An intelligent integrated design method for vacuum sintering process of super-large ceramic components

Abstract

This application proposes an intelligent integrated design method for vacuum sintering processes of ultra-large ceramic components, belonging to the field of vacuum furnace technology. The method includes: obtaining DFT aggregation parameters based on information about the sintering aids to be optimized, the correlation between the sintering aid information and the component chemical formulas of the DFT calculation object, and a second database; inputting the input features into the optimal machine learning model for prediction to obtain the relative density; optimizing the particle swarm optimization algorithm to find multiple candidate combinations; employing an active learning strategy to evaluate these combinations to obtain the uncertainty; extracting the N candidate combinations with the highest uncertainty as the verification objects for sintering experiments; and outputting the optimal combination of sintering aid information and sintering process parameters. This application can rapidly construct a preliminary and reliable composition-process-performance prediction model under limited data conditions, significantly shortening the initial research cycle of process development.

Description

Technical Field

[0001] This application relates to the field of vacuum furnace technology, and in particular to an intelligent integrated design method for vacuum sintering process of ultra-large ceramic components. Background Technology

[0002] With the rapid development of high-end equipment manufacturing, aerospace, and new energy fields, the demand for high-performance, large-size, and complex-structure ceramic components is becoming increasingly urgent. Vacuum sintering technology, due to its ability to effectively avoid material oxidation, inhibit harmful reactions, promote densification, and obtain uniform microstructures, has become a key process for preparing high-performance ceramics. The scaling up of vacuum sintering is a key technological focus, meeting the requirements for high-temperature resistance, corrosion resistance, and high strength of high-end materials under special working conditions. Designing intelligent integrated processes for vacuum sintering of silicon carbide ceramic components has become a crucial issue for related industries.

[0003] The performance of silicon carbide ceramic components is highly dependent on their final microstructure, which is directly determined by the complex sintering process. This process involves multiple interconnected key process parameters: particle size and gradation of the powder raw materials, type and content of the binder, impregnation temperature and time, heating rate during debinding, uniformity of the sintering temperature field, vacuum level, and holding time. Even minute deviations in any parameter can lead to stress concentration, deformation, cracking, or performance inhomogeneity in large components due to amplification effects, severely impacting the performance indicators of vacuum-sintered ceramics. Therefore, accurately coordinating and optimizing these process parameters to achieve precise control over the ceramic microstructure is a core technical challenge in ensuring the yield and performance consistency of silicon carbide ceramic components.

[0004] However, traditional vacuum sintering process design methods for silicon carbide ceramic components are typically based on empirical knowledge and require multiple trials for correction. This method is not only inefficient and resource-intensive, but while traditional trial-and-error methods may be applicable to the sintering design of small-sized ceramic parts, actual production of ultra-large ceramic components faces extremely high trial-and-error costs and physical limitations. During the vacuum sintering process of ultra-large components, due to severe thermal conduction hysteresis within the component, even small deviations in sintering process parameters are drastically amplified. This amplification effect easily leads to huge temperature gradients and uneven mass transfer between the surface and core of the component, resulting in localized densification differences, uneven shrinkage, and residual stress concentration, ultimately causing deformation, cracking, or extremely uneven performance distribution in large components. Therefore, existing vacuum sintering process design methods for silicon carbide ceramic components mainly suffer from the following problems:

[0005] 1. Data utilization efficiency is low. The setting of process parameters relies heavily on the personal experience and historical data of process personnel. For new components or new formulas, multiple time-consuming and costly iterative experiments are often required, resulting in extremely high trial-and-error costs.

[0006] 2. The phased and isolated optimization strategy usually treats powder preparation, molding, debinding, sintering and other steps as independent processes for optimization, ignoring the parameter coupling effect and linkage relationship of the whole process, making it difficult to achieve global optimization.

[0007] 3. Traditional methods are difficult to accurately predict and compensate for the temperature gradient and stress field changes caused by uneven heat transfer and mass migration during the sintering process of large components, which can easily lead to problems such as deformation, cracking and uneven performance distribution.

[0008] 4. The design cycle is lengthy. Once the size, shape, or material formula of the component changes, the entire process parameter system often needs to be redesigned. It lacks intelligent design capabilities with rapid response and efficient feedback optimization mechanisms.

[0009] Chinese patent CN119811526A, entitled "A Method for Composition and Process Design of Ductile Iron Based on Active Learning Strategy," effectively improves the design efficiency and prediction accuracy of composition and process parameters through data-driven modeling and active learning strategies. However, this method mainly focuses on modeling and optimizing the mapping relationship between composition, process, and performance. Its examples do not involve the real-time acquisition of sintering process parameters, the dynamic response characteristics of specific sintering equipment, or the closed-loop linkage between process setpoints and the actual execution state of the equipment. When facing the complex physical process of vacuum sintering of silicon carbide ceramic components, this method cannot solve the problem of how the ideal process curve adapts to the dynamic coupling of specific equipment, nor can it achieve the integrated design of equipment control and process parameters.

[0010] Chinese patent CN120636629A, entitled "Steel Process Design Method and Related Equipment," establishes a mapping model between process parameters and room temperature tensile properties, achieving data-driven intelligent process recommendation and effectively reducing the reliance on human experience in traditional trial-and-error methods. However, this method primarily focuses on static performance prediction within the existing historical data range, without addressing the proactive collection and supplementation mechanism of key descriptive data for specific complex processes. The database upon which this method relies is essentially a static collection of historical process-performance records, lacking a closed-loop optimization process capable of dynamically assessing the uncertainty of model predictions and proactively guiding new experiments to supplement the most informative data. Summary of the Invention

[0011] To address the shortcomings of existing technologies, this application provides an intelligent integrated design method for the vacuum sintering process of ultra-large ceramic components. This includes:

[0012] Step S1: Obtain the sintering process parameters to be optimized and the sintering aid information to be optimized;

[0013] Step S2: Determine the relative density, relative density weight, band gap optimization target, and band gap weight of silicon carbide ceramics;

[0014] Step S3: Based on the sintering aid information to be optimized, the correlation between the sintering aid information and the component chemical formula of the DFT calculation object, and the second database, the DFT polymerization parameters are obtained. The second database is used to store the DFT parameters.

[0015] Step S4: Transform the sintering process parameters to be optimized, the sintering aid information to be optimized, and the DFT polymerization parameters into input features that the machine learning model can recognize; input the input features into the optimal machine learning model for prediction to obtain the relative density of the combination of sintering aid information and sintering process parameters.

[0016] Step S5: Construct the particle swarm objective function to be minimized based on the relative density of the sintering aid information and the combination of sintering process parameters, the bandgap polymerization value in the DFT polymerization parameters, the relative density of silicon carbide ceramics, the relative density weight, the optimization objective of the bandgap, and the weight of the bandgap.

[0017] Step S6: The sintering aid information to be optimized and the sintering process parameter combination to be optimized are used as particle swarm optimization variables. The particle swarm optimization algorithm is used to optimize the particle swarm objective function to obtain multiple sets of candidate sintering aid information and sintering process parameter combinations.

[0018] Step S7: Using an active learning strategy, evaluate multiple sets of candidate sintering aid information and sintering process parameter combinations to obtain the uncertainty of each candidate sintering aid information and sintering process parameter combination;

[0019] Step S8: Extract the N groups of candidate sintering aid information and the optimal combination of sintering process parameters with the largest uncertainty as the objects of sintering experiment verification. If the target performance after sintering experiment verification reaches the preset value, output the corresponding candidate sintering aid information and sintering process parameter combination as the optimal sintering aid information and sintering process parameter combination. If the target performance after sintering experiment verification does not reach the preset value, save the sintering process parameters obtained after sintering experiment verification and the corresponding ceramic performance parameters after sintering to the first database, save the DFT parameters to the second database, return to step S4, and re-select the optimal machine learning model.

