Preparation method and system of alumina-based ceramic chisel with anti-metal contamination

Abstract

This invention discloses a method and system for preparing an alumina-based ceramic chopping tool resistant to metal contamination, belonging to the field of computer-aided manufacturing technology. The method includes: constructing an optimization space for alumina-based ceramic doping formulations; generating an initial formulation population and initializing anisotropic search step size; constructing an anisotropic neighborhood based on the anisotropic search step size; constructing a multidimensional optimization objective function, and iteratively optimizing the formulation population by combining a doping proxy model and the anisotropic neighborhood to obtain a target doping formulation; and completing raw material preparation, molding, sintering, and processing according to the target doping formulation to obtain an alumina-based ceramic chopping tool resistant to metal contamination. This invention solves the technical problems of low optimization efficiency, long experimental cycle, high R&D cost, and difficulty in achieving multi-performance synergistic optimization in the prior art for alumina-based ceramic chopping tools.

Description

Technical Field

[0001] This invention relates to the field of computer-aided manufacturing technology, specifically to a method and system for preparing an alumina-based ceramic chopping tool resistant to metal contamination. Background Technology

[0002] Alumina-based ceramics are widely used in the manufacture of ceramic cleavers for semiconductor packaging due to their high hardness, high wear resistance, good insulation properties, and chemical stability. In practical use, ceramic cleavers are in prolonged contact with metal wires, and their surfaces easily absorb and adhere metal particles, forming metal contamination. This contamination affects bonding quality, machining accuracy, and service life. Therefore, improving the resistance of alumina-based ceramic cleavers to metal contamination is of great significance.

[0003] Currently, the properties of alumina-based ceramics are typically improved by adjusting the dopant components and their content. However, complex interactions exist among the multiple dopant components, and the design space expands significantly with the increase in the number of dopant variables. Existing formulation development processes mostly rely on empirical screening and repeated experimental verification, which not only involves long development cycles and high experimental costs but also makes it difficult to simultaneously achieve resistance to metal fouling with comprehensive properties such as hardness, strength, and density. Therefore, how to efficiently obtain alumina-based ceramic chopping knife formulations that combine excellent resistance to metal fouling with comprehensive mechanical properties has become a pressing technical problem to be solved in this field. Summary of the Invention

[0004] This invention provides a method and system for preparing alumina-based ceramic chopping blades that resist metal contamination, thereby solving the technical problem that the development of alumina-based ceramic chopping blade formulations in the prior art relies on experience screening and extensive experimental verification, resulting in long research and development cycles, high experimental costs, and difficulty in obtaining formulations that combine excellent resistance to metal contamination and comprehensive mechanical properties.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a method for preparing an alumina-based ceramic chopping knife resistant to metal contamination, comprising: Based on prior experience from the target production line, the optimization space for the doping formulation of alumina-based ceramics is defined, and the optimization space for the doping formulation includes the range of mass percentage values ​​for each doping component. An initial formulation population is generated in the optimization space, and a search step size is initialized for each doping variable dimension in the doping formulation optimization space to obtain the anisotropic search step size. Based on the anisotropic search step size, an anisotropic distance metric is constructed, and the anisotropic neighborhood corresponding to each initial formula in the initial formula population is determined according to the anisotropic distance metric. A multidimensional optimization objective function is constructed, and the initial formulation population is iteratively optimized by combining the anisotropic neighborhood with the pre-constructed doping proxy model. Normalization constraint processing is performed in each iteration until the preset convergence condition is met, and the target doping formulation is output. According to the target doping formula, an alumina-based ceramic chopping knife resistant to metal contamination is obtained through a preset process flow treatment.

[0006] Secondly, the present invention provides a system for preparing an alumina-based ceramic chopping knife resistant to metal contamination, comprising: An optimization space construction module is used to define the optimization space of doping formulation for alumina-based ceramics based on the prior experience of the target production line. The optimization space of doping formulation includes the range of mass percentage values ​​of each doping component. The initial population generation module is used to generate an initial formulation population in the doping formulation optimization space, and initialize the search step size for each doping variable dimension in the doping formulation optimization space to obtain the anisotropic search step size. The neighborhood construction module is used to construct an anisotropic distance metric based on the anisotropic search step size, and determine the anisotropic neighborhood corresponding to each initial recipe in the initial recipe population according to the anisotropic distance metric. The formulation optimization module is used to construct a multidimensional optimization objective function and combine the anisotropic neighborhood with the pre-constructed doping proxy model to iteratively optimize the initial formulation population. In each iteration, normalization constraint processing is performed until the preset convergence condition is met, and the target doping formulation is output. The process preparation module is used to prepare an alumina-based ceramic chopping knife resistant to metal contamination by processing it through a preset process flow according to the target doping formula.

[0007] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention constructs a doping formulation optimization system driven by big data and surrogate models, combining an anisotropic neighborhood search mechanism and a multi-objective collaborative optimization method to achieve efficient optimization of doping formulations for alumina-based ceramic chopping tools. Compared with existing technologies, this invention can significantly improve formulation design efficiency, reduce experimental costs, and achieve a synergistic improvement in anti-metal fouling performance and mechanical properties, thereby obtaining alumina-based ceramic chopping tools with excellent anti-metal fouling performance, high hardness, high strength, and high reliability. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the process for preparing an alumina-based ceramic chopping knife resistant to metal contamination, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the formula optimization iteration mechanism provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a metal-resistant alumina-based ceramic chopping knife preparation system module provided in an embodiment of the present invention; In the attached diagram, the components represented by each number are as follows: The system includes: optimization space construction module 11, initial population generation module 12, neighborhood construction module 13, formula optimization module 14, and process preparation module 15. Detailed Implementation

[0009] This invention provides a method and system for preparing alumina-based ceramic chopping tools that are resistant to metal contamination, and specifically addresses the technical problems in the prior art, such as low efficiency in optimizing doping formulations, long experimental cycles, high R&D costs, and difficulty in achieving synergistic optimization of multiple properties.

[0010] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0011] Example 1, as Figure 1 and Figure 2 As shown, the present invention provides a method for preparing an alumina-based ceramic chopping knife resistant to metal contamination, comprising: S100: Based on the prior experience of the target production line, define the optimization space of the doping formula for alumina-based ceramics. The optimization space of the doping formula includes the range of mass percentage values ​​of each doping component. Step S100 in the method provided in this application embodiment includes: High-purity α-alumina with a purity of not less than 99.5% and an average particle size of 0.3 to 0.8 micrometers is used as the matrix material; Obtain historical formula experimental data, determine the composition of doping components based on historical formula experimental data, and define the corresponding doping formula optimization space. The doping formula optimization space includes multiple dimensions, and each dimension corresponds to the mass percentage of a doping component. Based on historical formulation experimental data, combined with statistical analysis methods and preset value amplification factors, the mass percentage range of each dopant component in the doping formulation optimization space is set.

[0012] In this step, the doping formulation optimization space refers to a multi-dimensional parameter space composed of the mass percentage of multiple doping components, where each dimension corresponds to the range of mass percentage values ​​of a doping component, which is used by the subsequent optimization algorithm to search for the target formulation. The amplification factor refers to the expansion factor set based on the statistical range of historical experimental data. It is used to appropriately expand the search boundary, avoid limiting the optimization process to the historical experimental area, and increase the probability of discovering potential superior formulations.

[0013] In this embodiment, high-purity α-alumina with a purity of not less than 99.5% and an average particle size of 0.3 to 0.8 micrometers is used as the matrix material. High-purity alumina can reduce the influence of impurities on the sintering process and final properties, while smaller and more uniform particle size is beneficial to improving sintering activity and densification, thus providing a stable material basis for subsequent doping modification.

[0014] Subsequently, historical formulation experimental data were acquired, and the composition of the doping components was determined based on this data. The historical formulation experimental data was derived from at least one of the following sources: completed alumina-based ceramic R&D experimental data, production trial data, and publicly available materials database data.

[0015] Specifically, statistical analysis was conducted on the types, mass percentages, and corresponding performance test results of doping elements in historical experimental samples to screen out doping components that have a significant impact on the resistance to metal contamination and overall mechanical properties, and to establish the corresponding optimization space for doping formulations.

[0016] Each dopant component corresponds to a variable dimension in the optimization space, and the value of each variable dimension represents the mass percentage of the corresponding dopant component in the final ceramic.

[0017] For example, when historical experimental data analysis results indicate that several doping components have a significant impact on the resistance to metal contamination, the doping components can be mapped to corresponding optimization dimensions to construct a multi-dimensional doping formulation optimization space. For instance, when four target doping components are selected, a first optimization dimension, a second optimization dimension, a third optimization dimension, and a fourth optimization dimension can be established respectively, thus forming a four-dimensional doping formulation optimization space; when five or more target doping components are selected, the space can be further expanded to form a higher-dimensional doping formulation optimization space.

[0018] Furthermore, based on historical experimental data, the minimum mass percentage, maximum mass percentage, average mass percentage, and standard deviation of each doped component were statistically analyzed.

[0019] The minimum mass percentage is used as the lower boundary, the maximum mass percentage is used as the upper boundary, and the average mass percentage and standard deviation are used to characterize the historical distribution characteristics of the corresponding doped components.

