A machine learning-based collaborative optimization method for high-capacity MLCC electrode paste
By constructing a multi-source heterogeneous database and machine learning model, combined with multi-objective optimization algorithms and physical information constraints, the problems of long development cycle and high cost of MLCC electrode paste were solved, realizing efficient and intelligent paste formulation and sintering process design, and improving the electrical performance and reliability of multilayer ceramic capacitors.
Patent Information
- Application Number
- CN202611127534.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-28
- Publication Date
- 2026-11-10
AI Technical Summary
In the existing technology, the development and process optimization of electrode paste for multilayer ceramic capacitors (MLCCs) rely on trial and error, resulting in long R&D cycles, high costs, difficulty in obtaining the global optimal solution, and difficulty in achieving the co-firing matching of the inner electrode and the ceramic dielectric, as well as the densification sintering and solderability of the outer electrode.
By constructing a multi-source heterogeneous database, a machine learning prediction model is used to establish the mapping relationship between slurry composition and process parameters to performance. Combined with multi-objective optimization algorithms and physical information constraints, global optimization is performed to obtain the Pareto optimal solution set. Furthermore, the training dataset is expanded by generating adversarial networks to achieve synergistic optimization of slurry formulation and sintering process.
Significantly shortens the R&D cycle, reduces costs, enables submicron internal electrode stacking and high-layer stacking, improves electrode density and solderability, enhances electrical performance and yield, ensures that model prediction results conform to physical laws, and strengthens the robustness and reliability of optimization schemes.
Smart Images

Figure CN123072576A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of electronic component material preparation and intelligent manufacturing technology, and more specifically, to a collaborative optimization method for high-capacity MLCC electrode paste based on machine learning. Background Technology
[0002] Multilayer ceramic capacitors are fundamental components of electronic devices, and their development trend is towards smaller size and larger capacitance. To achieve this goal, the thickness of the internal electrode needs to evolve from the micrometer level to the submicrometer level in order to stack more dielectric layers; at the same time, the co-firing compatibility between the internal electrode and the ceramic dielectric, as well as the densification sintering and solderability of the external electrode, have become core technical challenges.
[0003] In existing technologies, the development and process optimization of electrode pastes (including inner electrode nickel paste and outer electrode copper paste) mainly rely on trial and error. Researchers need to conduct a large number of experiments to adjust the particle size of metal powders, the ratio of organic carriers, and the content of additives (such as barium titanate powder in the inner paste), and combine them with different sintering temperature profiles to find a combination of formulations and processes that meets the following stringent conditions: 1. Co-firing compatibility: The shrinkage rate of the internal electrode and the ceramic medium is consistent during the sintering process (approximately 1200-1300℃), avoiding delamination and cracking.
[0004] 2. Microstructure: The inner electrode is dense, continuous, and pore-free; the outer electrode is dense and forms a good metallurgical bond with the inner electrode.
[0005] 3. Electrical properties: Low equivalent series resistance (ESR) of internal electrodes and high conductivity and solderability of external electrodes.
[0006] 4. Process adaptability: The paste has suitable rheological properties to meet the requirements of screen printing or dip coating.
[0007] This traditional development model is characterized by long development cycles, high costs, and low efficiency. Furthermore, due to the highly nonlinear coupling relationship between slurry components (metal powder, inorganic additives, organic carriers) and process parameters (heating rate, sintering temperature, and holding time), it is difficult to obtain the true optimal solution, which severely restricts the R&D iteration speed and performance improvement of high-capacity MLCCs.
[0008] Therefore, those skilled in the art are dedicated to providing a machine learning-based method for the synergistic optimization of high-capacity MLCC electrode slurry composition and process that can effectively solve the above-mentioned technical problems. Summary of the Invention
[0009] According to a first aspect of the embodiments of this specification, a collaborative optimization method for high-capacity MLCC electrode slurry based on machine learning is provided, the method comprising the following steps: A multi-source heterogeneous database is constructed to collect and integrate historical data from the research and development and production process of multilayer ceramic capacitor electrode pastes. The multi-source heterogeneous database includes paste composition characteristic data, process parameter characteristic data, and performance label data. A machine learning prediction model is trained using the multi-source heterogeneous database to construct a mapping relationship from the slurry component feature data and the process parameter feature data to the performance label data, thereby obtaining the prediction model. By setting an optimization objective and using the prediction model as a fitness function, a multi-objective optimization algorithm is employed to perform global optimization in a high-dimensional search space composed of slurry composition and process parameters, thereby obtaining a Pareto optimal solution set; and The optimization schemes in the Pareto optimal solution set are experimentally verified, and the verification data is fed back to the multi-source heterogeneous database to perform incremental learning or retraining on the prediction model.
[0010] In one possible implementation, the slurry composition characteristic data includes internal electrode slurry composition characteristic data and external electrode slurry composition characteristic data; the internal electrode slurry composition characteristic data includes nickel powder particle size distribution parameters, nickel powder specific surface area, nickel powder morphology, barium titanate powder addition amount, barium titanate powder particle size, type and ratio of organic carrier, and type and content of additives; the external electrode slurry composition characteristic data includes copper powder particle size distribution parameters, copper powder morphology, glass powder composition and content, glass powder softening point, and rheological property parameters of organic carrier.
[0011] In one possible implementation, the process parameter characteristic data includes parameters for the debinding stage and parameters for the sintering stage; the debinding stage parameters include the debinding heating rate, debinding temperature, debinding holding time, and oxygen partial pressure of the debinding atmosphere; the sintering stage parameters include parameters for the co-firing temperature curve of the inner electrode and parameters for the sintering temperature curve of the outer electrode; the parameters for the co-firing temperature curve of the inner electrode include the co-firing heating rate, maximum co-firing temperature, co-firing holding time, co-firing cooling rate, and oxygen partial pressure of the co-firing atmosphere; the parameters for the sintering temperature curve of the outer electrode include the sintering heating rate, maximum sintering temperature, sintering holding time, and sintering cooling rate; the performance label data includes physical performance data, electrical performance data, and process adaptability data; the physical performance data includes the inner electrode thickness, inner electrode continuity and density, interface state between the electrode and dielectric layer, outer electrode density, end coverage, and bonding strength; the electrical performance data includes capacity, loss tangent, insulation resistance, equivalent series resistance, and withstand voltage; the process adaptability data includes slurry viscosity, thixotropy, and pattern accuracy.