[0020] The optimal machine learning model was obtained through the following selection method:

[0021] Multiple machine learning models targeting different performance targets are trained based on a first database, a second database, and a correlation database. The first database stores historical sintering aid information, historical sintering process parameters, and corresponding post-sintering ceramic performance parameters. The second database stores DFT parameters, including: the chemical formula of the DFT calculation object, the ratio of metal to non-metal atoms, the first ionization energy, the metal atomic radius, the electronegativity ratio, thermodynamic stability, interfacial reactivity, charge transfer, and band gap. The correlation database stores the historical sintering aid information in the first database, the chemical formula of the corresponding DFT calculation object in the second database, and the correlation between other DFT parameters.

[0022] The training of multiple machine learning models with different target performance based on the first database, the second database, and the associated database includes:

[0023] Based on the historical sintering aid information in the first database and the association relationships in the related database, the component chemical formula of the DFT calculation object in the second database is obtained. Based on the component chemical formula of the DFT calculation object, the DFT polymerization parameters are obtained. The DFT polymerization parameters include: band gap polymerization value and thermodynamic stability polymerization value.

[0024] Historical sintering aid information, historical sintering process parameters, and DFT polymerization parameters from the first database are used as a sintering sample and as the input feature matrix for multiple machine learning models. The sintered ceramic performance parameters from the first database are used as the optimization targets for multiple machine learning models. The outputs of multiple machine learning models are used as the model prediction performance. The errors of the model prediction performance and the optimization targets of multiple machine learning models are used as the evaluation criteria for multiple machine learning models. The machine learning model with the smallest error is selected as the optimal machine learning model.

[0025] The process of obtaining DFT polymerization parameters based on the component chemical formulas of the DFT-calculated object includes:

[0026] When a historical sintering aid information corresponds to the component chemical formula of only one DFT calculation object, the DFT parameters are directly used as DFT polymerization parameters.

[0027] When a historical candidate sintering aid information corresponds to the component chemical formula of multiple DFT calculation objects, the DFT parameters corresponding to the component chemical formulas of multiple DFT calculation objects are averaged and polymerized to obtain the polymerization value of the DFT parameters. The polymerization value of the DFT parameters is then used as the final DFT polymerization parameter.

[0028] The bandgap polymerization value and the thermodynamic stability polymerization value are calculated using the following formulas:

[0029] Bandgap aggregation value:

[0030] ;

[0031] in, Let be the bandgap polymerization value in the i-th sintered sample. is the band gap value of the component chemical formula of the k-th DFT calculation object in the i-th sintered sample.

[0032] Thermodynamic stability of polymerization value:

[0033] ;

[0034] in, The thermodynamic stability polymerization value in the i-th sintered sample. is the thermodynamic stability value of the component chemical formula of the k-th DFT calculation object in the i-th sample.

[0035] The objective function of the particle swarm to be minimized is expressed as follows:

[0036] ;

[0037] Where F(x) is the particle swarm optimization objective function to be minimized; The relative density of the combination of sintering aid information and sintering process parameters predicted by the optimal machine learning model; This refers to the bandgap aggregation value in the DFT aggregation parameters; For the target relative density; E g , target W1 represents the target bandgap; W2 represents the relative density weight; W3 represents the bandgap weight.

[0038] The active learning strategy is used to evaluate multiple sets of candidate sintering aid information and sintering process parameter combinations to obtain the uncertainty of each candidate sintering aid information and sintering process parameter combination, including:

[0039] The random forest algorithm is used to predict the information of candidate sintering aids and the combination of sintering process parameters, respectively, and the predicted value of each tree is obtained.

[0040] The average of the squared deviations between the predicted values ​​of each tree and the average predicted value is calculated to obtain the uncertainty of the combination of candidate sintering aid information and sintering process parameters.

[0041] Secondly, this application proposes an electronic device, comprising: one or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the intelligent integrated design method for vacuum sintering process of ultra-large ceramic components.

[0042] Thirdly, this application proposes a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the intelligent integrated design method for vacuum sintering process of ultra-large ceramic components.

[0043] Beneficial technical effects:

[0044] 1. To address the challenge of scarce historical data in the research and development of ultra-large ceramic components, this invention employs a small-sample machine learning algorithm, which can quickly construct a preliminary and reliable composition-process-performance prediction model under limited data conditions, significantly shortening the initial research cycle of process development.

[0045] 2. This invention combines an active learning strategy, which quantifies the uncertainty of model predictions and intelligently selects the process points that are most effective in reducing the uncertainty of model cognition for priority experimental verification.

[0046] 3. This invention transforms the traditional, lengthy, and manual trial-and-error design process into a highly automated, data-driven rapid analysis and decision-making process, greatly improving the efficiency of process design and the speed of response to market demands. More importantly, the system built based on this method also has continuous learning capabilities. New data generated by each experiment is automatically fed back for model iteration and updates, enabling the system performance to continuously evolve with use and realizing the dynamic accumulation of sintering process experience. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 A flowchart of the intelligent integrated design method for vacuum sintering process of silicon carbide ceramic components provided in the embodiments of this application;

[0049] Figure 2 This is a schematic diagram illustrating the evaluation results of various machine learning models provided in the embodiments of the present invention for the density and band gap after sintering.

[0050] Figure 3 A first-principles bandgap structure diagram provided for embodiments of the present invention;

[0051] Figure 4 This is a schematic diagram of the simulation results of the optimal doped ion bandgap and PDOS (Partial Density of States) provided in the embodiments of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0053] This application proposes an intelligent integrated design method for vacuum sintering process of ultra-large ceramic components, involving the construction and evaluation of a silicon carbide ceramic sintering aid database (i.e., the first database), a DFT (Density Functional Theory) parameter database (i.e., the second database), and a machine learning model.

[0054] By establishing a first database, a second database, a correlation library of sintering aids, DFT polymerization parameters, sintering processes, and sintering performance, and the optimal machine learning model obtained through model evaluation and screening, the particle swarm optimization algorithm is used to achieve multi-objective optimization of silicon carbide ceramic sintering aid information and process parameters.

[0055] Example 1

[0056] To address the problems in the background technology, this invention provides an intelligent integrated design method for the vacuum sintering process of ultra-large ceramic components, such as... Figure 1 As shown, it includes the following steps:

[0057] Step S1: Obtain the sintering process parameters to be optimized and the sintering aid information to be optimized;

[0058] In this embodiment, the ultra-large ceramic component mainly refers to a large-scale complex structural component that exhibits a significant size effect during vacuum sintering due to its feature size exceeding conventional limitations. Specifically, it typically refers to silicon carbide ceramic components with a feature size of 0.5 meters or more, or a maximum feature thickness exceeding 30 millimeters. It is understood that the scope of application of this embodiment is particularly aimed at, but not limited to, the research and development and process optimization stage of the aforementioned ultra-large ceramic components. The method of this embodiment can also be used for non-ultra-large ceramic components, but the effect is more significant for ultra-large ceramic components.

[0059] The sintering process parameters to be optimized include: holding time, average sintering temperature, minimum sintering temperature, maximum sintering temperature, average heating rate, sintering pressure, and the logarithm of vacuum degree.