[0020] For example, for a certain target doped component, if the minimum mass percentage is 0.20 wt%, the maximum mass percentage is 1.00 wt%, the average mass percentage is 0.62 wt%, and the standard deviation is 0.15 wt%, then 0.20 wt% and 1.00 wt% can be used as the basic statistical boundaries.

[0021] Subsequently, a preset amplification factor is introduced to expand the basic statistical boundary.

[0022] The expanded search lower boundary is determined by the formula: Search lower boundary = minimum mass percentage - value amplification factor × standard deviation; the expanded search upper boundary is determined by the formula: Search upper boundary = maximum mass percentage + value amplification factor × standard deviation.

[0023] For example, when the amplification factor is 0.5, for the above target doped component, the expanded search lower boundary is 0.20-0.5×0.15=0.125wt%, and the expanded search upper boundary is 1.00+0.5×0.15=1.075wt%.

[0024] Finally, the search boundaries corresponding to each doping component are collectively used to form the doping formulation optimization space, which serves as the parameter boundary for the subsequent optimization search process.

[0025] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0026] In summary, this step screens key doping components using historical formulation experimental data and constructs a corresponding doping formulation optimization space. Simultaneously, it combines historical data statistical characteristics and value amplification factors to determine the search boundaries for each dimension. This ensures the rationality of the search space while enhancing the ability to explore potential superior formulations, thus providing a reliable parameter boundary basis for subsequent initial formulation population generation and target doping formulation optimization.

[0027] S200: Generate an initial formulation population in the optimization space, and initialize the search step size for each doping variable dimension in the doping formulation optimization space to obtain the anisotropic search step size; Step S200 in the method provided in this application embodiment includes: Based on historical formulation experimental data, an initial formulation population containing multiple initial formulations is generated in the optimization space. The mass percentage of each doped component in each initial formulation is normalized to meet the preset normalization constraints. Based on historical formulation experimental data, the normalized sensitivity corresponding to each doping variable dimension is calculated. Based on the normalized sensitivity and the preset basic step size scaling factor, the initial search step size corresponding to each doping variable dimension is calculated to obtain the anisotropic search step size.

[0028] In this step, the initial formulation population refers to a set of multiple candidate formulations generated within the doping formulation optimization space. Each candidate formulation contains the mass percentage parameters corresponding to each doping component. Normalized sensitivity is a relative quantitative index that measures the impact of a change in a doping variable on the target performance. It is used to characterize the importance of different doping variables to the optimization target. Anisotropic search step size refers to the set of search parameters with different search step sizes for different doping variable dimensions, which is used to improve the targeting and efficiency of subsequent optimization search processes.

[0029] In this embodiment, firstly, an initial formulation population is generated based on the doping formulation optimization space constructed in step S100. Specifically, according to the formulation distribution characteristics in historical formulation experimental data, multiple initial formulations are generated within the doping formulation optimization space to form the initial search population for subsequent optimization processes.

[0030] For example, when the doping formulation optimization space contains four doping variable dimensions, each initial formulation consists of four sets of doping component mass percentages. Multiple initial formulations covering different regions of the optimization space can be generated through random sampling, hierarchical sampling, or Latin hypercube sampling, thus forming an initial formulation population.

[0031] After generating the initial formulation, the mass percentage of each dopant component in each initial formulation is normalized to ensure that the dopant components maintain a reasonable proportional relationship and meet the composition constraints for subsequent formulation optimization.

[0032] Specifically, suppose an initial formula contains n doped components, and the mass percentage of each doped component is the first mass percentage, the second mass percentage, ..., the nth mass percentage, then first calculate the sum of the mass percentages of each doped component.

[0033] The percentage of total mass is calculated using the following formula: Total mass percentage = First mass percentage + Second mass percentage + ... + nth mass percentage.

[0034] Subsequently, the mass percentage of each dopant component was normalized.

[0035] The normalized value corresponding to the i-th dopant component is calculated according to the following formula: The normalized value of the i-th dopant component = mass percentage of the i-th dopant component ÷ total mass percentage.

[0036] After normalization, the normalized values ​​of each dopant component satisfy the following: First normalized value + Second normalized value + ... + nth normalized value = 1.

[0037] Furthermore, constraints are applied to each of the normalized initial formulas.

[0038] Specifically, when the sum of the normalized values ​​of each dopant component equals 1, the current initial formula is determined to meet the preset normalization constraint; otherwise, the proportional normalization process is re-executed for each dopant component until the sum of all normalized values ​​is 1.

[0039] Subsequently, the normalized sensitivity for each doping variable dimension was calculated based on historical formulation experimental data. The normalized sensitivity is used to reflect the degree of influence of changes in the corresponding doping variable on the target performance.

[0040] Furthermore, based on the normalized sensitivity and the preset basic step size scaling factor, the initial search step size corresponding to each doping variable dimension is calculated.

[0041] Specifically, different search scales are assigned to different variable dimensions based on the magnitude of the normalization sensitivity corresponding to each dopant variable.

[0042] Among them, variable dimensions with higher normalization sensitivity are given a smaller search step size to improve the search accuracy of key parameter regions; variable dimensions with lower normalization sensitivity are given a larger search step size to improve search efficiency and search range.

[0043] For example, when the normalization sensitivity of the first doping variable dimension is higher than that of the second doping variable dimension, the initial search step size corresponding to the first doping variable dimension is smaller than that corresponding to the second doping variable dimension, so that the optimization algorithm prioritizes fine-grained search of variables with greater influence.

[0044] Finally, the initial search step size corresponding to each doping variable dimension is combined to form the anisotropic search step size.

[0045] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0046] Step S200 in the method provided in this application embodiment further includes: Based on historical formulation experimental data and combined with response surface fitting method, the response relationship between anti-metal fouling performance and each dimension of doping variables was established. Based on the response relationship, the absolute value of the gradient sensitivity of the performance to each dimension of the doping variable is calculated; Using the maximum absolute value of the gradient sensitivity across all doping variable dimensions as a benchmark, the absolute value of the gradient sensitivity for each doping variable dimension is normalized to obtain the normalized sensitivity.

[0047] In this step, the response surface fitting method refers to a data modeling method that uses historical sample data to establish a mapping relationship between input variables and target performance, which is used to describe the influence of changes in various doping variables on the performance of metal contamination. The response relationship refers to the functional relationship between the anti-contamination performance of metals and each dimension of doping variables, as constructed by the response surface model. The absolute value of gradient sensitivity refers to the absolute value of the rate of change of the target performance relative to the corresponding doping variable in the response relationship, and is used to characterize the strength of the influence of the doping variable on the target performance. Normalized sensitivity refers to the relative influence index obtained after standardizing the absolute values ​​of the gradient sensitivity corresponding to each doping variable. It is used to eliminate the influence of differences in the dimensions and numerical ranges of different variables on the sensitivity evaluation results.

[0048] In this embodiment, the mass percentage of each dopant component in the historical formulation experimental data is first used as the input variable, and the corresponding anti-metal fouling performance index is used as the output variable. A performance prediction model is then constructed using response surface methodology. The anti-metal fouling performance index can be one or more of the following: metal adhesion force, metal residue area ratio, surface free energy change rate, or comprehensive anti-fouling evaluation index.

[0049] For example, when the doping formulation optimization space includes a first doping variable dimension, a second doping variable dimension, a third doping variable dimension, and a fourth doping variable dimension, historical formulation experimental data can be used to establish a multivariate response relationship between the anti-metal fouling performance and the four doping variable dimensions, thereby obtaining a performance response surface model covering the entire optimization space.

[0050] Subsequently, the absolute value of the gradient sensitivity of the performance to each doping variable dimension is calculated based on the response relationship.

[0051] Specifically, in the response surface model, the rate of change of the anti-metal fouling performance relative to each doping variable dimension is calculated, and its absolute value is taken as the absolute value of the gradient sensitivity of the corresponding dimension.

[0052] The larger the absolute value of the gradient sensitivity, the more significant the performance change can be caused by a small change in the corresponding doping variable, indicating that the variable has a higher degree of influence on the performance of metal contamination; the smaller the absolute value of the gradient sensitivity, the weaker the influence of the corresponding variable on the performance change.

[0053] For example, the absolute values ​​of the gradient sensitivity for the first, second, third, and fourth doping variable dimensions calculated by the response surface model are 0.82, 0.57, 0.31, and 0.24, respectively. This indicates that the first doping variable dimension has the most significant impact on the performance against metal contamination, while the fourth doping variable dimension has a relatively weaker impact.

[0054] Furthermore, using the maximum value among the absolute values ​​of gradient sensitivity corresponding to all doping variable dimensions as a benchmark, the absolute values ​​of gradient sensitivity corresponding to each doping variable dimension are normalized to obtain the normalized sensitivity.

[0055] Specifically, the normalized sensitivity for a given doping variable dimension is obtained by dividing the absolute value of the gradient sensitivity by the maximum value among all absolute gradient sensitivity values. Where: Normalized sensitivity = Absolute value of gradient sensitivity in the corresponding dimension ÷ Maximum value among the absolute values ​​of gradient sensitivity in all dimensions.

[0056] For example, when the absolute values ​​of the gradient sensitivity corresponding to the four doping variable dimensions are 0.82, 0.57, 0.31 and 0.24 respectively, and the maximum value of 0.82 is used as the normalization benchmark, the corresponding normalized sensitivities are 0.82÷0.82=1.00, 0.57÷0.82=0.695, 0.31÷0.82=0.378 and 0.24÷0.82=0.293 respectively.