[0012] In one possible implementation, the machine learning prediction model includes an ensemble learning model and a deep learning model; the ensemble learning model is used to process numerical and categorical features and output prediction results for the performance label data; the deep learning model includes a multilayer perceptron model and a convolutional neural network model.
[0013] In one possible implementation, the convolutional neural network model is used to process image-type data. It takes a scanning electron microscope (SEM) image of the sintered electrode microstructure as input, extracts image features from the SEM image, and directly predicts the equivalent series resistance or breakdown probability of the multilayer ceramic capacitor based on these image features. The ensemble learning model includes an XGBoost model, which processes numerical and categorical features. It takes the slurry composition feature data and the process parameter feature data as input and outputs predictions of capacitance and insulation resistance. The objective function of the XGBoost model is: in, The objective function value of the XGBoost model; It is a differentiable loss function; For the first The true label value of each sample; For the front The cumulative predicted value of each tree; For the first A regression tree; For the first Feature vectors of each sample; For regularization terms; This represents the number of training samples; This represents the current iteration round.
[0014] In one possible implementation, the machine learning prediction model further includes a physically based neural network model, which incorporates the sintering kinetics equations and thermal expansion coefficient matching principles as constraints into the model training process; the total loss function of the physically based neural network model is: in, The total loss function of the neural network model based on physical information; For data-driven prediction loss; The sintering kinetics constraint weighting coefficient; This is due to the sintering kinetic constraint loss; The coefficient of thermal expansion is used to match the constraint weighting coefficient; The loss is due to the constraint of matching the coefficient of thermal expansion; The sintering kinetic constraint loss The residuals were calculated using the following sintering kinetic equations: in, Relative density; Sintering time; Pre-exponential factors; Activation energy; It is the gas constant; Absolute temperature; The reaction order; The thermal expansion coefficient matching constraint loss Calculated using the following formula: in, The coefficient of thermal expansion of the electrode; The coefficient of thermal expansion of the dielectric layer; This is the threshold for the difference in thermal expansion coefficients.
[0015] In one possible implementation, before training the machine learning prediction model, a generative adversarial network (GAN) is used to generate virtual microstructure images that conform to the process logic to augment the training dataset; the GAN includes a generator and a discriminator, and the loss function of the generator is: The loss function of the discriminator is: in, Let the loss function of the generator be denoted as . The loss function of the discriminator; For the generator; The discriminator; These are actual scanning electron microscope images of the electrode microstructure. It is a random noise vector; For expectation operators.
[0016] In one possible implementation, the optimization objectives include maximizing the capacitance and insulation resistance of the multilayer ceramic capacitor, and minimizing the equivalent series resistance and internal electrode thickness, while satisfying co-firing matching constraints and process window robustness constraints; the multi-objective optimization algorithm is the Non-Dominated Sort Genetic Algorithm NSGA-II, which performs hierarchical screening of individuals in the population through non-dominated sorting and crowding distance; the formula for calculating the crowding distance is: in, The first in the population The crowding distance of each individual; For the first One optimization objective function; and The first The individual in the first Two adjacent individuals on the optimization objective; and The first The maximum and minimum values of the optimization objective function in the current population; To optimize the number of targets.
[0017] In one possible implementation, the multi-objective optimization algorithm further introduces a prediction uncertainty weighting mechanism during the optimization process. The prediction variance of the prediction model for each optimization objective is used as a measure of uncertainty, and optimization objectives with higher prediction uncertainty are assigned lower weight coefficients. The comprehensive objective function after uncertainty weighting is: in, The value of the robust synthesis objective function; For the first The weighting coefficients of each optimization objective; For the first One optimization objective function value; To optimize the number of objectives; the weighting coefficients Calculated using the following formula: in, For the prediction model for the first The prediction variance of each optimization objective.
[0018] In one possible implementation, when incrementally learning or retraining the prediction model, the distribution distance between the validation data and the existing data in the multi-source heterogeneous database is calculated, and an update strategy for the prediction model is determined based on the distribution distance; the distribution distance is calculated using the maximum mean difference. in, The distribution of existing data in the multi-source heterogeneous database; To verify the distribution of the data; For kernel functions; and To be from the distribution Samples sampled from the middle; and To be from the distribution Samples sampled from the middle; For the expectation operator; when the When the value is less than a preset threshold, incremental learning is performed on the prediction model; when the value is less than a preset threshold, incremental learning is performed on the prediction model. If the value is greater than or equal to the preset threshold, the prediction model is retrained.