[0060] The sintering aid information to be optimized includes: oxide liquid phase sintering aids, boron-carbon sintering aids, second phase reinforcing aids, resins or carbon source precursors, and organic binders or forming aids.

[0061] Step S2: Determine the relative density, relative density weight, band gap optimization target, and band gap weight of silicon carbide ceramics;

[0062] In this embodiment, the target relative density of the silicon carbide ceramic is set to 96%TD, with a relative density weight of 0.6; the target band gap is 2.572 eV, with a band gap weight of 0.4. The weighting reflects that this embodiment prioritizes ensuring the sintered ceramic material achieves a densification level close to 96%TD while ensuring the band gap is as close as possible to the target value.

[0063] Step S3: Based on the sintering aid information to be optimized, the correlation between the sintering aid information and the component chemical formula of the DFT calculation object, and the second database, the DFT polymerization parameters are obtained. The second database is used to store the DFT parameters.

[0064] In this embodiment, when a sintering aid information to be optimized corresponds to the chemical formula of a component of a single DFT calculation object, the DFT parameter of that component is directly used as the DFT polymerization parameter. When a sintering aid information to be optimized corresponds to the chemical formula of multiple DFT calculation objects, the DFT parameters of multiple components are averaged and polymerized to obtain the DFT polymerization parameter. For example, the Al2O3-Y2O3 system can be obtained by polymerizing the DFT parameters of two components, Al2O3 and Y2O3; the B4C-C system can be obtained by polymerizing the DFT parameters of B4C and C; and the Al2O3-Y2O3-TiO2 system can be obtained by polymerizing the DFT parameters of three components, Al2O3, Y2O3, and TiO2.

[0065] Step S4: Transform the sintering process parameters to be optimized, the sintering aid information to be optimized, and the DFT polymerization parameters into input features that the machine learning model can recognize; input the input features into the optimal machine learning model for prediction to obtain the relative density of the combination of sintering aid information and sintering process parameters.

[0066] In this embodiment, the input features are fed into the optimal machine learning model for prediction to obtain the relative density of the combination of sintering aid information and sintering process parameters, as expressed below:

[0067] ;

[0068] Where, ρ pred The relative density, expressed as %TD, represents the combination of sintering aid information and sintering process parameters; ENRX represents the optimal machine learning model; X represents the sintering process parameters to be optimized, the sintering aid information to be optimized, and the DFT polymerization parameters, which are transformed into input features that the machine learning model can recognize.

[0069] Machine learning models balance feature selection capability and model stability by simultaneously introducing L1 and L2 regularization terms into the loss function. Its general form is:

[0070] ;

[0071] Where Y is the true relative density, X is the sintering process parameters to be optimized, the sintering aid information to be optimized, and the DFT polymerization parameters converted into input features that the machine learning model can recognize, β is the coefficient of the machine learning model, λ is the regularization strength, and α is the ratio of L1 to L2 regularization. It is a first-order norm. It is a second-order norm.

[0072] The purpose of this machine learning model is to quickly predict the potential relative density level of candidate sintering aids based on information and process parameters, thereby avoiding conducting sintering experiments on each candidate combination individually. In this embodiment, the optimal machine learning model needs to be selected first, and the selection process is as follows:

[0073] Multiple machine learning models targeting different performance targets are trained based on a first database, a second database, and a correlation database. The first database stores historical sintering aid information, historical sintering process parameters, and corresponding post-sintering ceramic performance parameters. The second database stores DFT parameters, including: the chemical formula of the DFT calculation object, the ratio of metal to non-metal atoms, the first ionization energy, the metal atomic radius, the electronegativity ratio, thermodynamic stability, interfacial reactivity, charge transfer, and band gap. The correlation database stores the historical sintering aid information in the first database, the chemical formula of the corresponding DFT calculation object in the second database, and the correlation between other DFT parameters.

[0074] The training of multiple machine learning models with different target performance based on the first database, the second database, and the associated database includes:

[0075] Based on the historical sintering aid information in the first database and the association relationships in the related database, the component chemical formula of the DFT calculation object in the second database is obtained. Based on the component chemical formula of the DFT calculation object, the DFT polymerization parameters are obtained. The DFT polymerization parameters include: the polymerization value of the DFT parameter, the band gap polymerization value, and the thermodynamic stability polymerization value.

[0076] Historical sintering aid information, historical sintering process parameters, and DFT polymerization parameters from the first database are used as a sintering sample and as the input feature matrix for multiple machine learning models. The sintered ceramic performance parameters from the first database are used as the optimization targets for multiple machine learning models. The outputs of multiple machine learning models are used as the model prediction performance. The errors of the model prediction performance and the optimization targets of multiple machine learning models are used as the evaluation criteria for multiple machine learning models. The machine learning model with the smallest error is selected as the optimal machine learning model.

[0077] The process of obtaining DFT polymerization parameters based on the component chemical formulas of the DFT-calculated object includes:

[0078] When a historical sintering aid information corresponds to the component chemical formula of only one DFT calculation object, the DFT parameters are directly used as DFT polymerization parameters.

[0079] When a historical candidate sintering aid information corresponds to the component chemical formula of multiple DFT calculation objects, the DFT parameters corresponding to the component chemical formulas of multiple DFT calculation objects are averaged and polymerized to obtain the polymerization value of the DFT parameters. The polymerization value of the DFT parameters is then used as the final DFT polymerization parameter.

[0080] Specifically, in the implementation process, sintering process parameters, additive category coding parameters, and micro-polymerization parameters are set as inputs, and relative density and band gap performance parameters are set as outputs. Regression prediction models are constructed using small-sample machine learning algorithms, and the optimal machine learning model is determined.

[0081] Step S100: Set multiple performance target values ​​and their corresponding weights;

[0082] The target relative density of silicon carbide ceramics is set at 96%TD, with a relative density weight of 0.6; the target band gap is 2.572 eV, with a band gap weight of 0.4. This weighting reflects that, while ensuring the band gap is as close as possible to the target value, this embodiment prioritizes ensuring that the sintered ceramic material achieves a densification level close to 96%TD.

[0083] The relative density of the combination of sintering aid information and sintering process parameters is predicted and calculated using the optimal machine learning model, and the band gap is directly extracted from the micro parameters in the second database.

[0084] Step S101: Select the machine learning model to be optimized;

[0085] The selection of the optimal machine learning model is specifically implemented by treating experimental and literature data on silicon carbide ceramic sintering aids as historical data for the sintering of silicon carbide ceramic components. The specific implementation process includes the following steps:

[0086] Step S101.1: Collect experimental and literature data on silicon carbide ceramic sintering aids.

[0087] Historical sintering process parameters include holding time, average sintering temperature, heating rate, sintering pressure, and vacuum level. Corresponding post-sintering ceramic performance parameters include relative density, hardness, fracture toughness, and average grain size. Samples with complete parameters are entered into the first database; samples with missing parameters retain their original information as data to be supplemented.

[0088] Data on silicon carbide ceramic sintering systems were collected from literature, experimental data, and publicly available process records. The collected data should include information on sintering aids, sintering process parameters, and post-sintering ceramic performance parameters. Samples with high data completeness were entered into the first database as valid samples; for samples lacking certain performance parameters or DFT parameters, their original sintering aid and process information were retained for subsequent supplementation through DFT calculations or machine learning predictions.