[0057] Finally, the normalized sensitivity set corresponding to each doping variable dimension is obtained, and the normalized sensitivity is used as the basis for calculating the anisotropic search step size in step S200.

[0058] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0059] In summary, this step constructs an initial formulation population covering the optimization space and uses response surface methodology to quantify the impact of each doping variable on the performance against metal contamination, obtaining the normalized sensitivity for each variable. Furthermore, it combines the normalized sensitivity to determine differentiated search step sizes for each dimension, forming an anisotropic search step size. This allows for higher search accuracy for key variable dimensions and a larger search range for non-key variable dimensions, thereby improving the search efficiency, convergence speed, and target formulation optimization capability of the subsequent optimization process.

[0060] S300: Based on the anisotropic search step size, an anisotropic distance metric is constructed, and the anisotropic neighborhood corresponding to each initial recipe in the initial recipe population is determined according to the anisotropic distance metric. Step S300 in the method provided in this application embodiment includes: An anisotropic covariance matrix is ​​constructed using the square of the search step size corresponding to each doping variable dimension as the diagonal elements. The off-diagonal elements in the anisotropic covariance matrix are the covariance values ​​between the corresponding doping variable dimensions. The covariance values ​​are calculated based on the coupling coefficient between the corresponding doping variables and the corresponding search step size. Based on the anisotropic covariance matrix, each doping variable dimension is traversed, and the standard deviation is defined with the corresponding search step size as the standard deviation. An independent one-dimensional Gaussian distribution centered on the value of the current formulation in that dimension is formed, thus creating a one-dimensional Gaussian neighborhood. By fusing the probability density of single-dimensional Gaussian neighborhoods of all doping variable dimensions, an anisotropic neighborhood is obtained. Multiple initial formulations are iteratively processed to obtain the anisotropic neighborhood corresponding to each initial formulation.

[0061] In this step, the anisotropic covariance matrix refers to the matrix structure used to characterize the differences in search scales of different doping variable dimensions. Its diagonal elements correspond to the search step size information of each doping variable dimension, and the off-diagonal elements are used to characterize the correlation between different dimensions. A one-dimensional Gaussian neighborhood refers to the probability distribution region established around the current value of a certain doping variable dimension, used to describe the search range of parameter values ​​in the vicinity of that dimension; Anisotropic neighborhood refers to the local search region formed after integrating the search scales of all doping variables, and is used to characterize the parameter space with a high search probability near the current formulation.

[0062] In this embodiment, an anisotropic covariance matrix is ​​first constructed based on the search step size corresponding to each doping variable dimension and the correlation between each doping variable in the historical formula experimental data.

[0063] Specifically, the square of the search step size corresponding to each dopant variable dimension is used as the diagonal element of the covariance matrix. Simultaneously, based on historical formulation experimental data, the degree of synergistic correlation between each dopant variable is statistically analyzed. The coupling coefficient between any two dopant variables is calculated, and the off-diagonal elements in the covariance matrix are determined according to the coupling coefficient and the search step size of the corresponding dopant variable dimension. The covariance value corresponding to any two dopant variable dimensions is calculated as follows: Covariance = Coupling coefficient × First dimension search step size × Second dimension search step size.

[0064] The diagonal elements characterize the search scale of the corresponding dopant variable dimension; the off-diagonal elements characterize the strength of the coupling relationship between different dopant variables. When there is a strong synergistic effect between two dopant variables, the corresponding coupling coefficient takes a large positive value; when there is a competitive or inhibitory effect between two dopant variables, the corresponding coupling coefficient takes a negative value; when the mutual influence between two dopant variables is weak, the corresponding coupling coefficient is close to zero.

[0065] For example, when the doping formulation optimization space includes a first doping variable dimension, a second doping variable dimension, a third doping variable dimension, and a fourth doping variable dimension, and the corresponding search steps are 0.020, 0.035, 0.050, and 0.080, respectively, the diagonal elements of the corresponding covariance matrix are 0.0004, 0.001225, 0.0025, and 0.0064, respectively. For any two doping variable dimensions, the corresponding correlation coefficients are calculated based on historical formulation experimental data, and the corresponding off-diagonal elements are determined according to the coupling coefficient × first dimension search step size × second dimension search step size, thereby forming the corresponding anisotropic covariance matrix.

[0066] Subsequently, based on the anisotropic covariance matrix, the corresponding single-dimensional Gaussian neighborhood is constructed by traversing each doping variable dimension.

[0067] Specifically, for any doping variable dimension in the current formulation, an independent one-dimensional Gaussian distribution is established using the current value corresponding to that dimension as the distribution center and the corresponding search step size as the standard deviation. A smaller search step size results in a more concentrated Gaussian distribution, indicating that a fine-grained search strategy is employed for that dimension; a larger search step size results in a more gradual Gaussian distribution, indicating that a wide-range search strategy is employed for that dimension.

[0068] For example, for a given initial formulation, if the current value of the first doping variable dimension is 0.85wt%, and the corresponding search step size is 0.020, then a one-dimensional Gaussian distribution for the first dimension is established with 0.85wt% as the center and 0.020 as the standard deviation. If the current value of the second doping variable dimension is 1.30wt%, and the corresponding search step size is 0.035, then a one-dimensional Gaussian distribution for the second dimension is established with 1.30wt% as the center and 0.035 as the standard deviation. The remaining dimensions are established with corresponding one-dimensional Gaussian distributions in the same manner.

[0069] Furthermore, the single-dimensional Gaussian neighborhoods corresponding to all dopant variable dimensions are fused to obtain the anisotropic neighborhood corresponding to the current formulation.

[0070] Specifically, since the search step sizes for each doping variable dimension are independent in different covariance matrices, the probability densities of the Gaussian distributions corresponding to each dimension can be jointly calculated to obtain the joint probability density distribution of the current formulation in the multidimensional parameter space. Regions with higher joint probability density represent neighborhoods that are highly similar to the current formulation and have a higher search priority; regions with lower joint probability density represent parameter regions that are far from the current formulation.

[0071] For example, after establishing corresponding one-dimensional Gaussian distributions for each of the four doping variable dimensions, a corresponding four-dimensional anisotropic neighborhood can be obtained by fusing the probability density distributions of the four dimensions. In this case, a narrower probability distribution range is formed in the dimension with a smaller search step size, while a wider probability distribution range is formed in the dimension with a larger search step size, thus forming a search neighborhood with directional differences.

[0072] Finally, the above neighborhood construction process is repeated for multiple initial recipes in the initial recipe population to obtain the anisotropic neighborhood corresponding to each initial recipe.

[0073] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0074] In summary, this step constructs an anisotropic covariance matrix based on the search step size and establishes single-dimensional neighborhoods using the Gaussian distributions corresponding to each dopant variable dimension. These neighborhoods are then further integrated to form multi-dimensional anisotropic neighborhoods. This allows different dopant variable dimensions to participate in the optimization process according to their respective importance and search scale, thereby achieving fine-grained searching in the direction of key variables and broad-range exploration in the direction of non-key variables. This improves the characterization capability of the search space and the accuracy of neighborhood construction, providing a reliable neighborhood foundation for neighborhood difference variation, candidate solution generation, and adaptive search in the subsequent formulation optimization process.

[0075] S400: Construct a multidimensional optimization objective function and combine it with anisotropic neighborhood and pre-constructed doping proxy model to iteratively optimize the initial formulation population. In each iteration, normalization constraint processing is performed until the preset convergence condition is met, and the target doping formulation is output. Step S400 in the method provided in this application embodiment includes: The doping proxy model is updated based on the latest obtained formula experimental data; Construct a multidimensional optimization objective function, which includes at least six sub-objectives: metal adhesion force, surface free energy, Vickers hardness, surface roughness, bending strength, and relative density. Each sub-objective is normalized and then weighted and summed with preset weights. Based on the doping proxy model, the performance of each initial formulation in the initial formulation population is mapped to obtain the multidimensional predicted performance of each initial formulation, and the optimization target value of each initial formulation is calculated by combining the multidimensional optimization objective function. Based on the optimization target value, multiple initial formulas are traversed, and neighborhood differential mutation is performed in the corresponding anisotropic neighborhood. The neighborhood differential mutation results are synthesized, and normalization constraint processing is performed on the neighborhood differential mutation results to obtain multiple synthesized mutation vectors. Multiple synthetic mutation vectors are randomly crossed to obtain multiple cross mutation vectors. The synthetic mutation vectors and cross mutation vectors are used as supplementary solution sets. Each supplementary solution in the supplementary solution set is evaluated based on the doping surrogate model and the multidimensional optimization objective function. Based on the optimization objective value corresponding to each supplementary solution in the evaluation results, a greedy selection is performed, and an adaptive update combining the greedy selection result and the cumulative evolution path is performed on the anisotropic neighborhood. Determine whether the greedy selection result meets the preset convergence condition. If it does, output the current optimal formula as the target doping formula. If it does not meet the condition, return to perform neighborhood difference mutation based on the updated anisotropic neighborhood.