[0019] As described above, the present invention provides a collaborative optimization method for high-capacity MLCC electrode slurry based on machine learning, which has the following beneficial effects: 1. This invention constructs a multi-source heterogeneous database to structurally integrate the component characteristic data, process parameter characteristic data, and performance label data of internal electrode slurry and external electrode slurry. It uses machine learning prediction models to establish a mapping relationship from slurry composition and process parameters to final performance. It replaces a large number of physical experiments with AI proxy models, transforming the exploration of slurry formulation and process parameters from experience-based trial and error to intelligent optimization. This can shorten the R&D cycle by more than 50% and significantly reduce the costs of raw materials, equipment, and manpower. 2. This invention uses a prediction model as a fitness function and a multi-objective optimization algorithm to perform global optimization in a high-dimensional search space composed of slurry composition and process parameters. It can systematically handle the mutually restrictive technical indicators such as co-firing compatibility, electrical performance and process adaptability, and obtain the Pareto optimal solution set, that is, the combination of slurry formulation and process parameters that achieves the optimal balance among different performance indicators, rather than a local better solution, which helps to achieve the breakthrough of submicron internal electrode and higher layer stacking. 3. This invention constructs a neural network model based on physical information, and incorporates the sintering kinetic equation and the thermal expansion coefficient matching principle as constraints into the model training. This enables the model to accurately optimize the relationship between the amount of barium titanate powder added in the slurry and the sintering temperature curve, so as to achieve synchronous shrinkage of the electrode and the dielectric layer during the sintering process, fundamentally avoiding delamination and cracking. At the same time, it ensures that the model prediction results conform to physical laws and avoids invalid outputs that violate the sintering kinetics and thermal expansion matching principles. 4. This invention processes the SEM images of the electrode microstructure after sintering using a convolutional neural network model, automatically extracts information on micro-defects such as holes, necking, and discontinuities in the images, and directly predicts the equivalent series resistance or breakdown probability. This reflects the influence of microstructure on electrical performance more comprehensively than traditional numerical characterization. It can accurately design the sintering curve of the external electrode copper paste and obtain end electrodes with high densification and excellent solderability, thereby improving the capacity, reliability, and yield of MLCCs. 5. This invention expands the training dataset by generating virtual microstructure images that conform to the process logic through generative adversarial networks, solves the problem of insufficient model training caused by the scarcity of defect samples in reality, reduces the dependence on the collection of a large number of actual defect samples, shortens the data preparation cycle for model training, and improves the ability of convolutional neural network models to identify and predict defect features. 6. This invention introduces a prediction uncertainty weighting mechanism into a multi-objective optimization algorithm, using the prediction variance of the prediction model for each optimization objective as an uncertainty measure, and assigning lower weight coefficients to optimization objectives with higher prediction uncertainty. This makes the optimization algorithm pay more attention to performance indicators with higher prediction confidence, avoids deviation of optimization results from the actual optimum due to model prediction bias, and improves the robustness and reliability of the optimization scheme in practical applications. 7. This invention adaptively selects incremental learning or retraining update strategies by calculating the maximum mean difference between validation data and existing data. When the distribution difference between validation data and existing data is small, only incremental learning is performed to save computing resources. When the distribution difference is large, retraining is performed to ensure model accuracy, forming a closed loop of data, model, optimization, validation, and data. This enables continuous iteration of the model and continuous improvement of accuracy. At the same time, it integrates expert experience, experimental data, and optimization results into a continuously learning and evolving intelligent system, avoiding technical gaps caused by personnel turnover and realizing the digital management and efficient reuse of core process knowledge. 8. This invention uses machine learning prediction models to mine complex correlation patterns from high-dimensional data that are difficult for human experts to perceive. For example, when nickel powder with a specific particle size distribution works synergistically with additives of a specific concentration, an ultra-dense electrode can be obtained at a lower sintering temperature, providing a direction for developing new slurry formulations that break through the bottlenecks of existing technologies. Attached Figure Description
[0020] Figure 1 This is an overall flowchart of the high-capacity MLCC electrode slurry collaborative optimization method based on machine learning, as proposed in this invention.
[0021] Figure 2 This is a diagram of the multi-source heterogeneous database structure in this invention.
[0022] Figure 3 This is a diagram of the machine learning prediction model architecture in this invention.
[0023] Figure 4 This is a schematic flowchart of the multi-objective optimization algorithm in this invention.
[0024] Figure 5 This is a schematic flowchart of the model update strategy in this invention.
[0025] Figure 6 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation
[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments: In the description of this invention, it should be noted that the terms "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0027] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "setting," and "connection" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0028] This specification provides a collaborative optimization method for high-capacitance MLCC electrode paste based on machine learning. This method belongs to the field of electronic component material preparation and intelligent manufacturing technology. Specifically, it involves a method that uses artificial intelligence, especially machine learning algorithms, to collaboratively model and optimize the composition, physical properties, and associated sintering process of the inner and outer electrode pastes of multilayer ceramic capacitors (MLCCs). The following embodiments will be described in detail one by one.
[0029] This invention aims to solve the problems of long R&D cycles, high costs, and difficulty in obtaining the global optimal solution caused by relying on trial and error methods for slurry development and process matching in the prior art. By constructing a high-precision prediction model, it realizes the coordinated, efficient, and intelligent design of slurry formulation and sintering process.
[0030] See Figures 1 to 6 , Figure 1 A flowchart of a machine learning-based collaborative optimization method for high-capacity MLCC electrode slurry is shown according to an embodiment of the present invention, which specifically includes the following steps.
[0031] S110, Construct a multi-source heterogeneous database to collect and integrate historical data from the research and development and production process of multilayer ceramic capacitor electrode paste. The multi-source heterogeneous database includes paste composition characteristic data, process parameter characteristic data, and performance label data.
[0032] As an example, in this embodiment or other embodiments, the slurry composition characteristic data includes internal electrode slurry composition characteristic data and external electrode slurry composition characteristic data. The internal electrode slurry composition characteristic data includes nickel powder particle size distribution parameters (D10, D50, D90), nickel powder specific surface area, nickel powder morphology (spherical or flake-like), barium titanate powder (hereinafter referred to as BT powder) addition amount (wt%), barium titanate powder particle size, type and ratio of organic carrier (including resin, solvent, dispersant), and type and content of additives. The external electrode slurry composition characteristic data includes copper powder particle size distribution parameters, copper powder morphology, glass powder composition and content, glass powder softening point, and rheological property parameters of organic carrier (including viscosity and thixotropic index), which are not limited here.
[0033] As an example, in this embodiment or other embodiments, the process parameter characteristic data includes debinding stage parameters and sintering stage parameters. The debinding stage parameters include debinding heating rate, debinding temperature, debinding holding time, and debinding atmosphere oxygen partial pressure. The sintering stage parameters include internal electrode co-firing temperature curve parameters and external electrode sintering temperature curve parameters. The internal electrode co-firing temperature curve parameters include co-firing heating rate, maximum co-firing temperature, co-firing holding time, co-firing cooling rate, and co-firing atmosphere oxygen partial pressure. The external electrode sintering temperature curve parameters include sintering heating rate, maximum sintering temperature, sintering holding time, and sintering cooling rate. These are not limited here.
