV curing. Through multi-source data fusion and data-driven model training, the coupling features of the four stages are modeled. Through the end-to-end process link, the droplet motion and film forming process are accurately predicted. Combined with the multi-objective optimization model, the process parameter synergistic optimization is effectively completed. The method of the present application significantly improves the optimization efficiency, and introduces the Bayesian optimization strategy, which reduces the number of optimization iterations compared with the traditional algorithm, and improves the development efficiency. The method of the present application supports model transfer learning and environment transfer because it is a data-driven model, and is suitable for different ink materials, substrate types and printing paths. It can be widely used in the fields of flexible electronics, organic optoelectronics and other fields of inkjet printing manufacturing process optimization and quality control, and has good industrial popularization value. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A thin film encapsulation inkjet printing process optimization method flow chart is provided for the embodiments of the present application. Figure 2 A thin film encapsulation inkjet printing process optimization method flow chart is provided for the embodiments of the present application. Figure 3 A multi-source data fusion implementation schematic diagram is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0021] Embodiment one A thin film encapsulation inkjet printing process optimization method, as shown in Figure 1 and Figure 2 , comprising: S1, initializing process parameters; S2, input the voltage amplitude, pulse time, voltage trapezoidal wave interval time and voltage pulse rate in the process parameters into the droplet ejection stage data-driven model constructed by pre-training, simulate the droplet ejection process by the model, and predict the output droplet volume; input the predicted output droplet volume and the flow field flow rate, flow field temperature and flow field pressure in the process parameters into the droplet flight stage data-driven model constructed by pre-training, simulate the flight trajectory of the droplet in the flow field by the model, and predict the output droplet landing point accuracy; input the predicted output droplet landing point accuracy, droplet volume and the substrate contact angle and target droplet spacing in the process parameters into the droplet deposition stage data-driven model constructed by pre-training, simulate the droplet deposition morphology by the model, and predict the output wet film morphology gray scale image and wet film thickness; input the predicted output wet film morphology gray scale image and the UV power, illumination time and radiation intensity in the process parameters into the UV curing stage data-driven model constructed by pre-training, simulate the energy transfer and chemical reaction in the curing process by the model, and predict the output cured film morphology; wherein during training, multi-source data fusion is performed between the training sample data required by each data-driven model, including time sequence alignment and space matching; S3, using a multi-objective Bayesian optimization algorithm, based on the cured film morphology, taking the film thickness distribution error and edge straightness error as the minimization target, adjusting the process parameters of each stage, and re-executing S2.

[0022] Since some data cannot be obtained by measurement, it needs to be obtained by simulation, and the amount of data obtained by measurement is not enough for the construction of a data-driven model, and also needs to be obtained by simulation, which requires the alignment of the measured and simulated data in time and space. In addition, the data-driven models of different stages need to be trained independently, while the present embodiment needs to combine the four-stage data-driven models to form a full-life-cycle simulation, so the training data of the four-stage models need to be aligned. Therefore, in order to generate a low-difference sampling point set in a high-dimensional parameter space, the data-driven model required by the present embodiment method needs to perform multi-source data fusion between the training sample data required by each data-driven model during training, including time sequence alignment and space matching. This scheme performs multi-source data alignment and shared feature learning on the measured data and full-life-cycle simulation data.

[0023] In a specific implementation, a time alignment method (e.g., dynamic time warping, DTW) is used to eliminate time deviations caused by different sampling frequencies, trigger delays, and recording jitter, establish a consistent time axis across devices and stages; a spatial and geometric registration method (e.g., iterative closest point, ICP) is used to unify the coordinate system and reference benchmark, so that the data of each stage of jetting, flying, depositing, and solidifying correspond in the same spatial framework; a feature learning method (e.g., variational autoencoder, VAE) is used to extract low-dimensional and distinguishable shared features, realize the comparability and fusion basis of different stages and different types of data, generate a low-difference sampling point set in a high-dimensional parameter space, and use a correlation coefficient matrix to ensure the orthogonality of the sampling points.

[0024] End-to-end process link, the output of the previous stage as the input of the next stage: in the droplet jetting stage, the voltage amplitude, pulse time, interval time, and voltage pulse rate are input, the droplet volume and droplet speed are output through the constructed droplet jetting stage data-driven model, the droplet volume, flow field flow rate, and flow field temperature are input, the droplet landing point accuracy is output through the analysis of air resistance and flow field characteristics, and the position accuracy of the printed pattern is ensured. Through the constructed droplet deposition stage data-driven model, based on the substrate contact angle, droplet radius, and droplet spacing, the wet film thickness is output, providing a uniform initial film layer for UV curing. Through the constructed UV curing stage data-driven model, the UV power, radiation intensity, and illumination time are input, and the curing rate, curing uniformity, and film morphology are output through energy transfer and chemical reaction kinetics analysis.