[0076] In this step, the doping proxy model refers to the performance prediction model trained using historical formulation experimental data and the latest experimental data. It is used to characterize the mapping relationship between doping formulation and anti-metal fouling performance and comprehensive mechanical properties, and to achieve rapid prediction of candidate formulation performance. The multidimensional optimization objective function refers to a unified evaluation function constructed by integrating multiple performance indicators. It is used to convert performance indicators such as metal adhesion force, surface free energy, Vickers hardness, surface roughness, bending strength and relative density into unified optimization evaluation indicators to quantify the comprehensive performance level of different formulations. The cumulative evolution path refers to the historical vector that records the evolution direction of excellent solutions during continuous iteration. It is used to reflect the long-term evolution trend of the search process and serves as an important basis for subsequent adaptive updates of the search step size and adjustments to the anisotropic neighborhood.

[0077] In this embodiment, the doping proxy model is first updated based on the latest obtained formula experimental data.

[0078] Specifically, newly validated formulation data and historical formulation experimental data are used together as training samples to incrementally train the pre-constructed doping proxy model, thereby improving the model's prediction accuracy for the current optimization region.

[0079] For example, when the new experimental data includes several sets of formula samples that have completed sintering verification and corresponding performance test results, the new data is incorporated into the original training dataset, and the surrogate model is retrained so that the model can continuously reflect the latest experimental patterns.

[0080] Subsequently, a multidimensional optimization objective function was constructed. Specifically, metal adhesion force, surface free energy, Vickers hardness, surface roughness, flexural strength, and relative density were used as optimization objectives, and each was subjected to dimensionless normalization. Through positive normalization, negative normalization, and objective value normalization, each optimization index was uniformly converted into an evaluation form in which "the larger the value, the better the performance." Then, a weighted sum was performed according to preset weights to obtain the comprehensive optimization objective value.

[0081] Among them, negative normalization is used for indicators that are expected to decrease, such as metal adhesion force and surface free energy; positive normalization is used for indicators that are expected to increase, such as Vickers hardness, bending strength and relative density; and target interval optimization is used for surface roughness. For example, the normalization benchmark can be selected from the statistical extreme values ​​of historical experimental data, such as the 5% and 95% quantiles, or any one of the predicted values ​​of the surrogate model at the boundary of the optimization space.

[0082] Specifically, the surface roughness evaluation value is determined based on a preset target roughness range. When the surface roughness is within the target range, a higher evaluation value is assigned; when the surface roughness deviates from the target range, the corresponding evaluation value is reduced according to the degree of deviation.

[0083] For example, when the target surface roughness is set to 0.18 μm, if the surface roughness of a certain formulation is 0.17 μm or 0.19 μm, a higher roughness evaluation value is obtained; if the surface roughness is 0.05 μm or 0.40 μm, the evaluation value is reduced because it deviates significantly from the target roughness.

[0084] Preferably, a cross-coupling evaluation factor is added to the multidimensional optimization objective function. For example, when both Vickers hardness and relative density are high, these two high values ​​may jointly lead to an increase in micropores, thereby worsening metal adhesion, corresponding to a cross-coupling evaluation factor with synergistic reduction. When metal adhesion is low and surface free energy is low, it has a synergistic gain in resisting fouling, corresponding to a cross-coupling evaluation factor with synergistic gain. The magnitude of the cross-coupling evaluation factor is determined based on the degree of gain or loss of the synergistic effect. Preferably, the weights can be dynamically adjusted according to the performance range of the current formulation. For example, when Vickers hardness exceeds the threshold, its weight is adaptively reduced, and the weight of relative density is increased to avoid excessive pursuit of hardness leading to increased porosity.

[0085] Furthermore, the normalized sub-objectives are weighted and summed according to preset weights to obtain a multidimensional optimization objective function.

[0086] The preset weights are pre-set based on the target requirements of anti-metal fouling performance, bonding stability and mechanical reliability, and the sum of the weights corresponding to each sub-target is 1.

[0087] Specifically, the normalized metal adhesion force evaluation value, surface free energy evaluation value, Vickers hardness evaluation value, surface roughness evaluation value, bending strength evaluation value, and relative density evaluation value are multiplied by their respective weights, and the weighted results are summed to obtain the comprehensive optimization target value.

[0088] For example, when the normalized metal adhesion strength evaluation value of a certain formula is 0.92, the normalized surface free energy evaluation value is 0.85, the normalized Vickers hardness evaluation value is 0.88, the normalized surface roughness evaluation value is 0.90, the normalized flexural strength evaluation value is 0.87, and the normalized relative density evaluation value is 0.95, if the corresponding weights are 0.30, 0.20, 0.15, 0.10, 0.15, and 0.10 respectively, then the product of each evaluation value and the corresponding weight is calculated, and the results are accumulated to obtain the comprehensive optimization target value corresponding to the formula.

[0089] Subsequently, the updated doping proxy model was used to perform performance mapping on each initial formulation in the initial formulation population.

[0090] Specifically, each initial formula is input into the doping proxy model to obtain the corresponding predicted values ​​of metal adhesion force, surface free energy, Vickers hardness, surface roughness, flexural strength and relative density, and these values ​​are then substituted into the multidimensional optimization objective function to calculate the corresponding optimization objective value.

[0091] Furthermore, based on the optimization target value corresponding to each initial formula, neighborhood difference mutation is performed in the corresponding anisotropic neighborhood.

[0092] Specifically, multiple neighboring formulations are selected from the current formulation neighborhood, and the neighborhood difference mutation direction is constructed using the difference vectors of each doping variable between different formulations. The neighborhood difference mutation result is generated by combining the anisotropic search step size of the corresponding variable dimension, thereby generating a new candidate solution.

[0093] Since different dimensions of the anisotropic neighborhood correspond to different search scales, the mutation process can automatically adapt to the importance of different doping variables, performing fine-grained searches in the direction of key variables and large-scale explorations in the direction of non-key variables.

[0094] Furthermore, normalization constraints are applied to the neighborhood difference variation results.

[0095] Specifically, the mass percentage of each dopant component in each candidate formulation generated by neighborhood difference mutation is normalized to ensure that the composition ratio of each dopant component remains constant.

[0096] First, calculate the sum of the mass percentages of each dopant component in the current candidate formulation.

[0097] The percentage of total mass is calculated using the following formula: Total mass percentage = mass percentage of the first doped component + mass percentage of the second doped component + ... + mass percentage of the nth doped component.

[0098] Subsequently, the mass percentage of each dopant component was normalized.

[0099] The normalized value corresponding to the i-th dopant component is calculated according to the following formula: The normalized value of the i-th dopant component = mass percentage of the i-th dopant component ÷ total mass percentage.

[0100] After normalization, the normalized values ​​of each dopant component satisfy: First normalized value + Second normalized value + ... + nth normalized value = 1.

[0101] If any doped component after normalization exceeds the corresponding process allowable range, the doped component is corrected to the corresponding boundary value, and the remaining doped components are normalized again until all doped components meet the composition constraints and process constraints, thereby obtaining a synthetic variation vector that meets the constraints.

[0102] For example, when the mass percentages of the four doped components corresponding to the difference variation results in a certain neighborhood are 2.8%, 1.7%, 3.5%, and 2.4%, respectively, the total mass percentage is 10.4%. After proportional normalization, the normalized values ​​of each doped component are 0.269, 0.163, 0.337, and 0.231, respectively. The sum of the four normalized values ​​is 1, thus satisfying the preset normalization constraint.

[0103] Subsequently, a random dimension crossover operation is performed on multiple synthetic mutation vectors.

[0104] Specifically, one or more doping variable dimensions are randomly selected from multiple synthetic mutation vectors to perform cross-replacement, forming multiple cross-mutation vectors, and the cross-mutation vectors and the original mutation results together constitute an supplementary solution set.

[0105] For example, when a candidate solution contains six doping variable dimensions, two or three of the variable dimensions can be randomly selected to perform cross-substitution, thereby forming a new candidate formulation.

[0106] Furthermore, the doping surrogate model and multidimensional optimization objective function are used to evaluate all supplementary solutions in the supplementary solution set, and the corresponding optimization objective values ​​are calculated.

[0107] Then, a greedy choice is made based on the evaluation results.

[0108] Specifically, the supplementary solution is compared with the current solution. If the optimization objective value corresponding to the supplementary solution is better than that of the current solution, the supplementary solution is used to replace the current solution; otherwise, the current solution is retained and the process proceeds to the next iteration.

[0109] Simultaneously, based on the results of this round of greedy selection and the cumulative evolution path, the anisotropic neighborhood is adaptively updated. Specifically, if performance improvement is achieved in a certain direction for multiple consecutive iterations, the search range in that direction is appropriately expanded; if performance improvement is not achieved in multiple consecutive iterations, the search range in that direction is appropriately narrowed to improve local search accuracy.

[0110] Finally, determine whether the current greedy choice result meets the preset convergence conditions. Convergence conditions may include one or more of the following: the change in the target value is less than a preset threshold after a preset number of consecutive iterations, the maximum number of iterations is reached, or the performance of the optimal solution meets the target requirements.

[0111] For example, when the change in the optimization target value is less than 0.1% for 20 consecutive iterations, or when the number of iterations reaches 500, the optimization process can be determined to meet the convergence condition.

[0112] If the convergence condition is met, the formula with the largest current optimization objective value is output as the target doping formula; if the convergence condition is not met, the neighborhood difference mutation is performed based on the updated anisotropic neighborhood, and the next round of iterative optimization is continued until the final target doping formula is obtained.