[0089] The first database contains historical information on sintering aids, which mainly includes oxide liquid-phase sintering aids, boron-carbon sintering aids, second-phase reinforcing aids, resins or carbon source precursors, and organic binders or forming aids. Among them, oxide liquid-phase sintering aids include Al2O3 (aluminum oxide), Y2O3 (yttrium oxide), Al2O3-Y2O3, and Al2O3-Y2O3-TiO2 (titanium dioxide), whose main function is to promote liquid phase formation and mass transfer processes, thereby improving the degree of sintering densification; boron-carbon sintering aids include B4C (boron carbide), C (carbon), and B4C-C composite systems, whose main function is to remove the oxide film on the surface of SiC (silicon carbide) powder and promote solid-phase sintering; second-phase reinforcing aids include TiB2 (titanium diboride), TiC (titanium carbide), ZrB2 (zirconium diboride), and WC (tungsten carbide), whose main function is to improve the hardness, fracture toughness, or high-temperature stability of ceramic materials; resins or carbon source precursors include phenolic resins, polycarbosilanes, etc., which can produce residual carbon or ceramic products during sintering; organic binders or forming aids include PVA (Polyvinyl chloride). Alcohol (polyvinyl alcohol), PVB (polyvinyl butyral), PEG (polyethylene glycol), and glycerin, etc., mainly affect the forming, dispersion, impregnation, and structural uniformity of the preform before sintering. The process parameter data is shown in Table 1 of the first database.

[0090] Table 1. Data table of sintering process parameters;

[0091] time min total Insulation time or process time min temp means C Average sintering temperature ℃ temp min C Minimum sintering temperature ℃ temp max C Maximum sintering temperature ℃ heating rate C per min mean Average heating rate ℃ / min pressure MPa Sintering pressure MPa log10 vacuum Pa Logarithmic value of vacuum degree log10(Pa)

[0092] The performance parameter data is shown in Table 2 of the first database.

[0093] Table 2: Data table of ceramic performance parameters after sintering;

[0094] density g cm3 Density after sintering g / cm³ relative density pct TD relative density %TD hardness GPa hardness GPa grain size um Average grain size μm toughness MPam05 fracture toughness MPa·m¹ / ² strength MPa flexural strength MPa resistivity Ohm cm resistivity Ω·cm

[0095] Step S102: Construct a second database describing the intrinsic characteristics and electronic structure characteristics of sintering aids.

[0096] The physical quantities of single sintering aid components and composite aid components were obtained through first-principles calculations. Calculation parameters included the ratio of metal to non-metal atoms, first ionization energy, metal atomic radius, electronegativity ratio, thermodynamic stability, interfacial reactivity, charge transfer amount, and band gap.

[0097] Based on the first database, a second database is further constructed. The second database is used to describe the intrinsic elemental characteristics, electronic structure characteristics, interfacial reaction characteristics, and thermodynamic stability characteristics of sintering aids or related components.

[0098] The purpose of establishing the second database is to introduce microscopic electronic structure information, which is difficult to directly characterize in traditional sintering experimental data, into the machine learning model, so that the model can not only learn the statistical relationship between "process parameters and performance", but also combine the physical relationship between "additive chemical composition, electronic structure and interface behavior".

[0099] The DFT calculation objects in the second database include single sintering aid components and composite aid components. For example, for Al2O3-Y2O3 composite aids, DFT polymerization parameter records for Al2O3 and Y2O3 can be established separately; for the B4C-C system, DFT polymerization parameter records for B4C and C can be established separately; for second-phase reinforcing aids such as TiB2, TiC, ZrB2, and WC, DFT polymerization parameter records for the corresponding compounds can be established separately; for organic or carbon source aids such as phenolic resins, PVA, PEG, and glycerol, approximate DFT matching can be established based on the carbon source, Si-C structure, or related residual components that may be formed after sintering.

[0100] The fields in the second database are shown in Table 3.

[0101] Table 3: Data tables in the second database;

[0102] Formula Chemical formula of the component of the DFT calculation object Primary Metal Major metallic elements Primary Non Metal Major nonmetallic elements Num Metal Atoms Number of metal atoms M NM Atomic Ratio Metal / Nonmetal Atomic Ratio First Ionization Energy eV First ionization energy Metal Radius Angstrom Metal atomic radius Metal Valence Electrons Metal valence electron count Metal Electronegativity Electronegativity of metals M NM Electronegativity Ratio Electronegativity ratio of metals / nonmetals Metal Period Periodic number of metallic elements Metal Group Metal group numbers Metal Block Element Block Metal Block Code Element block encoding Y1 Thermo Stability eV Thermodynamic stability Y2 Interface Reactivity eV Interfacial reactivity Y3 ChargeTransfer e Charge transfer amount Y4 BandGap eV band gap

[0103] Among them, Y1 Thermo Stability eV is used to characterize the thermodynamic stability of the additive or additive-matrix related structure at the sintering temperature; Y2 Interface Reactivity eV is used to characterize the tendency of interfacial reaction between the sintering additive and the SiC matrix; Y3 ChargeTransfer eV is used to describe the charge transfer behavior between the additive and the matrix interface; and Y4 BandGap eV is used to describe the electronic band gap of the additive information or additive-matrix related structure. The above DFT parameters can be obtained through first-principles calculations, existing material databases, publicly available DFT results, or subsequent simulation calculations. In this embodiment, the second database includes 75 DFT component data points. This database can be further supplemented based on machine learning feature weights and particle swarm optimization results.

[0104] Step S103: Based on the historical sintering aid information in the first database and the association relationship in the association library, obtain the component chemical formula of the DFT calculation object in the second database, and obtain the DFT polymerization parameters based on the component chemical formula of the DFT calculation object.

[0105] In this embodiment, multiple microscopic component parameters corresponding to the same sample are aggregated by arithmetic mean or mass fraction weighted average to form a sample-level descriptor. The aggregated microscopic parameters are then combined with process parameters and performance parameters to construct a dataset relating components to processes and microscopic parameters to performance.

[0106] After obtaining the first and second databases, the two databases are correlated and matched. The core of this step is to convert the sintering aid names in the first database into standardized chemical components that can be matched with the Formula field in the second database. One or more DFT components corresponding to each sintering sample are expanded to form expanded correlation data of "sample—sintering aid—DFT component—DFT polymerization parameter". If a sample corresponds to multiple DFT components (i.e., the chemical formula of the component being calculated), then that sample corresponds to multiple records in the expanded table. The purpose of the expanded table is to retain the independent DFT information of each aid component, facilitating subsequent analysis of the contribution of different components to performance. Multiple DFT components matched to the same sample are aggregated to form a sample-level DFT descriptor. The preferred aggregation method is arithmetic mean, but weighted average based on aid content, molar ratio, or mass fraction can also be used. When the precise aid content is not given in the original data, the arithmetic mean is used as the default aggregation method.

[0107] Aggregated values ​​of DFT parameters:

[0108] ;

[0109] in, This represents the aggregation value of the j-th DFT parameter in the i-th sintered sample; n i This represents the number of component chemical formulas of the DFT calculation object matched by the i-th sintered sample; This represents the j-th DFT parameter of the component chemical formula of the k-th DFT calculation object in the i-th sintered sample.

[0110] The band gap aggregation value is:

[0111] ;

[0112] in, Let be the bandgap polymerization value in the i-th sintered sample. is the band gap value of the component chemical formula of the k-th DFT calculation object in the i-th sintered sample.

[0113] Thermodynamic stability of the polymer is:

[0114] ;

[0115] in, The thermodynamic stability polymerization value in the i-th sintered sample. is the thermodynamic stability value of the component chemical formula of the k-th DFT calculation object in the i-th sample.