[0025] The method of the embodiment connects the data-driven models of the four stages end to end, optimizes the parameters through end-to-end joint optimization and parameter optimization, constructs a multi-objective optimization model, takes the film thickness distribution error and edge straightness error as global optimization objectives, uses a multi-objective Bayesian optimization algorithm and automatic differentiation technology to jointly back-propagate gradient information, outputs a Pareto parameter set and its recommended process parameter settings in different production scenarios, completes the collaborative optimization of process parameters, and realizes the optimization and quality control of inkjet printing film manufacturing process.

[0026] In a specific implementation, the global objective function is defined as:

[0027] In the formula, α , β is a dynamic weight coefficient, is the edge curvature, which is used to quantify the straightness error, N represents the number of edge division segments, the gradient of the process parameters on the optimization objective is calculated through automatic differentiation, and the gradient is back-propagated.

[0028] The process parameter collaborative iterative optimization obtains available gradient information combined with automatic differentiation, or provides function values by a proxy model at non-differentiable points; a next batch of candidate parameter combinations is generated by a multi-objective acquisition strategy (such as an expected hypervolume improvement, etc.), experiments and simulation evaluations are performed, and data is backfilled; the proxy model and the acquisition strategy are iteratively updated until a convergence condition is reached; and a Pareto parameter set and recommended process parameter settings in different production scenarios are output, completing the collaborative optimization of process parameters.

[0029] The method of the embodiment realizes collaborative optimization of process parameters between different stages. Specifically, the thin film encapsulation inkjet printing process is divided into four core stages of droplet ejection, droplet flight, droplet deposition, and UV curing. Through multi-source data fusion and data-driven model training, the coupling features of the four stages are modeled. Through an end-to-end process link, the droplet motion and film formation process are accurately predicted. Combined with a multi-objective optimization model, the process parameter collaborative optimization is effectively completed. The embodiment significantly improves the optimization efficiency. By introducing automatic differentiation and Bayesian optimization strategies, the number of optimization iterations is reduced compared to traditional algorithms, improving development efficiency. Because the data-driven model is constructed, model transfer learning and environment transfer are supported, and the process environment changes such as different ink materials, substrate types, and printing paths are adapted. The method can be widely applied to inkjet printing process optimization and quality control in the fields of flexible electronics and organic optoelectronics, and has good industrial promotion value.

[0030] Before the method of the embodiment is implemented, process stage modeling is performed in advance. The inkjet printing process is divided into four core stages of droplet ejection, droplet flight, droplet deposition, and UV curing, and the process parameters of each stage are determined. Data-driven models are constructed for the physical characteristics and parameter coupling relationships of each stage. Data acquisition and dataset construction are required as a prerequisite for modeling. In the actual printing process, real-time acquisition of process parameter inputs is performed, and offline measurement and simulation are supplemented to construct the dataset.

[0031] In a specific implementation, the input parameters of each stage are obtained from the equipment and sensing system and the timestamps are recorded, and dimension unification and sensor calibration are completed. High-speed cameras are used to obtain droplet ejection process images, and pixel scale calibration and contour extraction are performed on the images to calculate the droplet volume. White light interference is used for film thickness measurement to obtain the wet film thickness and its spatial distribution. At the same time, full life cycle simulation is performed. Variables that cannot be directly measured (for example, the inkjet printing process belongs to the field of micro-nano manufacturing, and the ejected droplets are micron-sized, so even if a high-speed camera is used, the accurate trajectory changes of the droplets during flight cannot be accurately captured; during the UV curing process, sensors cannot be arranged inside the equipment to collect data) or have high measurement costs are obtained as supplementary simulation data, abnormal point rejection and missing value processing are completed, and the dataset is constructed.

[0032] As a preferred implementation, the droplet ejection phase data-driven model comprises a data processing unit and a Bi-LSTM model; wherein the data processing unit is configured to extract high-frequency features in the voltage amplitude and the voltage pulse rate, respectively, and the Bi-LSTM model is configured to simulate the droplet ejection process based on the pulse time, the voltage trapezoidal wave interval time, and the extracted high-frequency features in the voltage amplitude and the voltage pulse rate, and output the droplet volume and the initial flight speed. In the loss function used for training the droplet ejection phase data-driven model, a regularization term is configured, and the regularization term adopts a simplified form of the Navier-Stokes equation to constrain the droplet volume conservation.