[0034] As an example, in this embodiment or other embodiments, the performance label data includes physical performance data, electrical performance data, and process adaptability data. The physical performance data includes inner electrode thickness, inner electrode continuity and density (quantified by SEM image analysis), interface state between electrode and dielectric layer (presence or absence of delamination and voids), outer electrode density, end coverage, and bonding strength. The electrical performance data includes capacity, loss tangent (DF), insulation resistance (IR), equivalent series resistance (ESR), and withstand voltage (BDV). The process adaptability data includes slurry viscosity, thixotropy, and pattern accuracy (pattern accuracy of screen printing or dip coating), which are not limited here.
[0035] See Figure 2 , Figure 2 The structure of the multi-source heterogeneous database in this invention is shown. This database integrates slurry composition characteristics, process parameter characteristics, and performance label data in a structured manner, providing a data foundation for the training of subsequent machine learning models.
[0036] Table 1 shows an example of the structure of slurry composition feature data, process parameter feature data, and performance label data in the multi-source heterogeneous database. Data categories Feature Dimension Example of specific parameters Internal electrode slurry composition characteristics Nickel powder particle size distribution (D10, D50, D90), specific surface area, morphology, BT powder addition (wt%), BT powder particle size, organic carrier ratio, type and content of additives D50 = 0.15μm, BT addition = 8wt%, resin content = 5wt% External electrode paste composition characteristics Copper powder particle size distribution, morphology, glass powder composition and content, softening point, and rheological properties Glass powder softening point = 650℃, viscosity = 15 Pa·s Parameters of the glue removal stage Heating rate, glue discharge temperature, holding time, atmospheric oxygen partial pressure Heating rate = 2℃ / min, glue discharge temperature = 350℃ Internal electrode co-firing parameters Heating rate, maximum sintering temperature, holding time, cooling rate, and atmospheric oxygen partial pressure Maximum sintering temperature = 1280℃, holding time = 2h External electrode sintering parameters Heating rate, maximum sintering temperature, holding time, cooling rate Maximum sintering temperature = 780℃, holding time = 30min Physical property label Inner electrode thickness, continuity and density, interface bonding state, outer electrode density, end coverage, and bonding strength. Electrode thickness = 0.83 μm, continuity = 99% Electrical performance label Capacity, DF, IR, ESR, BDV ESR ≤ target value, IR ≥ target value Process Adaptability Label Slurry viscosity, thixotropy, and pattern accuracy Thixotropic index = 4.5 Table 1 S120, using the multi-source heterogeneous database to train a machine learning prediction model, constructing a mapping relationship from the slurry component feature data and the process parameter feature data to the performance label data, and obtaining the prediction model.
[0037] As an example, in this embodiment or other embodiments, the machine learning prediction model includes an ensemble learning model and a deep learning model. The ensemble learning model is used to process numerical and categorical features and output prediction results for the performance label data. The deep learning model includes a multilayer perceptron (MLP) model and a convolutional neural network (CNN) model, which are not limited here.
[0038] As an example, in this embodiment or other embodiments, the convolutional neural network model is used to process image-type data. It takes the scanning electron microscope (SEM) image of the sintered electrode microstructure as input, extracts image features (including microscopic defect information such as holes, necking, and discontinuities) from the SEM image, and directly predicts the equivalent series resistance or breakdown probability of the multilayer ceramic capacitor based on the image features. This utilizes the microscopic defect information such as holes, necking, and discontinuities in the image, which is more comprehensive than traditional numerical representation, and is not limited here.
[0039] As an example, in this embodiment or other embodiments, the ensemble learning model includes the XGBoost model, which is used to process numerical and categorical features, taking the slurry composition feature data and the process parameter feature data as inputs, and outputting predictions of capacity and insulation resistance, which are not limited here.
[0040] The objective function of the XGBoost model is: ;in, The objective function value of the XGBoost model is... The loss function is differentiable. For the first The true label value of each sample For the front The cumulative predicted value of the trees, For the first A tree of return, For the first The feature vector of each sample For regularization terms, The number of training samples. This represents the current iteration round.
[0041] According to the above description, in an exemplary embodiment, the machine learning prediction model further includes a physical information-based neural network (PINN) model, which incorporates the sintering kinetic equation and the thermal expansion coefficient matching principle as constraints into the model training process to improve the model's predictive extrapolation ability and physical rationality in sparse data regions.
[0042] The total loss function of the neural network model based on physical information is: ;in, Let be the total loss function of the neural network model based on physical information. For data-driven prediction loss, The sintering kinetics constraint weighting coefficient is... For sintering kinetic constraint loss, To match the constraint weighting coefficient for the coefficient of thermal expansion, The loss is due to the constraint of matching the coefficient of thermal expansion.
[0043] The sintering kinetic constraint loss The residuals were calculated using the following sintering kinetic equations: ;in, For relative density, For sintering time, Pre-exponential factor, For activation energy, The gas constant is... Absolute temperature This represents the reaction order.
[0044] The thermal expansion coefficient matching constraint loss Calculated using the following formula: ;in, is the coefficient of thermal expansion of the electrode. The coefficient of thermal expansion of the dielectric layer, This is the threshold for the difference in thermal expansion coefficients.
[0045] Therefore, by incorporating the sintering kinetic equation and the thermal expansion coefficient matching principle as constraints into the model training, it is possible not only to improve the model's predictive extrapolation ability in sparse data regions, but also to ensure the physical rationality of the model's prediction results and avoid prediction outputs that violate physical laws. At the same time, through the thermal expansion coefficient matching constraint, it is possible to effectively ensure the consistency of the shrinkage rate between the inner electrode and the ceramic medium during the sintering process, fundamentally avoiding the problem of delamination and cracking.
[0046] According to the above description, in an exemplary embodiment, before training the machine learning prediction model, a virtual microstructure image conforming to the process logic is generated using a generative adversarial network (GAN) to expand the training dataset and solve the problem of insufficient model training caused by the scarcity of defect samples in reality.