[0116] It is understood that the calculation of DFT polymerization parameters in step S3, including the bandgap polymerization value and the thermodynamic stability polymerization value, is consistent with the above calculation formula, and will not be repeated in this embodiment.

[0117] Through the above processing, a correlation library of sintering aids, DFT polymerization parameters, sintering processes, and sintering performance is finally obtained. In this embodiment, a correlation record of sample-sintering aid-DFT polymerization parameters is established using valid sintering samples from the first database and DFT component data from the second database.

[0118] Step S104: Perform data cleaning and normalization on the sintered samples. The normalization formula is:

[0119] ;

[0120] Where X represents the original data (i.e., a specific data point in the sintered sample), X min With X max These represent the minimum and maximum values ​​of a specific data item in the sintering sample. The category features of the sintering aids are subjected to one-thermal encoding to generate the input feature matrix and output target matrix for the machine learning model.

[0121] The sintered samples underwent data cleaning, missing value handling, encoding, and normalization. Data cleaning included deleting duplicate records, deleting obviously invalid records, correcting records with inconsistent units, identifying outliers, and performing necessary corrections. During encoding, the sintering aid categories were converted into category features recognizable by the machine learning model. Specifically, each sample was coded as either 0 or 1 for whether it contained oxide liquid-phase sintering aids, boron-carbon sintering aids, carbon sources or resin precursors, second-phase reinforcing aids, and organic binders. In this way, the model can identify both specific DFT polymerization parameters and the functional categories of the aids.

[0122] Step S105: Evaluate model performance using the 5-fold cross-validation method.

[0123] A 5-fold cross-validation method was employed to evaluate and optimize the constructed screening model, effectively preventing overfitting. First, the complete sample dataset was randomly and uniformly divided into five non-overlapping subsets. Then, in each iteration, one subset was selected as the validation set, and the remaining four subsets were combined as the training set. After all five iterations were completed, the arithmetic mean of the five performance metrics was calculated, which served as the final evaluation result measuring the model's overall prediction accuracy and generalization ability.

[0124] Specifically, the following steps are included:

[0125] Step S105.1: Use the sintering process parameters, additive category coding parameters, and DFT polymerization parameters as input features for the machine learning model. The input feature matrix is ​​represented as follows:

[0126] ;

[0127] Among them, X i P represents the input feature matrix of the i-th sample; i This indicates the sintering process parameters for this sample; A i This indicates the sintering aid category parameter for this sample; D i This represents the DFT aggregation parameters of the sample.

[0128] Sintering process parameters P i This includes holding time, average sintering temperature, minimum sintering temperature, maximum sintering temperature, average heating rate, sintering pressure, and logarithm of vacuum degree. Additive category parameter A. i This includes whether it contains oxide liquid-phase sintering aids, boron-carbon additives, carbon sources, second-phase reinforcing agents, and organic binders. DFT polymerization parameters D iThis includes the number of DFT-matched components, the number of metal atoms, the metal / non-metal atom ratio, the first ionization energy, the metal atomic radius, the number of metal valence electrons, the metal electronegativity, the metal / non-metal electronegativity ratio, the metal period number, the metal group number, the element block code, thermodynamic stability, interfacial reactivity, charge transfer amount, and band gap.

[0129] The model output Y consists of the ceramic performance parameters after sintering, including density (g / cm³), relative density (pct / TD), hardness (GPa), and grain size (µm). Because the data completeness varies for different performance parameters, a separate prediction model is built for each target performance. That is, the density model is trained using only samples with valid density data, the relative density model uses only samples with valid relative density data, the hardness model uses only samples with valid hardness data, and the grain size model uses only samples with valid grain size data.

[0130] Step S105.2: Based on the input feature matrix and target performance data, establish various machine learning regression models. The models used include OLS (Ordinary Least Squares), LASSO (Least Absolute Shrinkage and Selection Operator), Ridge (Ridge Regression), ENR (Elastic Net Regression), SVR (Support Vector Regression), RF (Random Forest Regression), GPR (Gaussian Process Regression), and ANN (Artificial Neural Network).

[0131] The model prediction relationship is represented as follows:

[0132] ;

[0133] Among them, Y pred X represents the model's predictive performance; X is the input feature matrix composed of sintering process parameters, additive category parameters, and DFT polymerization parameters; f is the corresponding machine learning model.

[0134] For each target performance, the data is divided into training, validation, and test sets, and Trainmse, Valmse, Testmse, and R² are calculated. Trainmse characterizes the model's fitting error to the training data, Valmse assists in model selection and hyperparameter tuning, Testmse evaluates the model's predictive ability on unseen samples, and R² measures the model's overall interpretability.

[0135] Mean squared error (MSE) is defined as:

[0136] ;

[0137] MSE is the mean squared error of the model's predictive performance and the optimization objective of various machine learning models.

[0138] The coefficient of determination R² is defined as:

[0139] ;

[0140] Where R² is the coefficient of determination for the model's predictive performance and the optimization objectives of various machine learning models, y i These are the actual values ​​(i.e., the sintered ceramic performance parameters in the first database). The predicted value (i.e., the output of multiple machine learning models). is the average of the true values, and n is the number of samples.

[0141] In addition, to evaluate the correlation between the model's predicted values ​​and the true values, the Pearson correlation coefficients for the training and test sets were calculated separately:

[0142] ;

[0143] in, is the average of the predicted values; n is the sample size; r is the Pearson correlation coefficient.

[0144] Step S105.3: Compare the predictive performance of different machine learning models;

[0145] Based on the association database constructed from the first and second databases, OLS, Lasso, Ridge, ENR, SVR, RF, GPR and ANN models were trained for the four target performances: density g cm3, relative density pct TD, hardness GPa and grain size um, respectively, and the model evaluation results are shown in Table 4.

[0146] Table 4 Evaluation results of the density g cm3 model;

[0147] OLS 0.0118 0.0326 0.0937 0.7728 Lasso 7.68E-10 0.00498 0.20068 0.67907 Ridge 0.11250 0.02162 0.14126 0.35540 ENR 0.22469 0.02669 0.14942 -0.05970 SVR 0.20103 0.00608 0.08785 0.14480 RF 0.10902 0.01876 0.10777 0.42360 GPR 2.24E-17 0.02603 0.14632 0.73965 ANN 0.00036 0.01307 0.08757 0.84587

[0148] Table 5 Evaluation results of the bandgap eV model;

[0149] OLS 0.000000 1.336269 8.00E-05 0.996102 Lasso 0.091042 1.003788 0.2494 -11.1111 Ridge 0.847723 1.179052 0.785787 -37.1586 ENR 1.026502 1.035241 0.78581 -37.1598 SVR 0.009553 0.875785 0.153631 -6.46047 RF 0.104446 0.264945 0.135835 -5.59631 GPR 0.000000 1.297824 0.165327 -7.02847 ANN 0.000000 0.6769 0.047584 -1.31073

[0150] Step S105.4: Based on the Train mse, Val mse, Test mse and R² evaluation results of the above models, determine the optimal machine learning model corresponding to different target performance.

[0151] In this embodiment, this step completes the construction and evaluation of the first database, the second database, the sintering aid-DFT polymerization parameters-sintering process-sintering performance association library, and various machine learning performance prediction models, providing a data foundation and model foundation for subsequent reverse design of sintering aid information and process parameters using particle swarm optimization algorithm.

