A Deep Learning-Based Intelligent Composition Screening Method for High Dielectric Constant Ceramic Specimens
By employing a deep learning-based approach combined with orthogonal/uniform design and a two-step sintering method, a closed-loop iterative system of small-sample training, model prediction, and experimental verification was constructed. This system solved the problem of low efficiency in traditional trial-and-error methods, enabling efficient screening of barium titanate-based ceramic compositions with high dielectric constants and low dielectric losses, thereby improving the model's prediction accuracy and production stability.
Patent Information
- Application Number
- CN202611611773.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-19
- Publication Date
- 2026-11-17
AI Technical Summary
Traditional trial-and-error methods are inefficient in optimizing the dielectric properties of barium titanate-based ceramics, have difficulty in multi-element synergistic optimization, have poor model prediction accuracy under small sample conditions, lack embedding of the physical mechanism of perovskite ceramics, and have not established a closed-loop iterative system of small sample training-model prediction-experimental verification, thus failing to effectively handle the risks of liquid phase sintering and nonlinear effects.
A deep learning-based approach was adopted, by designing 30-50 representative formulations to construct a training dataset, calculating physical descriptors, and training the model using random forest or XGBoost algorithms. Machine learning was then performed, and a closed-loop iterative process of small-sample training, model prediction, and experimental verification was achieved by combining orthogonal/uniform design and a two-step sintering method. This process screened out ceramic compositions with high dielectric constant and low dielectric loss.
It shortened the R&D cycle, reduced material costs, improved the model's generalization ability and prediction accuracy, ensured the stability and reliability of the selected formulations in actual production, reduced the failure rate of verification experiments, and accelerated the commercialization process of ceramic materials.
Smart Images

Figure CN123120156A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer-aided design technology, and particularly relates to a method for intelligent screening of high dielectric constant ceramic components based on deep learning, using BT / BCT-based materials. Background Technology
[0002] Barium titanate (BaTiO3, BT)-based ceramics are widely used in multilayer ceramic capacitors (MLCCs) and other fields due to their excellent dielectric properties. To further improve the dielectric constant (K value) and reduce the dielectric loss (tanδ), A-site or B-site doping modification is usually adopted, such as adding CaCO3, WO3, V2O5, ZrO2, Ho2O3, NiO, SiO2, Co2O3, etc. However, the variety of doping elements is large (often more than 9), and the molar concentration combination space is huge. The traditional "trial and error method" has a long experimental cycle and high cost, making it difficult to efficiently locate the optimal formula with high K (>3000) and low loss (<2%).
[0003] In recent years, machine learning-assisted material design has become a research hotspot. However, existing technologies are mostly based on pure data-driven approaches and lack the embedding of the physical mechanisms of perovskite ceramics, resulting in poor model generalization ability and inaccurate predictions under small sample conditions. In addition, existing methods have not established a closed-loop iterative system of "small sample training - model prediction - experimental verification" and lack targeted treatment for nonlinear effects such as the risk of liquid phase sintering of V2O5-SiO2 and core-shell structure. Summary of the Invention
[0004] This invention aims to provide a machine learning-assisted method for predicting and validating BT / BCT-based high-k ceramic formulations, addressing the problems of low efficiency, difficulty in multi-element synergistic optimization, and poor model prediction accuracy under small sample conditions in traditional trial-and-error methods. The specific solution is as follows: A deep learning-based intelligent screening method for the composition of BT / BCT-based high dielectric constant ceramics, the method specifically includes the following steps: Step S1: Using barium titanate or barium calcium titanate as the matrix, determine the type of dopant, the range of molar concentration of the dopant, and the doping mechanism. Step S2: Design the formula and preparation process, prepare the ceramic product, and test the dielectric constant K and dielectric loss tanδ to form a training dataset; Step S3: Calculate the physical descriptor for each recipe group; Step S4: Use the physical descriptor and dopant molar concentration of each formulation as the model input features to form a training set. The output targets are dielectric constant K and dielectric loss tanδ. Then, train the machine learning model. Step S5: Use the trained model to randomly generate virtual recipes and predict the dielectric constant K and dielectric loss tanδ, and perform virtual screening and Pareto optimal recipe acquisition. Step S6: Conduct experimental verification and model iteration of the optimal formula.