[0033] During the training process, the network weights of the model can be updated continuously by combining the measured data and the simulation data, and the model outputs the droplet volume and the initial flight speed.

[0034] As a preferred implementation, the droplet flight phase modeling discretizes the time-varying flow field into a dynamic graph structure, where each node attribute includes flow field velocity, flow field temperature, flow field pressure (preset parameters), etc. A graph neural network is used to simulate the flight trajectory of the droplet in the flow field, and an air resistance equation is introduced in the message passing function to simulate the influence of air resistance on the droplet falling process. During model training, the droplet landing point accuracy (minimizing the deviation between the actual landing point and the set target landing point) is used as the optimization target to optimize the graph neural network for trajectory simulation accuracy during the droplet flight process. Through iterative calculation, the ability of the graph neural network to accurately output the droplet landing point accuracy is trained based on the input parameters of the droplet volume, flow field velocity, flow field temperature, and flow field pressure.

[0035] As a preferred implementation, the droplet deposition phase data-driven model adopts a conditional generative adversarial architecture, wherein the generator in the conditional generative adversarial architecture adopts a U-Net network; the skip connection layer of the U-Net network connects the condition vector obtained by encoding the input of the driving model.

[0036] In a specific implementation, a hybrid architecture of a U-Net convolutional neural network and a conditional generative adversarial network is constructed, the substrate contact angle, the droplet radius, and the spacing are encoded into a condition vector, which is spliced to the U-Net skip connection layer, the generator outputs a wet film topography grayscale image, and the discriminator constrains the topography authenticity. The wet film thickness in the wet film topography grayscale image (i.e., the spatial distribution of the film thickness) is calculated by pixel integration, and cross-validated with the law of mass conservation to ensure that the model output wet film thickness is accurate and reliable, providing a uniform initial film layer for the UV curing stage.

[0037] The generator and the discriminator are both part of the conditional generative adversarial network framework, the generator generates the topography image, and the discriminator judges its authenticity, so as to continuously improve the fidelity and physical rationality of the topography generated by the generator through adversarial training.

[0038] As a preferred embodiment, the UV curing stage data-driven model adopts a convolutional neural network, which includes an input layer, a convolutional layer, a self-attention layer, and a fully connected layer, wherein the self-attention mechanism is used to capture long-range spatial dependencies; the UV curing stage data-driven model first converts the received wet film topography grayscale image into a voxelized 3D network, and then simulates the energy transfer and chemical reactions in the curing process based on the voxelized 3D network and the UV power, illumination time, and radiation intensity through the convolutional neural network. In the training process, multiple output targets are optimized simultaneously using multi-task learning to improve the overall performance of the model.

[0039] The various ways of multi-source data fusion are described below.

[0040] (1) Regarding the aforementioned time alignment, as shown in Figure 3 , for the measured data in the inkjet printing process: such as droplet ejection voltage timing (pulse time, interval time), UV curing power changes and other parameters, the dynamic time warping algorithm is used for time alignment.

[0041] Specifically, first, data preprocessing is performed, and the voltage pulse signal (voltage amplitude, pulse time, interval time, and voltage pulse rate, sequence ) of the droplet ejection stage and the flow field flow rate change data (sequence ) of the droplet flight stage are standardized to eliminate dimensional differences. The standardization formula is: , where is the standardized data, , are the maximum and minimum values of the sequence , respectively.

[0042] After data standardization, a distance matrix is constructed to calculate the Euclidean distance between two data points in each sequence, and the distance matrix of the two sequences is constructed , where m and n are the number of data points in the two sequences, and the element . The warping path is solved, and the cumulative distance of the warping path is minimized: the optimal warping path is found by dynamic programming, which satisfies the boundary conditions , and , , so that the cumulative distance of the warping path is minimized: .

[0043] Finally, interpolation or downsampling is performed on the two current sequences according to the optimal normalization path to make the measured data correspond to the simulation data in the time dimension, ensuring the consistency of the time sequence relationship of process parameters such as droplet jet pulse and droplet volume change, UV irradiation time and curing rate change.

[0044] (2) Regarding spatial matching, for spatial coordinate differences such as substrate position offset and landing point pose deviation, the iterative nearest point algorithm can be used to achieve accurate matching of multi-source spatial data.