[0152] Based on the evaluation results of the different machine learning models mentioned above, for relative density (pctTD), the ENR model has better overall performance under the existing small sample data conditions (e.g., Figure 2 (As shown). The ENR model possesses both L1 and L2 regularization capabilities, enabling it to maintain model stability even with a small sample size, a large number of input features, and some collinearity. Therefore, this embodiment selects the ENR model as the prediction surrogate model for relative density.

[0153] For the bandgap, this embodiment does not use a machine learning model for prediction, but directly uses the DFT calculation parameter Y4 BandGap eV from the second database as the bandgap descriptor. The reason for this setting is that the bandgap is an electronic structure parameter, and directly derived from the DFT calculation results has more physical meaning; at the same time, the bandgap parameters in the second database can be continuously updated as subsequent DFT simulation results are added, thereby improving the reliability of the optimization results.

[0154] Therefore, in this embodiment, the relative density is predicted by the selected ENR model, and the band gap is given by the DFT aggregation parameters in the second database. For each set of candidate sintering aid information and process combination, its machine learning input features are first constructed, and then the relative density is predicted by the output of the ENR model. At the same time, the band gap of the corresponding sintering aid information is read or aggregated from the second database.

[0155] Step S5: Construct the particle swarm objective function to be minimized based on the relative density of the sintering aid information and the combination of sintering process parameters, the bandgap polymerization value in the DFT polymerization parameters, the relative density of silicon carbide ceramics, the relative density weight, the optimization objective of the bandgap, and the weight of the bandgap.

[0156] In this embodiment, a particle swarm optimization algorithm is used to iteratively search for sintering aid information and process parameters. The sintering aid information and process parameters are obtained by solving the optimization objective of the silicon carbide ceramic component sintering process using a selected machine learning model.

[0157] Step S5.1: First, construct the particle swarm objective function to be minimized based on the weighted error:

[0158] Step S5.1.1: Determine the particle swarm optimization variables;

[0159] In the particle swarm optimization algorithm, each particle represents a set of sintering aid information and sintering process parameters to be optimized. This candidate combination contains both compositional and process information.

[0160] Information on sintering aids to be optimized includes TiB2, TiC, ZrB2, WC, Al2O3, Al2O3-Y2O3, B4C-C, Si-C, Al2O3-Y2O3-TiO2, and carbon source information related to phenolic resins. The corresponding sintering process or performance records for these candidate sintering aids can be found in the first database, or the corresponding DFT composition parameters can be found in the second database.

[0161] The sintering process parameters to be optimized include: holding time, average sintering temperature, minimum sintering temperature, maximum sintering temperature, average heating rate, sintering pressure, and logarithmic values ​​of vacuum. The ranges of continuous variables such as holding time, average sintering temperature, average heating rate, sintering pressure, and vacuum are not arbitrarily set, but determined based on the actual distribution of existing experimental or literature data in the first database. This avoids the particle swarm optimization algorithm finding combinations that significantly exceed the existing data range, thereby reducing the uncertainty brought about by extrapolation in machine learning models. For cases where pressure or vacuum is missing in certain data records, it can be reasonably supplemented based on similar sintering methods, similar equipment conditions, or the statistical range of existing data. However, this supplementation is only used to construct calculable optimization inputs and does not change the original experimental records themselves.

[0162] The sintering aid information and sintering process parameters to be optimized are used as variables in the particle swarm optimization (PSO) algorithm. The parameter ranges for these parameters are determined based on the actual distribution of existing experimental or literature data in the first database. For each candidate combination generated by the PSO algorithm, it is transformed into input features that the machine learning model can recognize. These input features include three parts: sintering process parameters, sintering aid category features, and DFT polymerization features. If a candidate combination only provides a single sintering temperature, this temperature is simultaneously used as the average sintering temperature, minimum sintering temperature, and maximum sintering temperature in the model input.

[0163] The sintering aid category feature describes whether the candidate sintering aid information contains oxide liquid-phase sintering aids, boron-carbon sintering aids, carbon sources or resin precursors, second-phase reinforcing aids, and organic binders. This category feature does not directly replace DFT polymerization parameters, but rather assists the model in identifying the functional differences of different types of aids in the sintering densification process.

[0164] When a candidate sintering aid corresponds to the chemical formula of only one component in a DFT calculation object, the DFT parameter of that component is directly used as the DFT polymerization parameter. When a candidate sintering aid corresponds to the chemical formula of multiple components in a DFT calculation object, the DFT parameters of multiple components are averaged and polymerized to obtain the DFT polymerization parameter. For example, the Al2O3-Y2O3 system can be obtained by polymerizing the DFT parameters of two components, Al2O3 and Y2O3; the B4C-C system can be obtained by polymerizing the DFT parameters of B4C and C; and the Al2O3-Y2O3-TiO2 system can be obtained by polymerizing the DFT parameters of three components, Al2O3, Y2O3, and TiO2.

[0165] The aggregation formula is:

[0166] ;

[0167] Wherein, DFTmean j (i) represents the aggregated value of the j-th DFT parameter in the i-th candidate combination, n i This indicates the number of component chemical formulas in the DFT calculation objects matched by the candidate combination. j (i,k) represents the j-th DFT parameter of the component chemical formula of the k-th DFT calculation object in the i-th candidate combination.

[0168] For each group of candidate sintering aids, its band gap is determined by the Y4 BandGap eV field in the second database.

[0169] When the candidate sintering aid information is a single component, such as TiB2, TiC, ZrB2, or WC, the DFT band gap value of that component in the second database is directly used. When the candidate sintering aid information is a composite component, such as Al2O3-Y2O3, B4C-C, or Al2O3-Y2O3-TiO2, the band gaps of each matching component are averaged and polymerized to serve as the band gap descriptor for the candidate sintering aid information.

[0170] The bandgap convergence formula is:

[0171] ;

[0172] in, n represents the bandgap aggregation value in the DFT aggregation parameters. A E represents the number of component chemical formulas of the DFT calculation object corresponding to the sintering aid information. g (A,k) represents the band gap of the component chemical formula of the k-th DFT calculation object in the sintering aid information.

[0173] If reliable DFT bandgap data is not yet available for a candidate sintering aid, this information can still be retained as a candidate in the current optimization, but the uncertainty of its bandgap objective term is high. For such candidate information, DFT simulation results should be prioritized in subsequent steps to improve the reliability of the optimization results.

[0174] Step S 5.1.2, construct the particle swarm optimization objective function:

[0175] In this embodiment, the relative density error and bandgap error are normalized and then summed according to their weights to obtain the objective function of the particle swarm optimization algorithm.

[0176] ;

[0177] Where F(x) is the particle swarm optimization objective function to be minimized; The relative density of the combination of sintering aid information and sintering process parameters predicted by the optimal machine learning model; This refers to the bandgap aggregation value in the DFT aggregation parameters; For the target relative density; E g , target W1 represents the target bandgap; W2 represents the relative density weight; W3 represents the bandgap weight.

[0178] Specifically, in this embodiment, since the target relative density is 96, 2.572 is the target bandgap, 0.6 is the relative density weight, and 0.4 is the bandgap weight, the objective function is:

[0179] ;

[0180] Where F(x) is the objective function to be minimized; The relative density is predicted by the ENR model; This refers to the bandgap aggregation value in the DFT aggregation parameters.

[0181] Since relative density and band gap have different units and orders of magnitude, this embodiment uses relative error form for normalization, so that the two targets can be compared in a weighted manner within the same objective function.