[0047] The generative adversarial network includes a generator and a discriminator, and the loss function of the generator is: The loss function of the discriminator is: ;in, Let be the loss function of the generator. Let be the loss function of the discriminator. For the generator, For the discriminator, These are actual scanning electron microscope images of the electrode microstructure. It is a random noise vector. For expectation operators.
[0048] Therefore, it can be seen that by generating virtual microstructure images that conform to the process logic through generative adversarial networks, the training dataset can be effectively expanded, the problem of insufficient model training caused by the scarcity of defect samples can be solved, thereby improving the ability of convolutional neural network models to identify and predict defect features, and also reducing the dependence on the collection of a large number of actual defect samples, significantly shortening the data preparation cycle for model training.
[0049] See Figure 3 , Figure 3 The architecture of the machine learning prediction model in this invention is shown. This architecture integrates the XGBoost model, the multilayer perceptron model, the convolutional neural network model, and optional physical information-based neural network models and generative adversarial networks to form a multi-model collaborative prediction system.
[0050] Through model training, an AI proxy model is obtained that can accurately predict the key performance of the final MLCC under given slurry composition and process parameters.
[0051] S130, Set the optimization objective, use the prediction model as the fitness function, and use a multi-objective optimization algorithm to perform global optimization in a high-dimensional search space composed of slurry composition and process parameters to obtain the Pareto optimal solution set.
[0052] As an example, in this embodiment or other embodiments, the optimization objectives include maximizing the capacitance and insulation resistance of the multilayer ceramic capacitor, and minimizing the equivalent series resistance and internal electrode thickness (to achieve miniaturization), while satisfying co-fired matching constraints (no delamination, cracking) and process window robustness constraints, which are not limited here.
[0053] As an example, in this embodiment or other embodiments, the multi-objective optimization algorithm is the non-dominated sorting genetic algorithm NSGA-II, which performs hierarchical screening of individuals in the population through non-dominated sorting and crowding distance, and is not limited here.
[0054] The formula for calculating the congestion distance is: ;in, The first in the population The crowding distance of each individual For the first An optimization objective function, and The first The individual in the first Two adjacent individuals on the optimization objective, and The first The maximum and minimum values of the optimization objective function in the current population. To optimize the number of targets.
[0055] The optimization algorithm automatically recommends a set of Pareto optimal solutions through iteration, which is a series of slurry formulations and process parameter combinations that achieve the best balance among different performance indicators.
[0056] According to the above description, in an exemplary embodiment, the multi-objective optimization algorithm also introduces a prediction uncertainty weighting mechanism during the optimization process, using the prediction variance of the prediction model for each optimization objective as an uncertainty measure, and assigning a lower weight coefficient to optimization objectives with higher prediction uncertainty.
[0057] The comprehensive objective function after uncertainty weighting is: ;in, The value of the robust synthesis objective function. For the first The weighting coefficients of each optimization objective. For the first The value of the optimization objective function. To optimize the number of targets.
[0058] The weighting coefficient Calculated using the following formula: ;in, For the prediction model for the first The prediction variance of each optimization objective.
[0059] Therefore, by introducing a prediction uncertainty weighting mechanism, we can reduce the impact of optimization objectives with high prediction uncertainty on the optimization results, improve the robustness and reliability of optimization schemes in practical applications, and enable optimization algorithms to pay more attention to performance indicators with high prediction confidence, thus avoiding deviation of optimization results from the actual optimum due to model prediction bias.
[0060] See Figure 4 , Figure 4 The flowchart of the multi-objective optimization algorithm in this invention is shown, including steps such as population initialization, non-dominated sorting, crowding calculation, selection, crossover, mutation, and uncertainty weighting.
[0061] S140, Experimentally verify the optimization scheme in the Pareto optimal solution set, and feed the verification data back to the multi-source heterogeneous database to perform incremental learning or retraining on the prediction model, forming a closed loop of "data-model-optimization-verification-data".
[0062] According to the above description, in an exemplary embodiment, when incrementally learning or retraining the prediction model, the distribution distance between the verification data and the existing data in the multi-source heterogeneous database is calculated, and an update strategy for the prediction model is determined based on the distribution distance.
[0063] The distribution distance is calculated using the maximum mean difference (MMD): in, The distribution of existing data in the multi-source heterogeneous database. To verify the distribution of the data, For kernel function, and To be from the distribution Samples sampled from the middle, and To be from the distribution Samples sampled from the middle, For expectation operators.
[0064] When the MMD² is less than a preset threshold, the prediction model is incrementally learned; when the MMD² is greater than or equal to the preset threshold, the prediction model is retrained.
[0065] Therefore, by calculating the distribution distance based on the maximum mean difference, we can adaptively select the update strategy of incremental learning or retraining, which can improve the efficiency of model update while ensuring model accuracy. Moreover, when the distribution difference between the validation data and the existing data is small, we can only perform incremental learning to save computing resources, and when the distribution difference is large, we can retrain to ensure model accuracy.
[0066] See Figure 5 , Figure 5 The process of the model update strategy in this invention is shown, including steps such as calculating the distribution distance, determining the threshold, selecting the update strategy, and model iteration.
[0067] In summary, this invention provides a machine learning-based collaborative optimization method for high-capacity MLCC electrode slurries. First, a multi-source heterogeneous database is constructed to collect and integrate slurry composition feature data, process parameter feature data, and performance label data. Then, a machine learning prediction model is trained using this database to establish a mapping relationship from slurry composition feature data and process parameter feature data to performance label data, thus obtaining the prediction model. Next, an optimization objective is set, and the prediction model is used as a fitness function. A multi-objective optimization algorithm is then used to globally optimize within a high-dimensional search space composed of slurry composition and process parameters to obtain a Pareto optimal solution set. Simultaneously, the optimization schemes in the Pareto optimal solution set are experimentally verified, and the verification data is fed back to the multi-source heterogeneous database for incremental learning or retraining of the prediction model. Therefore, this method, by constructing a closed loop of data, model, optimization, verification, and data, can achieve collaborative, efficient, and intelligent design of slurry formulation and sintering process, significantly shortening the R&D cycle and reducing R&D costs. Furthermore, this method, through a multi-objective optimization algorithm, can systematically handle the interdependent technical indicators such as co-firing compatibility, electrical performance, and process adaptability, obtaining a globally optimal slurry-process solution, thus facilitating breakthroughs in submicron internal electrode fabrication and higher-layer stacking. Simultaneously, this method, through a physically based neural network model, can improve the model's predictive extrapolation ability and physical plausibility in sparse data regions. Moreover, by expanding the training dataset using generative adversarial networks, this method can address the problem of insufficient model training caused by scarce defect samples. Additionally, by introducing a prediction uncertainty weighting mechanism, this method can improve the robustness and reliability of the optimization scheme in practical applications, and through an adaptive update strategy based on the maximum mean difference, it can achieve continuous model iteration and continuous improvement in accuracy.