[0045] Specifically, the measured point cloud of droplet landing point coordinates is acquired across the entire substrate: With simulated point clouds ,in , representing the measured and simulated landing points in three-dimensional space, respectively. Then, for each ,pass Nearest neighbor algorithm search The closest point in the middle To form an initial set of corresponding point pairs Calculate the optimal transformation matrix based on corresponding point pairs. ,in Let be a rotation matrix. Let be the translation vector, and minimize the objective function: The above optimization problem is solved using singular value decomposition to obtain the rotation matrix. With translation vector .

[0046] Then iterative optimization is performed, applying the transformation matrix. For measured point cloud Perform the transformation and update the point cloud coordinates. Find the corresponding point pairs again and iteratively calculate the transformation matrix until the mean square error of two iterations is found. The convergence condition is met: ,in To preset the accuracy threshold, it is usually taken as... This is to ensure the accuracy of spatial coordinate correction, such as substrate position and landing point pose.

[0047] (3) Regarding the fusion of process parameters across stages, a shared feature space can be constructed using a variational autoencoder, which can realize the fusion of data between different stages and the fusion of measured data and simulation data.

[0048] Specifically, the first step is to perform data modality segmentation, dividing the preset parameter data required for model input in the dataset into four modalities: electrical signal data during the droplet ejection stage. (Voltage amplitude, pulse duration, etc.), flow field data during the droplet flight phase (Flow rate, temperature, etc.) Geometric data of the deposition stage (Substrate contact angle), energy data during the curing stage (Radiation intensity, duration of illumination, etc.)

[0049] Next, a VAE model containing an encoder and a decoder is constructed. The encoder is used to map the modal data to a shared latent space. ,in The encoder outputs mean and variance; the decoder uses the latent variables... The input data is reconstructed, and the loss function is defined as follows: ,in To reconstruct the loss, For KL divergence regularization, This is the balance coefficient.

[0050] After VAE training is completed, the encoder maps the data of each modality to the latent space. To obtain feature vectors of uniform dimension This enables the comparability of data from different modalities.

[0051] As a preferred implementation, multi-source data fusion also includes parameter importance analysis, using graph attention networks to establish a process chain relationship graph, with each stage model as a node and the transmission relationship between parameters as an edge, and performing feature importance attribution and sensitivity analysis on the output of the trained model, such as the SHAP value analysis method, to guide the directional adjustment of process parameters.

[0052] (4) Regarding parameter correlation analysis, a graph attention network is used to establish a process chain relationship diagram to quantify the correlation strength between process parameters.

[0053] Specifically, a graph structure is constructed, using the four process stage models as nodes. These correspond to the spraying, flight, deposition, and solidification stages, respectively, with the transfer relationships between process parameters serving as edges. edge weight Representation phase The output parameters of the stage The strength of the influence of the input parameters.

[0054] Through attention mechanisms, computation nodes Its neighboring nodes Attention weights between This indicates the difference in importance of different node features in association analysis:

[0055] in For nodes The feature vectors of each node represent the node's information, including the input process parameters and the stage model prediction output information for each stage. For learnable attention parameters, For nodes The neighborhood group, The activation function is used to calculate the nodes. The attention weights between nodes and their neighbors reflect the differences in the importance of different node features in association analysis.

[0056] Based on the attention mechanism weights, aggregate neighbor features and update the node feature vector: ,in The weight matrix is ​​composed of nodes. All corresponding weights constitute, For activation function, The neighbor of node i is represented by the neighbor corresponding to the next stage. The node features are updated iteratively through multiple layers of GAT to capture the nonlinear correlation between stages.

[0057] Node feature similarity is obtained by measuring the similarity between node feature vectors, based on attention weights. Based on the similarity of node features, the parameter transfer strength between each process stage is calculated to form an association strength matrix, which is used to quantify the coupling relationship of multi-stage processes.

[0058] Next, the parameter boundary effect is analyzed: the SHAP value is used to quantify the impact of each process parameter on the global optimization objective. Its SHAP value is defined as: ,in For parameter set, Its subset For the global objective function, the expectation operator This represents the average contribution across all combinations of parameter subsets. Based on the ranking of SHAP values, key influencing parameters are identified to guide the targeted adjustment of process parameters.

[0059] The GAT correlation strength matrix characterizes the correlation strength between process stages (or parameters), while SHAP value analysis explains the contribution of a particular input parameter to the final prediction result. The two are related: the correlation strength matrix obtained from GAT quantification, various process parameters input at each stage, and the printed result parameters all serve as the SHAP analysis model. The input features, the introduction of the SHAP value method, and the model The input features (i.e., the process parameters modeled by the GAT association) are decomposed to determine the contribution of each parameter to the global target, and the SHAP value of each parameter is calculated to determine the key influence parameters, which are used to guide the directional adjustment of the process parameters, whether in the training sample data set construction stage or in the parameter adjustment link in the parameter optimization stage.