[0182] Step S6: The sintering aid information and the sintering process parameter combination to be optimized are used as particle swarm optimization variables. The particle swarm optimization algorithm is used to optimize the particle swarm objective function to obtain multiple sets of candidate sintering aid information and sintering process parameter combinations.

[0183] In this embodiment, the particle swarm optimization algorithm is used to search for the optimal combination. After the objective function is constructed, the particle swarm optimization algorithm is used to iteratively search for sintering aid information and process parameters. The core idea of ​​the particle swarm optimization algorithm is that each particle represents a candidate sintering combination, and each particle continuously updates its position based on its own historical best result and the group's historical best result during the search process, thereby gradually moving towards a region with a smaller objective function.

[0184] The particle velocity update formula is:

[0185] ;

[0186] The particle position update formula is:

[0187] ;

[0188] Among them, v i (k) represents the velocity of the i-th particle in the k-th iteration. To represent the velocity of the i-th particle in the (k+1)-th iteration, Let represent the position of the i-th particle in the k-th iteration, ω represent the inertia weight, c1 represent the individual learning factor, c2 represent the group learning factor, r1 represent the first random number between 0 and 1, and r2 represent the second random number between 0 and 1. Let g represent the historical best position of the i-th particle. best This represents the group's historical best position among all particles.

[0189] In this embodiment, the particle swarm optimization process is implemented using the Python programming language and runs in the PyCharm integrated development environment. The program is derived from the data preprocessing, model training, and model evaluation workflow established in Embodiment 1, with the addition of a particle swarm optimization module. The goal of this program is not to replace experiments, but rather to quickly search for candidate sintering aids and process combinations that best approximate the target performance based on existing databases and machine learning models, providing a priority ranking for subsequent DFT supplementary calculations and experimental verification.

[0190] During the iteration process, for each candidate combination generated, the program automatically performs the following operations: converts the candidate combination into model input features; calls the ENR model to predict the relative density; reads or aggregates bandgap parameters from the second database; calculates the weighted objective function; compares the current result with the historical best result; and updates the particle position and velocity. After multiple iterations, the algorithm outputs the candidate combination with the smallest objective function.

[0191] Through the above particle swarm optimization process, under the current database and model conditions, we obtained 8 sets of sintering aid information and process combinations that minimize the objective function.

[0192] To further improve the accuracy and interpretability of the optimization results, it is necessary to determine the supplementary direction of subsequent DFT simulations based on the model feature weights, the composition of the objective function, and the current optimization results.

[0193] Since the optimization objective in this embodiment directly includes the band gap, Y4 BandGap eV is the most prioritized DFT parameter to supplement. For candidate sintering aids with a small objective function but uncertain band gap data, band gap calculation or recalculation should be performed first to confirm the reliability of their electronic structure parameters.

[0194] In addition, intrinsic characteristics of elements such as electronegativity, metal / nonmetal electronegativity ratio, atomic radius, first ionization energy, group number, and element block code also need to be retained in the second database. Although these parameters may not be directly used as part of the objective function, they can help machine learning models identify periodic patterns and differences in chemical interactions among different auxiliary elements.

[0195] Regarding information on candidate sintering aids, DFT calculation results for TiB2 / SiC were prioritized for supplementation, as TiB2 represents the optimal information obtained from current PSO optimization. Secondly, DFT results for Si-C / SiC or carbon source-related information were supplemented, as the objective function value for this information is second only to TiB2. Subsequent supplementation could include TiC-Si-Al / SiC, etc. , and The DFT results, including information such as [list of information], should be re-entered into the second database and used to update the sintering aid-DFT parameter-process-performance correlation library.

[0196] Step S7: Employing an active learning strategy, evaluate multiple sets of candidate sintering aid information and sintering process parameter combinations to obtain the uncertainty of each candidate sintering aid information and sintering process parameter combination, including:

[0197] Step S7.1: Use the random forest algorithm to predict the information of candidate sintering aids and the combination of sintering process parameters respectively, and obtain the predicted value of each tree;

[0198] Step S7.2: Calculate the average of the squared deviations between the predicted value and the average predicted value of each tree to obtain the uncertainty of the combination of candidate sintering aid information and sintering process parameters.

[0199] In one optional implementation, the method further includes selecting sintering aid information and process combinations that need to be experimentally verified based on an active learning strategy.

[0200] After obtaining information on eight candidate sintering aids and process combinations, instead of directly conducting experiments on all combinations, an active learning strategy was used to determine the priority verification targets. The purpose of active learning is to reduce the number of experiments and prioritize limited experimental resources for the candidate combinations that are most valuable for improving the model. Furthermore, the final model with high accuracy optimized based on the active learning strategy was used to further optimize the composition-process that meets the performance targets.

[0201] Based on the target relative density of 96%TD and the target band gap of 2.572eV, the composition and smelting process parameters of ductile iron that meet the target performance were designed and verified by experiments. The experiments showed that the designed SiC ceramic had a target relative density of 96%TD and a target band gap of 2.572eV.

[0202] Specifically, the uncertainty of candidate combinations is evaluated based on an active learning strategy, and the information of the three candidate sintering aids with the highest uncertainty is extracted as priority verification objects.

[0203] While keeping parameters such as maximum depth constant, the number of trees in the extreme random tree algorithm was changed sequentially to 100, 140, 180, 220, and 260, and the performance prediction values ​​of 8 candidate combinations were calculated for each. The formula is:

[0204] ;

[0205] Where M is the number of trees in the random forest, which is the number of models in the ensemble; m is the index of the m-th tree, m = 1, 2, …, M; The m-th tree represents the predicted value of the candidate sintering aid information and sintering process parameter combination; This is the average predicted value for all M trees.

[0206] Calculate the mean square error of the predicted values ​​of different models, and obtain the average of the squared deviations between the predicted values ​​of each tree and the average predicted value, i.e., the uncertainty U.

[0207] Step S8: Extract the N groups of candidate sintering aid information and the optimal combination of sintering process parameters with the largest uncertainty as the objects of sintering experiment verification. If the target performance after sintering experiment verification reaches the preset value, output the corresponding candidate sintering aid information and sintering process parameter combination as the optimal sintering aid information and sintering process parameter combination. If the target performance after sintering experiment verification does not reach the preset value, save the sintering process parameters obtained after sintering experiment verification and the corresponding ceramic performance parameters after sintering to the first database, save the DFT parameters to the second database, return to step S4, and re-select the optimal machine learning model.

[0208] In this embodiment, the information of the three candidate sintering aids with the highest uncertainty and the optimal combination of sintering process parameters are extracted as the priority verification targets. First-principles calculation parameters for band gap and density of states (such as...) are used to supplement the missing sintering aid information. Figure 3 , Figure 4 As shown in the figure, corresponding sintering experiments were conducted to verify the results. The composition, process, microscopic parameters, and performance data obtained from the experiments were added to the original dataset, and step S4 was re-executed to update and correct the machine learning model. Finally, the iterative output of silicon carbide ceramic sintering aid information and process parameters that meet the target performance was obtained.

[0209] Example 2

[0210] This embodiment proposes an electronic device, including: one or more processors, and a memory, the memory being used to store instructions, which, when executed by the one or more processors, cause the one or more processors to perform the instructions.

[0211] The electronic device can be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements an intelligent integrated design method for vacuum sintering of ultra-large ceramic components as described in the embodiments. It is understood that the electronic device may also include input / output (I / O) interfaces and communication components.

[0212] The processor is used to execute all or part of the steps in the intelligent integrated design method for vacuum sintering of ultra-large ceramic components as described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in the electronic device, as well as application-related data.