[0005] Furthermore, the type of dopant, the range of molar concentrations, and the doping mechanism include: The dopant is BaCO3, with a molar concentration of 0~1.5%. The doping mechanism is A-site excess to compensate for barium volatilization. The dopant is WO3, with a molar concentration of 0-0.5%, and the doping mechanism is B-site donor doping. 6+ Replace Ti 4+ ; The dopant V2O5 has a molar concentration of 0~0.5% and acts as a liquid-phase sintering aid. The dopant ZrO2, with a molar concentration of 0~0.5%, is used for doping via B-site modification, which broadens the dielectric temperature spectrum. The dopant Ho2O3 has a molar concentration of 0-1.0% and its doping mechanism is amphoteric doping at the A / B sites, with acceptor behavior. The dopant is CaCO3, with a molar concentration of 0-1.5%, and the doping mechanism is A-site modification and moving Curie point. The dopant is NiO, with a molar concentration of 0~0.5%, and the doping mechanism is B-site acceptor doping. 2+ Replace Ti 4+ ; The dopant SiO2 has a molar concentration of 0-1.5%, and the doping mechanism is grain boundary modification to form an insulating layer. The dopant is Co2O3, with a molar concentration of 0~0.5%. The doping mechanism is B-site acceptor doping, and it is a sintering aid.
[0006] Furthermore, in step S2, 30 to 50 formulations are designed using orthogonal design or uniform design methods, including single doping and multi-element complexes, where multi-element complexes include known synergistic combinations of different dopants. Furthermore, the formulation includes: Nine single-doped formulations (each dopant is individually measured for its upper molar concentration limit and 1 / 2 molar concentration limit), totaling 18 formulations; There are 15 combinations of two elements, including: Groups 1-3 consist of V2O5 + SiO2, with molar concentrations of 0.1 mol% V2O5 + 0.3 mol% SiO2, 0.25 mol% V2O5 + 0.75 mol% SiO2, and 0.4 mol% V2O5 + 1.2 mol% SiO2, respectively. Groups 4-6 are CaCO3+SiO2, with concentrations of 0.3 mol% CaCO3+0.3 mol% SiO2, 0.75 mol% CaCO3+0.75 mol% SiO2, and 1.2 mol% CaCO3+1.2 mol% SiO2. Groups 7-9 consist of WO3 + Ho2O3, with concentrations of 0.1 mol%WO3 + 0.2 mol%Ho2O3, 0.25 mol%WO3 + 0.5 mol%Ho2O3, and 0.4 mol%WO3 + 0.8 mol%Ho2O3. Groups 10-12 are NiO+Co2O3, with concentrations of 0.1 mol% NiO+0.1 mol% Co2O3, 0.25 mol% NiO+0.25 mol% Co2O3, and 0.4 mol% NiO+0.4 mol% Co2O3. Groups 13~15 are BaCO3+CaCO3, take 0.3mol%BaCO3+0.3mol%CaCO3, 0.75mol%BaCO3+0.75mol%CaCO3, 1.2mol%BaCO3+1.2mol% CaCO3; There are 9 combinations of the three elements, including: Groups 16 to 18 are V2O5+SiO2+CaCO3, take 0.1mol%V2O5+0.3mol%SiO2+0.3mol%CaCO3, 0.25mol%V2O5+0.75mol%SiO2+0.75mol%CaCO3, 0.4mol%V2O5+1.2mol%SiO2+1.2mol%CaCO3; Groups 19-21 consist of WO3 + Ho2O3 + NiO, with the following concentrations: 0.1 mol% WO3 + 0.2 mol% Ho2O3 + 0.1 mol% NiO, 0.25 mol% WO3 + 0.5 mol% Ho2O3 + 0.25 mol% NiO, and 0.4 mol% WO3 + 0.8 mol% Ho2O3 + 0.4 mol% NiO. Groups 22 to 24 are BaCO3+CaCO3+SiO2, take 0.3mol%BaCO3+0.3mol%CaCO3+0.3mol%SiO2, 0.75mol%BaCO3+0.75mol%CaCO3+0.75mol%SiO2, 1.2mol%BaCO3+1.2mol%CaCO3+1.2mol%SiO2; Seven sets of boundary extreme formulations: Group 1: x1 (BaCO3) = 1.5 mol%, x6 (CaCO3) = 1.5 mol%. Group 2: x3 (V2O5) = 0.5 mol%, x8 (SiO2) = 1.5 mol% Group 3: x2 (WO3) = 0.5 mol%, x5 (Ho2O3) = 1.0 mol% Group 4: x4 (ZrO2) = 0.5 mol%, x7 (NiO) = 0.5 mol% Group 5: x9 (Co2O3) = 0.5 mol%, x1 (BaCO3) = 1.5 mol% Group 6: x3 (V2O5) = 0.5 mol%, x6 (CaCO3) = 1.5 mol%, x8 (SiO2) = 1.5 mol% Group 7: x2 (WO3) = 0.5 mol%, x5 (Ho2O3) = 1.0 mol%, x9 (Co2O3) = 0.5 mol%.