[0113] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0114] Step S400 in the method provided in this application embodiment further includes: Based on the search step size of each supplementary solution in each doped variable dimension, an independent Gaussian random perturbation is applied to the corresponding dimension to obtain the Gaussian perturbation vector; In the anisotropic neighborhood corresponding to each supplementary solution, two formulations are randomly selected, the difference vector between the two formulations in each doping variable dimension is calculated, and the difference vector is scaled by a preset scaling factor to obtain the difference vector. Based on the Mahalanobis distance between each supplementary solution and other solutions in the neighborhood, the influence weight of other solutions is calculated using the Gaussian kernel function, and the direction offset vector of other solutions relative to the current recipe is weighted and summed using the influence weight to obtain the neighborhood shared vector. The Gaussian perturbation vector, the difference vector, and the neighborhood shared vector are superimposed to obtain the synthetic mutation vector.

[0115] In this step, the Gaussian perturbation vector refers to a random search vector generated based on the search step size corresponding to the dimension of each doping variable, which is used to enhance the local search capability. The difference vector is a direction vector composed of the parameter differences between two recipes in the neighborhood, which is used to guide the search process to move towards the potentially superior region; The neighborhood shared vector refers to the group guidance direction formed by multiple excellent formulations in the neighborhood; the synthetic mutation vector refers to the final mutation direction vector formed by fusing the Gaussian perturbation vector, the difference vector, and the neighborhood shared vector.

[0116] In this embodiment, firstly, after obtaining the supplementary solution set, a Gaussian perturbation vector is generated based on the search step size of each supplementary solution in each doping variable dimension.

[0117] Specifically, for each doping variable dimension of the current supplementary solution, an independent Gaussian distribution is constructed using the search step size of the corresponding dimension as the standard deviation, and perturbation values ​​are randomly sampled from the corresponding Gaussian distribution. Subsequently, the perturbation values ​​of each dimension are combined to form a Gaussian perturbation vector.

[0118] Since different dimensions correspond to different search step sizes, the magnitude of the perturbations generated by each dimension varies. For key variable dimensions with smaller search step sizes, smaller random perturbations are generated; for variable dimensions with larger search step sizes, relatively larger random perturbations are generated, thereby maintaining the consistency of the search process with the anisotropic neighborhood.

[0119] Subsequently, two neighborhood formulations are randomly selected from the anisotropic neighborhood corresponding to the current supplementary solution, and the parameter differences between the two in each doping variable dimension are calculated to obtain the difference vector.

[0120] Furthermore, the difference vector is scaled using a preset scaling factor to obtain the difference vector.

[0121] Specifically, for any dimension of the doping variable, its difference components can be expressed as: Differential component = scaling factor × (second neighborhood recipe parameter value - first neighborhood recipe parameter value) The scaling factor is used to control the intensity of the differential search.

[0122] For example, when the parameter value corresponding to the first neighborhood recipe in a certain dimension is 0.80wt%, the parameter value corresponding to the second neighborhood recipe is 1.00wt%, and the scaling factor is 0.5, then the difference component corresponding to that dimension is 0.10wt%.

[0123] Furthermore, a neighborhood shared vector is constructed based on the Mahalanobis distance between the current supplementary solution and other formulations in the neighborhood. Specifically, the Mahalanobis distance between the current supplementary solution and each neighborhood formulation is first calculated using the anisotropic covariance matrix.

[0124] The smaller the Mahalanobis distance, the closer the corresponding neighborhood formula is to the current formula; the larger the Mahalanobis distance, the more significant the difference between the corresponding neighborhood formula and the current formula.

[0125] Subsequently, the Mahalanobis distance was substituted into the Gaussian kernel function to calculate the influence weight corresponding to each neighborhood formulation. Among them, the neighborhood formulations with closer distances have larger influence weights, while the neighborhood formulations with farther distances have smaller influence weights.

[0126] Furthermore, the directional offset vectors of other recipes in the neighborhood relative to the current recipe are weighted and summed using the influence weights to obtain the neighborhood shared vector.

[0127] Specifically, for the current recipe, its neighborhood shared vector is calculated according to the following formula: The neighborhood shared vector = the sum of the "influence weight × direction offset vector" of all neighborhood recipes ÷ the sum of the influence weights of all neighborhood recipes. Where: Directional offset vector = neighborhood recipe parameter vector - current recipe parameter vector.

[0128] For example, when there are multiple neighborhood solutions with optimization objective values ​​better than the current recipe, the neighborhood shared vector will point more towards the direction where these superior recipes cluster, thereby increasing the probability of the search process converging towards high-quality regions.

[0129] Finally, the Gaussian perturbation vector, the difference vector, and the neighborhood shared vector are superimposed to obtain the final synthetic mutation vector. Specifically: The composite mutation vector = Gaussian perturbation vector + difference vector + neighborhood shared vector.

[0130] Among them, the Gaussian perturbation vector provides local random exploration capability; the difference vector provides information exchange capability between populations; and the neighborhood sharing vector provides collaborative guidance capability for excellent solutions in the neighborhood.

[0131] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0132] Step S400 in the method provided in this application embodiment further includes: Based on the evaluation results, supplementary solutions whose improvement in the target value compared to the worst individual in the initial formula population exceeds a preset improvement threshold are selected as successful mutation solutions. Calculate the ratio of the variation length of each successful mutation solution relative to the parent solution in each doping variable dimension to the current search step size in that dimension, and normalize it based on the preset path learning rate to obtain the path increment value corresponding to each doping variable dimension. Based on the preset path learning rate, the cumulative evolution path of each doping variable dimension in the parent solution is weighted and summed to obtain the cumulative evolution path of each doping variable dimension in the current iteration. Based on the ratio of the absolute value of the cumulative evolution path to the expected modulus of the standard Gaussian distribution corresponding to that dimension, the search step size of each doped variable dimension is updated multiplicatively in the form of a natural exponential. Based on the doping surrogate model and successful mutation solution, the gradient sensitivity of each doping variable dimension is calculated, and the normalized sensitivity of each doping variable dimension is updated according to the gradient sensitivity. Based on the updated normalized sensitivity, the sensitivity of the multiplicative updated search step size is corrected, and the anisotropic neighborhood is updated based on the corrected search step size.

[0133] In this step, a successful mutation solution refers to a supplementary solution whose optimization target value reaches the preset improvement requirement after evaluation by the doping proxy model. It is used to characterize the effective evolutionary direction generated in the current iteration process. The cumulative evolutionary path refers to the historical vector that records the successful search directions during continuous iterations. It is used to reflect the long-term evolutionary trend in the optimization process and guide the subsequent adjustment of the search scale. Sensitivity correction refers to the process of adjusting the search step size differently based on the degree of influence of each doping variable dimension on the target performance, in order to achieve adaptive search of key and non-key variables.

[0134] In this embodiment, firstly, after evaluating the supplementary solution set, successful mutation solutions are determined based on the optimization target value. Specifically, the optimization target value corresponding to each supplementary solution is compared with the optimization target value corresponding to the worst individual in the current initial formula population, and the improvement amount between the two is calculated. Wherein: Improvement Amount = Target value of supplementary solution optimization - Target value of the worst individual in the current population optimization.

[0135] When the improvement exceeds the preset improvement threshold, the corresponding supplementary solution is determined to be a successful mutation solution; otherwise, it is determined to be an ordinary supplementary solution and will not participate in the subsequent neighborhood adaptive update process.

[0136] For example, when the optimization objective value corresponding to the worst individual in the current population is 0.62 and the preset improvement threshold is 0.05, and the optimization objective value corresponding to a certain supplementary solution is 0.71, its improvement amount is 0.09, which is greater than the preset improvement threshold. Therefore, the supplementary solution is identified as a successful mutation solution.

[0137] Subsequently, the path increment value corresponding to the successful mutation solution is calculated. Specifically, for each successful mutation solution, its variation length relative to the parent solution in each doping variable dimension is calculated, and the variation length is calculated as a ratio to the current search step size in the corresponding dimension to obtain the standardized mutation magnitude.

[0138] Furthermore, the standardized variation magnitude is normalized by combining the preset path learning rate to obtain the path increment value for the corresponding dimension. Wherein: Path increment value = Path learning rate × (variable asynchronous length ÷ current search step size).

[0139] The above processing can eliminate the influence of differences in the units of different variables and search scales on the path update results.

[0140] Subsequently, the cumulative evolutionary path is updated based on the path learning rate. Specifically, the path increment value obtained in the current iteration is weighted and fused with the cumulative evolutionary path saved in the previous generation to obtain the cumulative evolutionary path corresponding to the current iteration. Wherein: Current cumulative evolutionary path = (1 - path learning rate) × previous generation cumulative evolutionary path + path increment value.

[0141] By incorporating historical path information, the search direction can not only depend on the result of the current mutation, but also inherit the evolutionary trend that has been continuously effective in multiple rounds of optimization.

[0142] Preferably, based on the objective laws of materials science, when weighting and fusing the path increment value obtained in the current iteration with the cumulative evolution path saved in the previous generation, the influence of strong coupling relationships between dopant variables is introduced. The off-diagonal elements in the anisotropic covariance matrix are used as coupling influence factors between dimensions. Combined with the path increment value of each dopant variable dimension, a weighted calculation considering the coupling effect is performed, where the weights are determined based on multiple coupling influence factors. Specifically, the calculation objects for the coupled path increment value are selected based on a set of coupled variables defined under prior knowledge, ensuring that when the step size of a certain dimension within the group changes significantly, the step sizes of other related dimensions within the group are adjusted synchronously.