[0213] The processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the intelligent integrated design method for vacuum sintering process of ultra-large ceramic components described in the above embodiments.

[0214] Example 3

[0215] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0216] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the intelligent integrated design method for vacuum sintering process of ultra-large ceramic components described in various embodiments of this application.

[0217] The aforementioned storage media include: flash memory, hard disks, multimedia cards, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disks, optical discs, servers, APP (Application) application stores, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the aforementioned intelligent integrated design method for vacuum sintering processes of ultra-large ceramic components.

[0218] Example 4:

[0219] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned intelligent integrated design method for vacuum sintering process of ultra-large ceramic components.

[0220] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0221] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0222] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of equivalent technology of this disclosure, then the intent of this disclosure also includes such modifications and variations.

Claims

1. A smart integrated design method for vacuum sintering process of ultra-large ceramic components, characterized in that, include: Step S1: Obtain the sintering process parameters to be optimized and the sintering aid information to be optimized; Step S2: Determine the relative density, relative density weight, band gap optimization target, and band gap weight of silicon carbide ceramics; Step S3: Based on the sintering aid information to be optimized, the correlation between the sintering aid information and the component chemical formula of the DFT calculation object, and the second database, the DFT polymerization parameters are obtained. The second database is used to store the DFT parameters. Step S4: Transform the sintering process parameters to be optimized, the sintering aid information to be optimized, and the DFT polymerization parameters into input features that the machine learning model can recognize; input the input features into the optimal machine learning model for prediction to obtain the relative density of the combination of sintering aid information and sintering process parameters. Step S5: Construct the particle swarm objective function to be minimized based on the relative density of the sintering aid information and the combination of sintering process parameters, the bandgap polymerization value in the DFT polymerization parameters, the relative density of silicon carbide ceramics, the relative density weight, the optimization objective of the bandgap, and the weight of the bandgap. Step S6: The sintering aid information to be optimized and the sintering process parameter combination to be optimized are used as particle swarm optimization variables. The particle swarm optimization algorithm is used to optimize the particle swarm objective function to obtain multiple sets of candidate sintering aid information and sintering process parameter combinations. Step S7: Using an active learning strategy, evaluate multiple sets of candidate sintering aid information and sintering process parameter combinations to obtain the uncertainty of each candidate sintering aid information and sintering process parameter combination; Step S8: Extract the N groups of candidate sintering aid information and the optimal combination of sintering process parameters with the largest uncertainty as the objects of sintering experiment verification. If the target performance after sintering experiment verification reaches the preset value, output the corresponding candidate sintering aid information and combination of sintering process parameters as the optimal combination of sintering aid information and sintering process parameters. If the target performance after sintering experiment verification does not reach the preset value, the sintering process parameters obtained after sintering experiment verification and the corresponding ceramic performance parameters after sintering are saved to the first database, the DFT parameters are saved to the second database, and the process returns to step S4 to re-select the optimal machine learning model.

2. The intelligent integrated design method for vacuum sintering process of ultra-large ceramic components according to claim 1, characterized in that, The optimal machine learning model was obtained through the following selection method: Multiple machine learning models targeting different performance targets are trained based on a first database, a second database, and a correlation database. The first database stores historical sintering aid information, historical sintering process parameters, and corresponding post-sintering ceramic performance parameters. The second database stores DFT parameters, including: the chemical formula of the DFT calculation object, the ratio of metal to non-metal atoms, the first ionization energy, the metal atomic radius, the electronegativity ratio, thermodynamic stability, interfacial reactivity, charge transfer, and band gap. The correlation database stores the historical sintering aid information in the first database, the chemical formula of the corresponding DFT calculation object in the second database, and the correlation between other DFT parameters.

3. The intelligent integrated design method for vacuum sintering process of ultra-large ceramic components according to claim 2, characterized in that, The training of multiple machine learning models with different target performance based on the first database, the second database, and the associated database includes: Based on the historical sintering aid information in the first database and the association relationships in the related database, the component chemical formula of the DFT calculation object in the second database is obtained. Based on the component chemical formula of the DFT calculation object, the DFT polymerization parameters are obtained. The DFT polymerization parameters include: band gap polymerization value and thermodynamic stability polymerization value. Historical sintering aid information, historical sintering process parameters, and DFT polymerization parameters from the first database are used as a sintering sample and as the input feature matrix for multiple machine learning models. The sintered ceramic performance parameters from the first database are used as the optimization targets for multiple machine learning models. The outputs of multiple machine learning models are used as the model prediction performance. The errors of the model prediction performance and the optimization targets of multiple machine learning models are used as the evaluation criteria for multiple machine learning models. The machine learning model with the smallest error is selected as the optimal machine learning model.

4. The intelligent integrated design method for vacuum sintering process of ultra-large ceramic components according to claim 3, characterized in that, The process of obtaining DFT polymerization parameters based on the component chemical formulas of the DFT-calculated object includes: When a historical sintering aid information corresponds to the component chemical formula of only one DFT calculation object, the DFT parameters are directly used as DFT polymerization parameters. When a historical candidate sintering aid information corresponds to the component chemical formula of multiple DFT calculation objects, the DFT parameters corresponding to the component chemical formulas of multiple DFT calculation objects are averaged and polymerized to obtain the polymerization value of the DFT parameters. The polymerization value of the DFT parameters is then used as the final DFT polymerization parameter.

5. The intelligent integrated design method for vacuum sintering process of ultra-large ceramic components according to claim 3, characterized in that, The bandgap polymerization value and the thermodynamic stability polymerization value are calculated using the following formulas: Bandgap aggregation value: ； in, Let be the bandgap polymerization value in the i-th sintered sample. The band gap value of the component chemical formula of the k-th DFT calculation object in the i-th sintered sample; Thermodynamic stability of polymerization value: ； in, The thermodynamic stability polymerization value in the i-th sintered sample. is the thermodynamic stability value of the component chemical formula of the k-th DFT calculation object in the i-th sample.

6. The intelligent integrated design method for vacuum sintering process of ultra-large ceramic components according to claim 1, characterized in that, The objective function of the particle swarm to be minimized is expressed as follows: ； Where F(x) is the particle swarm optimization objective function to be minimized; The relative density of the combination of sintering aid information and sintering process parameters predicted by the optimal machine learning model; This refers to the bandgap aggregation value in the DFT aggregation parameters; For the target relative density; E g , target W1 represents the target bandgap; W2 represents the relative density weight; W3 represents the bandgap weight.

7. The intelligent integrated design method for vacuum sintering process of ultra-large ceramic components according to claim 1, characterized in that, The active learning strategy is used to evaluate multiple sets of candidate sintering aid information and sintering process parameter combinations to obtain the uncertainty of each candidate sintering aid information and sintering process parameter combination, including: The random forest algorithm is used to predict the information of candidate sintering aids and the combination of sintering process parameters, respectively, and the predicted value of each tree is obtained. The average of the squared deviations between the predicted values ​​of each tree and the average predicted value is calculated to obtain the uncertainty of the combination of candidate sintering aid information and sintering process parameters.

8. An electronic device, characterized in that, include: One or more processors, and a memory for storing instructions that, when executed by the one or more processors, cause the one or more processors to perform the intelligent integrated design method for vacuum sintering process of ultra-large ceramic components as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed, cause the processor to perform the intelligent integrated design method for vacuum sintering process of ultra-large ceramic components as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by the processor, they implement the intelligent integrated design method for vacuum sintering process of ultra-large ceramic components as described in any one of claims 1 to 7.

Images