[0007] Furthermore, the preparation process and subsequent testing include: Ingredients and ball milling: Weigh the raw materials according to the formula, use anhydrous ethanol as the medium, ball-to-material ratio of 2:1, and ball mill for 12 hours. The raw material accuracy is 0.0001g, and the raw material purity is ≥99.9%. Pre-sintering synthesis: Hold at 1000~1100℃ for 2 hours to form the perovskite main crystal phase; Granulation and molding: Polyvinyl alcohol is added for granulation, and the mixture is pressed into Φ10×1.5mm discs under a pressure of 200MPa; Sintering: Hold at 1300~1400℃ for 2 hours, then cool to room temperature in the furnace; Electrode and Testing: Silver paste was applied, and the dielectric constant K and dielectric loss tanδ were measured at a frequency of 1 kHz. Each sample was repeated 3 times and the average value was taken.
[0008] Further, step S3, calculating the physical descriptor, includes: 1) Average A / B site ion radius ratio r A / r B : ; Where, r A Represents the weighted average radius of A-site ions; c Ai The mole fraction of the i-th type of A-site dopant ion relative to the total number of A-site ions; Ba is the matrix 2+ The mole fraction of the total A-position; r Ba Ba is the matrix 2+ The Shannon effective ionic radius of the ion; ri The radius of the i-th type of A-site doped ion r B Represents the weighted average radius of the B-site ion; c Bj denoted as the mole fraction of the j-th type of B-site dopant ion relative to the total number of B-site ions; Ti as the matrix 4+ The mole fraction of ions in the total B sites; r Ti Ti as the matrix 4+ The Shannon effective ionic radius of the ion; r j Let be the radius of the j-th B-site doped ion; 2) Valence equilibrium factor ΔV: ; Among them, c i It is the molar concentration of the i-th donor ion; v i It is the valence of the i-th donor ion; v host It is the valence of the matrix ion that has been substituted; c j It is the molar concentration of the j-th acceptor ion; v j It is the valence of the j-th acceptor ion; 3) Electronegativity difference Δχ: ; in The average electronegativity of the A-site ions in the perovskite structure; The average electronegativity of the B-site ion in the perovskite structure; The calculation formula is: ; Where n k The mole fraction of the kth ion occupying the A site refers to the proportion of the kth ion occupying the A site in the total amount of substance of all A-site ions. χ k Pauling electronegativity of the k-th ion occupying the A site; The calculation formula is: ; Where n m The mole fraction of the m-th ion occupying the B-site refers to the proportion of the m-th ion occupying the B-site in the total amount of substance of all B-site ions. χm Pauling electronegativity of the m-th ion occupying the B site; 4) Tolerance factor t: ; Where, r O The radius of the oxygen ion; r A The weighted average radius of the A-site ion; r B is the weighted average radius of the B-site ion.