[0143] Furthermore, the search step size for each dimension is updated based on the cumulative evolutionary path. Specifically, the ratio between the absolute value of the current cumulative evolutionary path and the expected modulus of the corresponding standard Gaussian distribution is calculated, and a multiplicative update is performed using the natural exponential function. Wherein: Updated search step size = current search step size × Nature Index (Degree of path deviation).

[0144] The degree of path deviation is determined by the ratio between the absolute value of the cumulative evolutionary path and the expected modulus of the standard Gaussian distribution.

[0145] When the cumulative evolutionary path continuously increases, it indicates that the current search direction has strong consistency, and the search step size increases accordingly; when the cumulative evolutionary path continuously decreases, it indicates that the search direction tends to fluctuate randomly, and the search step size decreases accordingly, thereby improving the local search accuracy.

[0146] For example, when the absolute value of the cumulative evolutionary path in a certain dimension is greater than the expected modulus of the corresponding standard Gaussian distribution, the natural exponent term is greater than 1, and the corresponding search step size is amplified; conversely, the corresponding search step size is reduced.

[0147] Subsequently, the normalized sensitivity is updated based on the doping surrogate model and the successful mutation solution. Specifically, the successful mutation solution is input into the doping surrogate model, and the gradient rate of change of the target performance with respect to each doping variable dimension is calculated, and the corresponding gradient sensitivity is obtained.

[0148] Furthermore, the maximum gradient sensitivity among all current doping variable dimensions is used as the normalization benchmark to normalize the gradient sensitivity of each dimension, thereby obtaining the updated normalized sensitivity.

[0149] The higher the normalization sensitivity, the more significant the impact of the corresponding variable on the target performance; the lower the normalization sensitivity, the lower the importance of the corresponding variable.

[0150] Finally, the search step size after the multiplicative update is adjusted based on the updated normalized sensitivity, and the anisotropic neighborhood is updated accordingly.

[0151] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0152] Step S400 in the method provided in this application embodiment further includes: Acquire historical formula experimental data, which includes multiple historical formulas and their corresponding historical performance test data; Based on historical formulation experimental data, a doping proxy model based on a regression model was constructed; Based on historical formulation experimental data, a doping proxy model was trained and validated.

[0153] First, historical formulation experimental data were obtained. Specifically, the historical formulation experimental data included multiple historical formulations and their corresponding historical performance test data. Each set of historical formulations included the mass percentage of the alumina matrix and each dopant component. The historical performance test data included performance indicators such as metal adhesion force, surface free energy, Vickers hardness, flexural strength, relative density, and surface roughness after sintering of the corresponding formulation.

[0154] In this embodiment, the historical formula experimental data came from the enterprise's R&D experimental database, pilot production database, and laboratory testing database, and a total of 860 sets of historical experimental samples were collected.

[0155] Subsequently, the historical formulation experimental data were preprocessed. Specifically, the content of each dopant component was normalized, and the normalized value was calculated according to the following formula: Normalized value = (current variable value - historical minimum value) ÷ (historical maximum value - historical minimum value) The historical maximum and historical minimum values ​​are derived from the statistical results of all historical samples.

[0156] Furthermore, to avoid outlier experimental data reducing model fitting accuracy, outlier removal was performed on historical performance test data. First, the mean and standard deviation of each performance index were calculated. When a sample satisfies: Performance deviation > three standard deviations.

[0157] The sample is then identified as an abnormal sample and deleted.

[0158] Among them, three standard deviations are derived from the statistical Pauta criterion (3σ criterion), which can cover approximately 99.73% of normal sample data and effectively eliminate experimental errors and detect outliers.

[0159] After data preprocessing, a doping surrogate model is constructed. In this embodiment, a Gaussian process regression model is preferably used as the doping surrogate model. Since the number of ceramic doping experimental samples is limited and there is a nonlinear coupling relationship between the doping components, the Gaussian process regression model can achieve high prediction accuracy under small sample conditions and can output predicted values ​​and prediction confidence levels, thus making it suitable as a doping surrogate model.

[0160] Specifically, the doping proxy model includes an input layer, a Gaussian process mapping layer, and an output layer. The input layer is used to input the mass percentage of each dopant component. The Gaussian process mapping layer uses a Matérn 5 / 2 kernel function to establish a nonlinear mapping relationship between the input variables and performance indicators. The output layer outputs the predicted values ​​of each performance indicator. In this embodiment, corresponding Gaussian process regression models are established for metal adhesion force, Vickers hardness, surface free energy, flexural strength, and relative density, respectively.

[0161] Subsequently, the doping surrogate model was trained using historical formulation experimental data. Specifically, all samples were divided into a 70% training set, a 15% validation set, and a 15% test set. The training set was used for model parameter learning, the validation set for adjusting model parameters, and the test set for validating model performance. During model training, the negative logarithmic marginal likelihood value was used as the objective function, and the L-BFGS optimization algorithm was used to iteratively update parameters such as the kernel function length scale, signal variance, and noise variance. Model training ended when the objective function decreased by less than 0.001 after 20 consecutive iterations. The 0.001 threshold was derived from historical training results; when the objective function decreased below this value, the model's prediction accuracy improved by less than 0.2%, making further training less meaningful, and therefore it was used as the convergence threshold.

[0162] After model training, the doped surrogate model is validated. Specifically, the coefficient of determination, root mean square error, and mean relative error of the model are calculated, and a coefficient of determination of no less than 0.95 and a mean relative error of no more than 5% are set as model validation criteria. The thresholds for both the coefficient of determination and the mean relative error are determined through cross-validation using historical formulation experimental data. This embodiment tests the model prediction performance corresponding to different validation thresholds. When the coefficient of determination reaches above 0.95 and the mean relative error is controlled within 5%, the model prediction error is lower than the repeated test error of the material sintering experiment, and it also has good generalization ability. Therefore, the above thresholds are used as the model validation criteria.

[0163] For example, this embodiment uses 860 sets of historical formulation experimental data to complete model training, of which 602 sets are used as the training set, 129 sets as the validation set, and 129 sets as the test set. After 178 iterations, the model converges, and the final model determination coefficient reaches 0.972, with an average relative error of 3.68%, meeting the preset validation criteria. Therefore, the trained doping surrogate model is used for rapid prediction of the performance of each candidate formulation in the subsequent formulation optimization process.

[0164] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0165] In summary, this step establishes a search neighborhood that matches the importance of each doping variable by constructing an anisotropic covariance matrix. It then combines a Gaussian process regression surrogate model, a multidimensional optimization objective function, and a neighborhood difference mutation mechanism to rapidly predict and iteratively optimize the performance of candidate doping formulations. Simultaneously, it enhances global exploration and local development capabilities by combining Gaussian perturbation, difference search, and neighborhood sharing guidance. Furthermore, it adaptively updates the search step size and anisotropic neighborhood based on successful mutation solutions, cumulative evolution paths, and sensitivity correction mechanisms. This achieves synergistic optimization of anti-metal contamination performance, surface properties, and mechanical properties. While reducing the number of experiments and R&D costs, it improves the search efficiency, optimization accuracy, and convergence stability of the target doping formulation, ultimately obtaining a high-performance alumina-based ceramic cleaver doping formulation.

[0166] S500: According to the target doping formula, an alumina-based ceramic chopping knife resistant to metal contamination is produced through a preset process flow treatment.

[0167] In this embodiment, firstly, the raw materials are prepared according to the target doping formula output in step S400. Specifically, high-purity α-alumina powder is used as the matrix material, and the raw materials to be mixed are accurately weighed according to the mass percentage of each doping component determined by the target doping formula.

[0168] For example, when the target doping formulation includes an alumina matrix and several doping components, each raw material can be weighed according to its corresponding mass percentage, and the weighing error can be controlled within a preset range to ensure the consistency of the composition in the subsequent preparation process.

[0169] Subsequently, the raw materials to be mixed are subjected to a uniform mixing process. Specifically, alumina powder and each dopant component are added together into a mixing device and ball-milled using a dispersion medium to ensure that each dopant component is evenly distributed among the alumina particles.

[0170] Further, the mixed slurry is dried and granulated. Specifically, the liquid phase medium in the slurry is removed by spray drying, vacuum drying, or freeze drying to obtain granulated powder that meets the requirements of subsequent molding.

[0171] Subsequently, the green body is formed using granulated powder. Specifically, ceramic green bodies can be prepared by methods such as compression molding, cold isostatic pressing, injection molding, or tape casting.

[0172] For example, for cleaving products, a combination of molding and cold isostatic pressing is preferred to improve the uniformity of the blank density and dimensional stability.

[0173] Furthermore, the formed ceramic green body undergoes degreasing and sintering treatment. Specifically, organic additives in the green body are removed by gradually increasing the temperature, and densification sintering is completed under a preset sintering regime, so that the alumina matrix and various doped components form a stable microstructure.

[0174] The sintering temperature, holding time, and heating / cooling rates can be adjusted according to the specific doping system to obtain the target grain size and density.