[0068] Example 1
[0069] This embodiment takes the optimization of the internal electrode nickel paste of a certain type of high-capacity MLCC as an example to describe the method of the present invention in detail. The goal is to reduce the thickness of the internal electrode from 1.0 μm to 0.8 μm while maintaining the capacity and ESR.
[0070] In step S110, 500 sets of data are extracted from the historical database. Each set includes nickel powder D50, BT powder addition amount, resin content, maximum sintering temperature, holding time, and corresponding electrode thickness, continuity and MLCC capacity and ESR after sintering.
[0071] In step S120, the XGBoost algorithm is used to train a prediction model with slurry composition and process parameters as input. This model can accurately predict electrode thickness (RMSE < 0.05 μm), capacity (error < 3%), and ESR. At the same time, before training, a virtual microstructure image that conforms to the process logic is generated using a generative adversarial network to expand the training dataset, and a neural network model based on physical information is constructed. The sintering kinetic equation and the thermal expansion coefficient matching principle are incorporated as constraints into the model training process.
[0072] In step S130, the optimization objectives are set as follows: ① Electrode thickness ≤ 0.85 μm; ② Electrode continuity ≥ 98%; ③ Capacity ≥ target value; ④ ESR ≤ target value. The NSGA-II algorithm is used to perform 100 iterations of optimization in the input parameter space. At the same time, a prediction uncertainty weighting mechanism is introduced in the optimization process to assign lower weight coefficients to optimization objectives with higher prediction uncertainty.
[0073] Optimization results: The algorithm recommends three Pareto optimal solutions. One of the solutions uses spherical nickel powder with a D50 of 0.15 μm and 8 wt% BT powder. A stepped sintering curve is adopted (first rapidly heated to 1100℃, then slowly heated to 1280℃ and held for 2 hours). The model predicts that the electrode thickness under this solution is 0.82 μm, with a continuity of 99%, and the capacity and ESR both meet the requirements.
[0074] In step S140, the slurry was prepared according to the recommended scheme and a co-firing experiment was conducted. The actual measured electrode thickness was 0.83 μm. SEM showed that the electrode was continuous and dense without delamination. The MLCC electrical performance test results were highly consistent with the model prediction. The verification data was fed back to the database. The calculated MMD² between the verification data and the existing data was less than the preset threshold. Incremental learning was performed on the prediction model for the next round of model iteration.
[0075] The performance verification data for Example 1 are shown in Table 2: Testing items Target value Model predictions Measured value Electrode thickness (μm) ≤0.85 0.82 0.83 Electrode continuity (%) ≥98 99 99 capacity ≥ target value satisfy satisfy ESR ≤ target value satisfy satisfy Delamination / Cracking none none none Table 2 The XGBoost model prediction accuracy verification data in Example 1 are shown in Table 3: Predicting targets RMSE relative error Electrode thickness 0.042μm 5.1% capacity — <3% ESR — <5% Table 3 The comparison of the three schemes in the Pareto optimal solution set in Example 1 is shown in Table 4:
[0076] Table 4 Example 1
[0077] This embodiment takes the densification of the external electrode copper paste and the optimization of solderability as an example. It further explains the method of the present invention to address the problem that there are holes after the external electrode of a large-size high-capacitance MLCC is burned in, which leads to the risk of short circuit caused by tin seepage during subsequent electroplating.
[0078] In step S110, 200 sets of copper paste data are collected, including copper powder particle size, glass powder softening point, sintering temperature curve (heating rate, maximum temperature, holding time), and corresponding copper layer density (porosity quantified by image analysis) and solderability test results.
[0079] In step S120, the SEM image and process parameters are used as input to train a hybrid model of CNN and MLP. The CNN branch extracts hole features in the image, the MLP branch processes the process parameters, and finally the predicted density and solderability scores are jointly output.
[0080] In step S130, with the goal of minimizing porosity and maximizing weldability score, the Bayesian optimization algorithm is used to efficiently explore the process parameter space.
[0081] Optimization results: The algorithm recommends an asymmetric sintering curve, which involves rapidly heating (40℃ / min) to 650℃, then slowly heating (5℃ / min) to 780℃ and holding at that temperature for 30 minutes, followed by rapid cooling. The model predicts that this process can reduce porosity to below 1% and exhibit excellent weldability.
[0082] In step S140, the external electrode is applied according to this process. SEM shows that the copper layer is dense with a porosity of only 0.8%. Subsequent electroplating and welding tests did not find any tin penetration or cold solder joints, which verified the effectiveness of the optimization scheme. The verification data was fed back to the database. After calculation, the MMD² between the verification data and the existing data was greater than the preset threshold, and the prediction model was retrained.