[0060] Through the above four exemplary fusion methods, the spatiotemporal alignment, cross-modal fusion and parameter association analysis of multi-source data in the thin film packaging inkjet printing process can be realized, ensuring the input-output matching of each stage model and the accuracy of the optimization process.

[0061] According to the scheme of the embodiment, a thin film packaging inkjet printing process optimization system can be designed to execute the thin film packaging inkjet printing process optimization method described in embodiment one, which includes a segmented modeling module, a data acquisition module, a multi-source data fusion module, a joint optimization controller, an adaptive learning unit and a computer control unit.

[0062] Specifically, the segmented modeling module is used to implement the droplet ejection, flight, deposition and UV curing stage models and perform the prediction tasks of each stage model; the data acquisition module is used to collect real-time input process parameters and measurement data and simulation data to form a data set; the data fusion module integrates the DTW algorithm, the ICP algorithm, the VAE and the graph attention network for multi-source data alignment and feature fusion; the joint optimization controller has a multi-objective Bayesian optimization algorithm and an automatic differentiation backpropagation mechanism, and uses the multi-objective Bayesian optimization algorithm to perform full-process collaborative optimization through the four-stage data-driven models; the computer control unit drives the inkjet printing equipment through control signals, receives the optimization parameters output by the joint optimization controller, converts them into control instructions executable by the equipment, and realizes model-driven parameter closed-loop control, thereby completing the collaborative optimization of the thin film packaging inkjet printing process.

[0063] Embodiment two A thin film packaging inkjet printing process optimization device includes a memory and a processor, the memory stores a computer program, and the processor executes the computer program to realize the steps of the method described in embodiment one.

[0064] The related technical solutions are the same as above, and will not be repeated here.

[0065] Those skilled in the art will readily understand that the above description is only of the preferred embodiments of the present application and is not intended to limit the present application, and any modifications, equivalent replacements and improvements made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method of optimizing a thin film encapsulation inkjet printing process, characterized in that, The method comprises the following steps: S1, initializing process parameters; S2, inputting the voltage amplitude, pulse time, voltage trapezoidal wave interval time and voltage pulse rate in the process parameters into a droplet ejection stage data-driven model constructed by pre-training, simulating the droplet ejection process by the model, and predicting the output droplet volume; inputting the predicted output droplet volume and the flow field flow rate, flow field temperature and flow field pressure in the process parameters into a droplet flight stage data-driven model constructed by pre-training, simulating the flight trajectory of the droplet in the flow field by the model, and predicting the output droplet landing point accuracy; inputting the predicted output droplet landing point accuracy, droplet volume and substrate contact angle and target droplet spacing in the process parameters into a droplet deposition stage data-driven model constructed by pre-training, simulating the droplet deposition morphology by the model, and predicting the output wet film morphology grayscale image and wet film thickness; inputting the predicted output wet film morphology grayscale image and UV power, illumination time and radiation intensity in the process parameters into a UV curing stage data-driven model constructed by pre-training, simulating the energy transfer and chemical reaction in the curing process by the model, and predicting the output cured film morphology; During training, multi-source data fusion is performed between the training sample data required by each data-driven model, including time sequence alignment, space matching and cross-stage process parameter fusion; S3, using a multi-objective Bayesian optimization algorithm, based on the cured film morphology, minimizing the film thickness distribution error and edge straightness error as the target, adjusting the process parameters of each stage, and re-executing S2 until the termination condition is reached.

2. A method of optimizing a thin film encapsulated inkjet printing process as claimed in claim 1, wherein, The droplet ejection stage data-driven model comprises a data processing unit and a Bi-LSTM model; wherein the data processing unit is used to extract high-frequency features in the voltage amplitude and voltage pulse rate respectively, and the Bi-LSTM model is used to simulate the droplet ejection process based on the pulse time, voltage trapezoidal wave interval time and the extracted high-frequency features in the voltage amplitude and voltage pulse rate, and output the droplet volume and initial flight speed; In the loss function used for training the droplet ejection stage data-driven model, a regularization term is configured, and the regularization term adopts a simplified form of the Navier-Stokes equation to constrain the droplet volume conservation.