[0175] Subsequently, the sintered ceramic blank undergoes machining and precision finishing. Specifically, the working end structure of the cleaver is formed through processes such as grinding, ultra-precision grinding, polishing, and tip trimming, ensuring that the tip size, surface roughness, and geometric accuracy meet the usage requirements.

[0176] For example, a diamond grinding wheel can be used to finely grind the sintered ceramic blank, and combined with an ultra-precision polishing process to obtain a nanoscale surface roughness.

[0177] Finally, the finished ceramic chopping knife undergoes performance testing. When the metal adhesion force, surface free energy, Vickers hardness, bending strength, and relative density meet the preset requirements, an alumina-based ceramic chopping knife resistant to metal contamination is obtained.

[0178] In summary, this step applies the optimized target doping formulation to the entire process of raw material preparation, mixing and dispersion, granulation, debinding and sintering, and precision machining of alumina-based ceramic cleavers. This achieves uniform distribution and stable solid solution of the doping components in the ceramic matrix, forming a dense and uniform microstructure. Simultaneously, combined with precise sintering control and ultra-precision machining, the cleaver possesses lower metal adhesion tendency, lower surface free energy, higher hardness and flexural strength, and higher density. This effectively improves the cleaver's resistance to metal contamination, dimensional stability, and long-term service life during semiconductor packaging bonding, achieving a reliable transformation into high-performance, metal-contamination-resistant alumina-based ceramic cleaver products with optimized formulation.

[0179] Example 2, as Figure 3 As shown, this invention provides a system for preparing an alumina-based ceramic chopping knife resistant to metal contamination. The system includes: The optimization space construction module 11 is used to define the doping formula optimization space of alumina-based ceramics based on the prior experience of the target production line. The doping formula optimization space includes the range of mass percentage values ​​of each doping component. The initial population generation module 12 is used to generate an initial formulation population in the doping formulation optimization space, and initialize the search step size for each doping variable dimension in the doping formulation optimization space to obtain the anisotropic search step size. The neighborhood construction module 13 is used to construct an anisotropic distance metric based on the anisotropic search step size, and determine the anisotropic neighborhood corresponding to each initial recipe in the initial recipe population based on the anisotropic distance metric. The formulation optimization module 14 is used to construct a multidimensional optimization objective function and combine anisotropic neighborhood with a pre-constructed doping proxy model to iteratively optimize the initial formulation population. In each iteration, normalization constraint processing is performed until the preset convergence condition is met, and the target doping formulation is output. The process preparation module 15 is used to prepare an alumina-based ceramic chopping knife resistant to metal contamination by processing it through a preset process flow according to the target doping formula.

[0180] In this embodiment of the invention, the optimized space construction module 11 is further configured to: High-purity α-alumina with a purity of not less than 99.5% and an average particle size of 0.3 to 0.8 micrometers is used as the matrix material; Obtain historical formula experimental data, determine the composition of doping components based on historical formula experimental data, and define the corresponding doping formula optimization space. The doping formula optimization space includes multiple dimensions, and each dimension corresponds to the mass percentage of a doping component. Based on historical formulation experimental data, combined with statistical analysis methods and preset value amplification factors, the mass percentage range of each dopant component in the doping formulation optimization space is set.

[0181] In this embodiment of the invention, the initial population generation module 12 is further configured to: Based on historical formulation experimental data, an initial formulation population containing multiple initial formulations is generated in the optimization space. The mass percentage of each doped component in each initial formulation is normalized to meet the preset normalization constraints. Based on historical formulation experimental data, the normalized sensitivity corresponding to each doping variable dimension is calculated. Based on the normalized sensitivity and the preset basic step size scaling factor, the initial search step size corresponding to each doping variable dimension is calculated to obtain the anisotropic search step size.

[0182] Based on historical formulation experimental data, the normalized sensitivity corresponding to each doping variable dimension is calculated, including: Based on historical formulation experimental data and combined with response surface fitting method, the response relationship between anti-metal fouling performance and each dimension of doping variables was established. Based on the response relationship, the absolute value of the gradient sensitivity of the performance to each dimension of the doping variable is calculated; Using the maximum absolute value of the gradient sensitivity across all doping variable dimensions as a benchmark, the absolute value of the gradient sensitivity for each doping variable dimension is normalized to obtain the normalized sensitivity.

[0183] In this embodiment of the invention, the neighborhood construction module 13 is further configured to: An anisotropic covariance matrix is ​​constructed using the square of the search step size corresponding to each doping variable dimension as the diagonal elements. The off-diagonal elements in the anisotropic covariance matrix are the covariance values ​​between the corresponding doping variable dimensions. The covariance values ​​are calculated based on the coupling coefficient between the corresponding doping variables and the corresponding search step size. Based on the anisotropic covariance matrix, each doping variable dimension is traversed, and the standard deviation is defined with the corresponding search step size as the standard deviation. An independent one-dimensional Gaussian distribution centered on the value of the current formulation in that dimension is formed, thus creating a one-dimensional Gaussian neighborhood. By fusing the probability density of single-dimensional Gaussian neighborhoods of all doping variable dimensions, an anisotropic neighborhood is obtained. Multiple initial formulations are iteratively processed to obtain the anisotropic neighborhood corresponding to each initial formulation.

[0184] In this embodiment of the invention, the formula optimization module 14 is further used for: The doping proxy model is updated based on the latest obtained formula experimental data; Construct a multidimensional optimization objective function, which includes at least six sub-objectives: metal adhesion force, surface free energy, Vickers hardness, surface roughness, bending strength, and relative density. Each sub-objective is normalized and then weighted and summed with preset weights. Based on the doping proxy model, the performance of each initial formulation in the initial formulation population is mapped to obtain the multidimensional predicted performance of each initial formulation, and the optimization target value of each initial formulation is calculated by combining the multidimensional optimization objective function. Based on the optimization target value, multiple initial formulas are traversed, and neighborhood differential mutation is performed in the corresponding anisotropic neighborhood. The neighborhood differential mutation results are synthesized, and normalization constraint processing is performed on the neighborhood differential mutation results to obtain multiple synthesized mutation vectors. Multiple synthetic mutation vectors are randomly crossed to obtain multiple cross mutation vectors. The synthetic mutation vectors and cross mutation vectors are used as supplementary solution sets. Each supplementary solution in the supplementary solution set is evaluated based on the doping surrogate model and the multidimensional optimization objective function. Based on the optimization objective value corresponding to each supplementary solution in the evaluation results, a greedy selection is performed, and an adaptive update combining the greedy selection result and the cumulative evolution path is performed on the anisotropic neighborhood. Determine whether the greedy selection result meets the preset convergence condition. If it does, output the current optimal formula as the target doping formula. If it does not meet the condition, return to perform neighborhood difference mutation based on the updated anisotropic neighborhood.

[0185] In summary, this invention, by setting up an optimization space construction module, an initial population generation module, a neighborhood construction module, a formulation optimization module, and a process preparation module, first constructs a doping formulation optimization space based on historical formulation experimental data, and establishes an anisotropic search step size by combining normalized sensitivity; further, it utilizes the anisotropic covariance matrix to construct an anisotropic neighborhood that matches the importance of each doping variable, realizing differentiated search for different doping variables; simultaneously, it uses a Gaussian process regression surrogate model, a multidimensional optimization objective function, neighborhood difference mutation, random dimension crossover, and a greedy selection mechanism to efficiently iteratively optimize candidate formulations, and combines cumulative evolution path to achieve adaptive adjustment of search scale and search neighborhood; finally, it obtains a target doping formulation that balances low metal adhesion, low surface free energy, high hardness, high flexural strength, high relative density, and excellent surface quality, and completes the preparation of anti-metal-fouling alumina-based ceramic chopping blades through the process preparation module, thereby effectively reducing the problems of numerous experiments, long development cycles, and low optimization efficiency in traditional formulation development, and improving the intelligence level, optimization accuracy, and engineering application value of high-performance alumina-based ceramic chopping blade formulation design.

[0186] It should be noted that the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0187] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0188] This specification and accompanying drawings are merely illustrative examples of the invention and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its scope. Therefore, if such modifications and modifications fall within the scope of the invention and its equivalents, the invention is intended to include these modifications and modifications.

Claims

1. A method for preparing an aluminum oxide-based ceramic chopper knife resistant to metal contamination, characterized in that, include: Based on prior experience of the target production line, the optimization space of doping formulation for alumina-based ceramics is defined. The optimization space of doping formulation includes the range of mass percentage values ​​of each doping component, and the sum of the mass percentages of each doping component satisfies a preset normalization constraint. An initial formulation population is generated in the optimization space, and a search step size is initialized for each doping variable dimension in the doping formulation optimization space to obtain the anisotropic search step size. Based on the anisotropic search step size, an anisotropic distance metric is constructed, and the anisotropic neighborhood corresponding to each initial formula in the initial formula population is determined according to the anisotropic distance metric. A multidimensional optimization objective function is constructed, and the initial formulation population is iteratively optimized by combining the anisotropic neighborhood with the pre-constructed doping proxy model. Normalization constraint processing is performed in each iteration until the preset convergence condition is met, and the target doping formulation is output. According to the target doping formula, an alumina-based ceramic chopping knife resistant to metal contamination is obtained through a preset process flow treatment.