[0083] The performance verification data for Example 2 are shown in Table 5: Testing items Before optimization Model predictions Measured value Porosity (%) >3 <1 0.8 Solderability rating generally Excellent Excellent Tin seepage risk have none none Risk of cold solder joints have none none Table 5 The key parameters of the asymmetric sintering curve in Example 2 are shown in Table 6: stage Heating rate (°C / min) Target temperature (°C) Insulation time Rapid heating phase 40 650 — Slow heating phase 5 780 30min Rapid cooling section — — — Table 6 The comparative examples illustrate the superiority of the method of the present invention over the traditional trial-and-error method, comparing the traditional method with the method of the present invention in terms of R&D cycle and performance improvement, as shown in Table 7: Comparison Projects Traditional trial and error method Method of the present invention R&D cycle 6-12 months 2-3 months (shortened by more than 50%) Number of experiments 100-200 times 10-20 times (for verification purposes only) Raw material consumption high Low (reduced by more than 70%) Optimal solution quality Local Optimum Global Pareto Optimality Multi-objective collaborative capability Weak (difficult to balance all aspects simultaneously) Strong (systematic handling of mutually restrictive indicators) Knowledge reusability Low (depends on personal experience) High (digital accumulation and continuous learning) personnel dependence High (Technology gap caused by core personnel turnover) Low (digital management of process knowledge) Table 7 The method of this invention replaces a large number of physical experiments with AI models, transforming the exploration of slurry formulation and process parameters from "experience-based trial and error" to "intelligent optimization".
[0084] Compared to traditional trial-and-error methods, this invention significantly shortens the R&D cycle and reduces R&D costs, shortening the cycle by more than 50% and significantly reducing raw material, equipment, and labor costs. It enables multi-objective collaborative global optimization, systematically addressing mutually constraining technical indicators such as the co-firing compatibility of internal electrodes, the solderability of external electrodes, and electrical performance and process adaptability, resulting in a globally optimal slurry-process solution. It improves product performance and consistency by precisely optimizing the amount of BT powder added to the internal slurry, achieving ideal "synchronous shrinkage" of the electrode and dielectric layer during sintering, fundamentally avoiding delamination and cracking, and accurately designing the sintering curve of the external electrode copper paste to obtain end electrodes with high densification and excellent solderability. It can uncover tacit knowledge and new formulations, extracting complex correlation patterns from high-dimensional data that are difficult for human experts to perceive, providing direction for developing entirely new slurry formulations that break through existing technological bottlenecks. It enables the digital accumulation and reuse of process knowledge, integrating expert experience, experimental data, and optimization results into a continuously learning and evolving intelligent system, avoiding technological gaps caused by personnel turnover.
[0085] Figure 6 A structural block diagram of a computing device 400 according to one embodiment of this specification is shown. The components of the computing device 400 include, but are not limited to, a memory 410 and a processor 420. The processor 420 is connected to the memory 410 via a bus 430, and a database 450 is used to store data.
[0086] The computing device 400 also includes an access device 440, which enables the computing device 400 to communicate via one or more networks 460. Examples of these networks include a Public Switched Telephone Network (PSTN), a Local Area Network (LAN), a Wide Area Network (WAN), a Personal Area Network (PAN), or a combination of communication networks such as the Internet. The access device 440 may include one or more of any type of wired or wireless network interface (e.g., a Network Interface Card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) interface, a Wi-MAX interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.
[0087] In one embodiment of this specification, the aforementioned components of the computing device 400 and Figure 6 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 6 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.
[0088] The computing device 400 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 400 can also be a mobile or stationary server.
[0089] The processor 420 executes computer-executable instructions, which, when executed by the processor, implement the steps of the aforementioned machine learning-based high-capacity MLCC electrode slurry co-optimization method. The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the aforementioned machine learning-based high-capacity MLCC electrode slurry co-optimization method belong to the same concept. Details not described in detail in the technical solution of the computing device can be found in the description of the technical solution of the aforementioned method.
[0090] This specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the aforementioned machine learning-based high-capacity MLCC electrode slurry co-optimization method. The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium and the technical solution of the aforementioned machine learning-based high-capacity MLCC electrode slurry co-optimization method belong to the same concept. Details not described in detail in the technical solution of the storage medium can be found in the description of the technical solution of the aforementioned method.
[0091] This specification also provides a computer program in one embodiment, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the above-described machine learning-based high-capacity MLCC electrode slurry co-optimization method. The above is an illustrative scheme of a computer program according to this embodiment. It should be noted that the technical solution of this computer program and the technical solution of the above-described machine learning-based high-capacity MLCC electrode slurry co-optimization method belong to the same concept. Details not described in detail in the computer program's technical solution can be found in the description of the technical solution of the above method.
[0092] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.
[0093] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content contained in the computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.
[0094] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.
[0095] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0096] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.
Claims
1. A collaborative optimization method for high-capacity MLCC electrode paste based on machine learning, characterized in that, The method includes the following steps: A multi-source heterogeneous database is constructed to collect and integrate historical data from the research and development and production process of multilayer ceramic capacitor electrode pastes. The multi-source heterogeneous database includes paste composition characteristic data, process parameter characteristic data, and performance label data. A machine learning prediction model is trained using the multi-source heterogeneous database to construct a mapping relationship from the slurry component feature data and the process parameter feature data to the performance label data, thereby obtaining the prediction model. By setting an optimization objective and using the prediction model as a fitness function, a multi-objective optimization algorithm is employed to perform global optimization in a high-dimensional search space composed of slurry composition and process parameters, thereby obtaining a Pareto optimal solution set; and The optimization schemes in the Pareto optimal solution set are experimentally verified, and the verification data is fed back to the multi-source heterogeneous database to perform incremental learning or retraining on the prediction model.
2. The high-capacity MLCC electrode paste co-optimization method based on machine learning according to claim 1, characterized in that, The slurry composition characteristic data includes the internal electrode slurry composition characteristic data and the external electrode slurry composition characteristic data; the internal electrode slurry composition characteristic data includes nickel powder particle size distribution parameters, nickel powder specific surface area, nickel powder morphology, barium titanate powder addition amount, barium titanate powder particle size, type and ratio of organic carrier, and type and content of additives; the external electrode slurry composition characteristic data includes copper powder particle size distribution parameters, copper powder morphology, glass powder composition and content, glass powder softening point, and rheological property parameters of organic carrier.