3. A method of optimizing a thin film encapsulated inkjet printing process as claimed in claim 1, wherein, The droplet flight stage data-driven model adopts a graph neural network, wherein an air resistance equation is configured in the message passing function of the graph neural network to simulate the influence of air resistance on the droplet falling process; During training of the droplet flight stage data-driven model, the minimum deviation between the actual landing point and the set target landing point is used as the optimization target to optimize the trajectory simulation accuracy of the graph neural network in the droplet flight process.

4. A method of optimizing a thin film encapsulated inkjet printing process as recited in claim 1, wherein, The droplet deposition stage data-driven model adopts a conditional generative adversarial architecture, wherein the generator in the conditional generative adversarial architecture adopts a U-Net network; the skip connection layer of the U-Net network connects the condition vector obtained by encoding the input of the data-driven model.

5. A thin film encapsulated inkjet printing process optimization method as claimed in claim 1, wherein, The UV curing stage data-driven model adopts a convolutional neural network, and the network structure comprises an input layer, a convolutional layer, a self-attention layer and a fully connected layer, wherein the self-attention mechanism is used to capture long-range spatial dependencies; The UV curing stage data-driven model first converts the received wet film topography gray image into a voxelized 3D network, and then simulates the energy transfer and chemical reaction in the curing process based on the voxelized 3D network and the UV power, illumination time and radiation intensity through a convolutional neural network.

6. A thin film encapsulated inkjet printing process optimization method as claimed in claim 1, wherein, The global objective function corresponding to the minimum film thickness distribution error and edge straightness error is: wherein, is a global objective function value; is a dynamic weight coefficient; is an edge curvature for quantifying edge straightness error; N represents the number of edge segments; is a predicted film thickness vector; is a target film thickness vector; represents the L2 norm difference between and represents the profile height of the th edge segment; represents the lateral coordinate position of the th edge segment; When adjusting the process parameters of each stage, the gradient of the process parameters to the optimization target is calculated by automatic differentiation, and back propagation is performed.

7. A method of optimizing a thin film encapsulated inkjet printing process according to any one of claims 1 to 6, wherein The dynamic time warping algorithm is used for time alignment, and the iterative closest point algorithm is used for spatial matching of multi-source spatial data.

8. A method of thin film encapsulation inkjet printing process optimization according to any one of claims 1 to 6, wherein, The implementation mode of the cross-stage process parameter fusion is: The process parameters required for inputting each model are taken as a type of modal data, and a total of four types of modal data are divided; the VAE model including an encoder and a decoder is trained using the four types of modal data, after the training is completed, the encoder in the VAE model is used to map each modal data to the hidden space respectively to obtain the feature vectors of the unified dimension as the input required for training each model.

9. A method of optimizing a thin film encapsulated inkjet printing process according to any one of claims 1 to 6, wherein, The multi-source data fusion also includes parameter importance analysis, and the implementation mode is: A graph attention network is used to establish a process chain relationship graph, wherein four process stage data-driven models are taken as nodes , a parameter transmission relationship between process stages is taken as an edge in the graph, and an edge weight represents an influence strength of an output parameter of a stage on an input parameter of a stage ; Through the attention mechanism, the computing node The attention weight between the computing node and its neighbor nodes , indicating the difference in the importance of different node features in the correlation analysis; wherein,​ In the formula, is a feature vector of a node , and the feature vector of each node represents information of the node, including input parameters and predicted output information of the process stage data-driven model corresponding to the node; is an attention parameter; is a neighbor set of a node , and only contains nodes corresponding to the next stage node of the process stage; is an activation function; According to the attention mechanism weight aggregation neighbor feature, the node feature vector is updated: , wherein, is a weight matrix, which is composed of corresponding to all weights , is an activation function, and the node feature vector is updated through multi-layer GAT iteration to capture the nonlinear correlation between stages. The similarity between the node feature vectors is measured to obtain a node feature similarity Based on the node feature similarity and the attention weight A parameter transmission strength between the process stages is calculated to form a correlation strength matrix, which is used to quantify the coupling relationship of the multi-stage process. The correlation strength matrix, the actual printing process parameters and the actual printing result parameters of each stage are taken as the input features of the SHAP analysis model, the SHAP value of each process parameter is calculated to obtain the influence of each process parameter on the global objective, and the key influence parameters are determined to guide the directional adjustment of the process parameters.

10. A thin film encapsulation inkjet printing process optimization apparatus, characterized by, The memory stores a computer program, and the processor implements the steps of the method of any one of claims 1-9 when executing the computer program.

Images