2. The method for preparing an alumina-based ceramic chopping knife resistant to metal contamination as described in claim 1, characterized in that, Based on prior experience from the target production line, an optimization space for the doping formulation of alumina-based ceramics is defined. This optimization space includes the range of mass percentage values ​​for each dopant component, including: High-purity α-alumina with a purity of not less than 99.5% and an average particle size of 0.3 to 0.8 micrometers is used as the matrix material; Historical formulation experimental data is obtained, the composition of doping components is determined based on the historical formulation experimental data, and a doping formulation optimization space is defined accordingly. The doping formulation optimization space includes multiple dimensions, and each dimension corresponds to the mass percentage of a doping component. Based on the historical formulation experimental data, combined with statistical analysis methods and preset value amplification factors, the mass percentage range of each dopant component in the doping formulation optimization space is set respectively.

3. The method for preparing an alumina-based ceramic chopping knife resistant to metal contamination as described in claim 1, characterized in that, An initial formulation population is generated within the optimization space, and a search step size is initialized for each doping variable dimension in the doping formulation optimization space to obtain the anisotropic search step size, including: Based on historical formulation experimental data, an initial formulation population containing multiple initial formulations is generated within the optimization space, wherein the mass percentage of each doped component in each initial formulation satisfies a preset normalization constraint after normalization processing. Based on the historical formulation experimental data, the normalized sensitivity corresponding to each doping variable dimension is calculated. Based on the normalized sensitivity and the preset basic step size scaling factor, the initial search step size corresponding to each doping variable dimension is calculated to obtain the anisotropic search step size.

4. The method for preparing an alumina-based ceramic chopping knife resistant to metal contamination as described in claim 3, characterized in that, Based on the historical formulation experimental data, the normalized sensitivity corresponding to each doping variable dimension is calculated, including: Based on the historical formulation experimental data, and combined with the response surface fitting method, the response relationship between the anti-metal fouling performance and each doping variable dimension was established. Based on the aforementioned response relationship, the absolute value of the gradient sensitivity of the performance to each doping variable dimension is calculated; Using the maximum absolute value of the gradient sensitivity across all doping variable dimensions as a benchmark, the absolute value of the gradient sensitivity for each doping variable dimension is normalized to obtain the normalized sensitivity.

5. The method for preparing an alumina-based ceramic chopping knife resistant to metal contamination as described in claim 1, characterized in that, Based on the anisotropic search step size, an anisotropic distance metric is constructed, and the anisotropic neighborhood corresponding to each initial recipe in the initial recipe population is determined according to the anisotropic distance metric, including: An anisotropic covariance matrix is ​​constructed using the square of the search step size corresponding to each doping variable dimension as the diagonal elements. The off-diagonal elements in the anisotropic covariance matrix are the covariance values ​​between the corresponding doping variable dimensions. The covariance values ​​are calculated based on the coupling coefficient between the corresponding doping variables and the corresponding search step size. Based on the anisotropic covariance matrix, each doping variable dimension is traversed, and with the corresponding search step size as the standard deviation, an independent one-dimensional Gaussian distribution centered on the value of the current formulation in that dimension is defined to form a single-dimensional Gaussian neighborhood. The probability density of the single-dimensional Gaussian neighborhood of all doping variable dimensions is fused to obtain the anisotropic neighborhood. Multiple initial formulations are iteratively processed to obtain the anisotropic neighborhood corresponding to each initial formulation.

6. The method for preparing an alumina-based ceramic chopping knife resistant to metal contamination as described in claim 1, characterized in that, A multidimensional optimization objective function is constructed, and combined with the anisotropic neighborhood and the pre-constructed doping surrogate model, the initial formulation population is iteratively optimized until a preset convergence condition is met, outputting the target doping formulation, including: The doping proxy model is updated based on the latest obtained formula experimental data; A multidimensional optimization objective function is constructed, which includes at least six sub-objectives: metal adhesion force, surface free energy, Vickers hardness, surface roughness, bending strength, and relative density. Each sub-objective is normalized and then weighted and summed with a preset weight. Based on the doping proxy model, a performance mapping is performed on each initial formulation in the initial formulation population to obtain the multidimensional predicted performance corresponding to each initial formulation. Combined with the multidimensional optimization objective function, the optimization objective value corresponding to each initial formulation is calculated. Based on the optimization target value, multiple initial formulas are traversed, and neighborhood differential mutation is performed in the corresponding anisotropic neighborhood. The neighborhood differential mutation results are synthesized, and normalization constraint processing is performed on the neighborhood differential mutation results to obtain multiple synthesized mutation vectors. Random dimension crosses are performed on multiple synthetic mutation vectors to obtain multiple cross mutation vectors. The synthetic mutation vectors and the cross mutation vectors are used as supplementary solution sets. Each supplementary solution in the supplementary solution set is evaluated based on the doping surrogate model and the multidimensional optimization objective function. Based on the optimization objective value corresponding to each supplementary solution in the evaluation results, a greedy selection is performed, and the anisotropic neighborhood is adaptively updated by combining the greedy selection result and the cumulative evolution path. Determine whether the greedy selection result satisfies the preset convergence condition. If it does, output the current optimal formula as the target doping formula. If it does not, return to perform neighborhood difference mutation based on the updated anisotropic neighborhood.

7. The method for preparing an alumina-based ceramic chopping knife resistant to metal contamination as described in claim 6, characterized in that, Based on the optimization target value, multiple initial formulas are traversed, and neighborhood differential mutation is performed in the corresponding anisotropic neighborhood. Neighborhood differential mutation results are then synthesized, and normalization constraint processing is applied to these results to obtain multiple synthesized mutation vectors, including: Based on the search step size of each supplementary solution in each doped variable dimension, an independent Gaussian random perturbation is applied to the corresponding dimension to obtain the Gaussian perturbation vector; In the anisotropic neighborhood corresponding to each supplementary solution, two formulations are randomly selected, the difference vector between the two formulations in each doping variable dimension is calculated, and the difference vector is scaled by a preset scaling factor to obtain the difference vector. Based on the Mahalanobis distance between each supplementary solution and other solutions in the neighborhood, the influence weight of the other solutions is calculated using a Gaussian kernel function, and the influence weight is used to weight and sum the directional offset vector of the other solutions relative to the current formula to obtain the neighborhood shared vector. The Gaussian perturbation vector, the difference vector, and the neighborhood shared vector are superimposed to obtain the synthetic mutation vector.

8. The method for preparing an alumina-based ceramic chopping knife resistant to metal contamination as described in claim 6, characterized in that, Based on the optimization objective value corresponding to each supplementary solution in the evaluation results, a greedy selection is performed, and the anisotropic neighborhood is adaptively updated by combining the greedy selection result and the cumulative evolutionary path, including: Based on the evaluation results, supplementary solutions whose optimization target value improves by more than a preset improvement threshold compared to the optimization target value of the worst individual in the initial formula population are selected as successful mutation solutions. Calculate the ratio of the variation length of each successful mutation solution relative to the parent solution in each doping variable dimension to the current search step size in that dimension, and normalize it based on the preset path learning rate to obtain the path increment value corresponding to each doping variable dimension. Based on the preset path learning rate, the path increment value and the cumulative evolution path of the corresponding doping variable dimension in the parent solution are weighted and summed to obtain the cumulative evolution path of each doping variable dimension in the current iteration. Based on the ratio of the absolute value of the cumulative evolution path to the expected modulus of the standard Gaussian distribution corresponding to that dimension, the search step size of each doped variable dimension is multiplicatively updated in the form of a natural exponential. Based on the doping proxy model and the successful mutation solution, the gradient sensitivity of each doping variable dimension is calculated, and the normalized sensitivity of each doping variable dimension is updated according to the gradient sensitivity. Based on the updated normalized sensitivity, the search step size after the multiplicative update is adjusted for sensitivity, and the anisotropic neighborhood is updated based on the adjusted search step size.

9. The method for preparing an alumina-based ceramic chopping knife resistant to metal contamination as described in claim 1, characterized in that, The pre-construction of the doped proxy model includes: Acquire historical formula experimental data, which includes multiple historical formulas and their corresponding historical performance test data; Based on the historical formulation experimental data, a doping proxy model based on a regression model is constructed; The doping proxy model is trained and validated based on the historical formulation experimental data.

10. A system for preparing an alumina-based ceramic chopping knife resistant to metal contamination, characterized in that, A method for preparing an alumina-based ceramic chopping knife resistant to metal contamination as described in any one of claims 1 to 9 includes: An optimization space construction module is used to define the optimization space of doping formulation for alumina-based ceramics based on the prior experience of the target production line. The optimization space of doping formulation includes the range of mass percentage values ​​of each doping component. The initial population generation module is used to generate an initial formulation population in the doping formulation optimization space, and initialize the search step size for each doping variable dimension in the doping formulation optimization space to obtain the anisotropic search step size. The neighborhood construction module is used to construct an anisotropic distance metric based on the anisotropic search step size, and determine the anisotropic neighborhood corresponding to each initial recipe in the initial recipe population according to the anisotropic distance metric. The formulation optimization module is used to construct a multidimensional optimization objective function and combine the anisotropic neighborhood with the pre-constructed doping proxy model to iteratively optimize the initial formulation population. In each iteration, normalization constraint processing is performed until the preset convergence condition is met, and the target doping formulation is output. The process preparation module is used to prepare an alumina-based ceramic chopping knife resistant to metal contamination by processing it through a preset process flow according to the target doping formula.

Images