3. The high-capacity MLCC electrode paste co-optimization method based on machine learning according to claim 1, characterized in that, The process parameter characteristic data includes parameters for the debinding stage and parameters for the sintering stage. The debinding stage parameters include the debinding heating rate, debinding temperature, debinding holding time, and oxygen partial pressure in the debinding atmosphere. The sintering stage parameters include parameters for the co-firing temperature curve of the inner electrode and the sintering temperature curve of the outer electrode. The co-firing temperature curve parameters of the inner electrode include the co-firing heating rate, maximum co-firing temperature, co-firing holding time, co-firing cooling rate, and oxygen partial pressure in the co-firing atmosphere. The sintering temperature curve parameters of the outer electrode include the sintering heating rate, maximum sintering temperature, sintering holding time, and sintering cooling rate. The performance label data includes physical performance data, electrical performance data, and process adaptability data. The physical performance data includes the inner electrode thickness, inner electrode continuity and density, interface state between the electrode and dielectric layer, outer electrode density, end coverage, and bonding strength. The electrical performance data includes capacity, loss tangent, insulation resistance, equivalent series resistance, and withstand voltage. The process adaptability data includes slurry viscosity, thixotropy, and pattern accuracy.
4. The high-capacity MLCC electrode paste co-optimization method based on machine learning according to claim 1, characterized in that, The machine learning prediction model includes an ensemble learning model and a deep learning model; the ensemble learning model is used to process numerical and categorical features and output prediction results for the performance label data; the deep learning model includes a multilayer perceptron model and a convolutional neural network model.
5. The high-capacity MLCC electrode paste co-optimization method based on machine learning according to claim 4, characterized in that, The convolutional neural network model is used to process image-type data. It takes scanning electron microscope (SEM) images of the sintered electrode microstructure as input, extracts image features from the SEM images, and directly predicts the equivalent series resistance or breakdown probability of the multilayer ceramic capacitor based on these image features. The ensemble learning model includes an XGBoost model, which processes numerical and categorical features. It takes the slurry composition feature data and the process parameter feature data as input and outputs predictions for capacitance and insulation resistance. The objective function of the XGBoost model is: in, The objective function value of the XGBoost model; It is a differentiable loss function; For the first The true label value of each sample; For the front The cumulative predicted value of each tree; For the first A regression tree; For the first Feature vectors of each sample; For regularization terms; This represents the number of training samples; This represents the current iteration round.
6. The high-capacity MLCC electrode paste co-optimization method based on machine learning according to claim 4, characterized in that, The machine learning prediction model also includes a physical information-based neural network model, which incorporates the sintering kinetics equation and the thermal expansion coefficient matching principle as constraints into the model training process; the total loss function of the physical information-based neural network model is: in, The total loss function of the neural network model based on physical information; For data-driven prediction loss; The sintering kinetics constraint weighting coefficient; This is due to the sintering kinetic constraint loss; The coefficient of thermal expansion is used to match the constraint weighting coefficient; The loss is due to the constraint of matching the coefficient of thermal expansion; The sintering kinetic constraint loss The residuals were calculated using the following sintering kinetic equations: in, Relative density; Sintering time; Pre-exponential factors; Activation energy; It is the gas constant; Absolute temperature; The reaction order; The thermal expansion coefficient matching constraint loss Calculated using the following formula: in, The coefficient of thermal expansion of the electrode; The coefficient of thermal expansion of the dielectric layer; This is the threshold for the difference in thermal expansion coefficients.
7. The high-capacity MLCC electrode paste co-optimization method based on machine learning according to claim 4, characterized in that, Before training the machine learning prediction model, a generative adversarial network (GAN) is used to generate virtual microstructure images that conform to the process logic to expand the training dataset; the GAN includes a generator and a discriminator, and the loss function of the generator is: The loss function of the discriminator is: in, The loss function of the generator; The loss function of the discriminator; For the generator; The discriminator; These are actual scanning electron microscope images of the electrode microstructure. It is a random noise vector; For expectation operators.
8. The high-capacity MLCC electrode paste co-optimization method based on machine learning according to claim 1, characterized in that, The optimization objectives include maximizing the capacitance and insulation resistance of the multilayer ceramic capacitor, and minimizing the equivalent series resistance and internal electrode thickness, while satisfying co-firing matching constraints and process window robustness constraints. The multi-objective optimization algorithm is the Non-Dominated Sort Genetic Algorithm NSGA-II, which uses non-dominated sorting and crowding distance to perform hierarchical screening of individuals in the population. The formula for calculating the crowding distance is: in, The first in the population The crowding distance of each individual; For the first One optimization objective function; and The first The individual in the first Two adjacent individuals on the optimization objective; and The first The maximum and minimum values of the optimization objective function in the current population; To optimize the number of targets.
9. The high-capacity MLCC electrode paste co-optimization method based on machine learning according to claim 8, characterized in that, The multi-objective optimization algorithm also introduces a prediction uncertainty weighting mechanism during the optimization process, using the prediction variance of the prediction model for each optimization objective as an uncertainty measure, and assigning lower weight coefficients to optimization objectives with higher prediction uncertainty; the comprehensive objective function after uncertainty weighting is: in, The value of the robust synthesis objective function; For the first The weighting coefficients of each optimization objective; For the first One optimization objective function value; To optimize the number of objectives; the weighting coefficients Calculated using the following formula: in, For the prediction model for the first The prediction variance of each optimization objective.
10. The high-capacity MLCC electrode paste co-optimization method based on machine learning according to claim 1, characterized in that, When incrementally learning or retraining the prediction model, the distribution distance between the validation data and the existing data in the multi-source heterogeneous database is calculated, and an update strategy for the prediction model is determined based on the distribution distance; the distribution distance is calculated using the maximum mean difference. in, The distribution of existing data in the multi-source heterogeneous database; To verify the distribution of the data; For kernel functions; and To be from the distribution Samples sampled from the middle; and To be from the distribution Samples sampled from the middle; For the expectation operator; when the When the value is less than a preset threshold, incremental learning is performed on the prediction model; when the value is less than a preset threshold, incremental learning is performed on the prediction model. If the value is greater than or equal to the preset threshold, the prediction model is retrained.