[0009] Furthermore, step S4, machine learning model training, includes: Choose either Random Forest or XGBoost algorithm. The input features include 9 doping concentrations + 4 physical descriptors (13 dimensions in total), and the output targets are dielectric constant K and dielectric loss tanδ. Leave-one-out cross-validation was used, and the evaluation metrics included R² score, mean squared error (MSE), and feature importance ranking. Regularization is used during model training to prevent overfitting; The nine doping concentrations refer to the molar concentrations of BaCO3, WO3, V2O5, ZrO2, Ho2O3, CaCO3, NiO, SiO2, and Co2O3, respectively. The four physical descriptors are: average A / B site ion radius ratio (r) A / r B ), valence equilibrium factor (ΔV), electronegativity difference (Δχ), and tolerance factor (t).
[0010] Further, step S5 includes: Using the trained model, 10,000 candidate formulations that meet the range of doping molar concentrations are randomly generated using the Monte Carlo method, and K and tanδ are predicted for each candidate formulation. The objective function is set as K≥3000 and tanδ≤2%. Pareto front screening is used to obtain the top 10 formulations with the best performance.
[0011] Further, step S6 includes: Select 3 to 5 of the top 10 formulas for experimental preparation and testing; Add the newly measured real data to the model's training set, retrain the model, and repeat step S5 to form a closed loop of "small sample training - model prediction - experimental verification - data expansion - model iteration" until the formula with model prediction accuracy R²≥0.85 and K≥3000 and tanδ≤2% is stably reproduced. Stable reproducibility refers to the following: under the same formula and process conditions, samples prepared and tested independently for three consecutive times have an average dielectric constant K of ≥3000 and a relative standard deviation of ≤5%, and an average dielectric loss tanδ of ≤2% and a relative standard deviation of ≤10%.
[0012] Furthermore, the sintering process is replaced by a two-step sintering method: the shaped sample is heated to 1350°C at a heating rate of 5°C / min and held for 10-15 minutes, then cooled to 1250°C at a rapid cooling rate of 10-20°C / min and held at this temperature for 4-6 hours, and finally cooled to room temperature in the furnace. To prevent abnormal liquid-phase sintering caused by high concentrations of both V2O5 and SiO2, combinations of V2O5 > 0.3 mol% (molar concentration) and SiO2 > 1.0 mol% (molar concentration) are excluded in the Monte Carlo generation process. During the model training phase, the sintering temperature T is set. s It is added to the model as the 14th dimension input feature.
[0013] Compared with the prior art, the present invention has at least one of the following technical effects: 1. This invention employs a closed-loop iterative strategy of "small sample training - model prediction - experimental verification." It trains a machine learning model using experimental data from 30-50 representative formulations, and then utilizes Monte Carlo virtual screening to quickly locate the optimal region from tens of thousands of candidate formulations. Only 3-5 verifications are needed to obtain a high-K, low-loss formulation. This compresses the R&D process, which originally required hundreds of trial-and-error experiments, to less than 50, resulting in shorter R&D cycles, reduced material costs, and accelerated commercialization of high-K ceramic materials.
[0014] 2. This invention specifically sets up 15 two-element combinations and 9 three-element compound formulations in the orthogonal / homogeneous design (covering V2O5+SiO2 liquid-phase sintering pairs, CaCO3+SiO2 core-shell structure pairs, WO3+Ho2O3 amphoteric competition pairs, etc.), and supplements them with 7 boundary extreme formulations, enabling the model to learn the interaction and nonlinear response when multiple elements coexist. The model of this invention can accurately identify synergistic enhancement or inhibition effects. For example, the risk of abnormal grain growth when V2O5 and SiO2 are both at high concentrations can be automatically avoided by the model, and the experimental reproducibility of high-K formulations in the prediction results is good.
[0015] 3. This invention embeds the physical mechanisms of perovskite ceramics into feature engineering, calculating four physical descriptors: average A / B site ion radius ratio, valence equilibrium factor, electronegativity difference, and tolerance factor. These, along with nine doping concentrations, constitute a 13-dimensional input feature, endowing the model with prior knowledge of crystal chemistry. Simultaneously, it employs regularization hyperparameters (max) from random forests or XGBoost. depth Limitation, minsamples_split Cross-validation and grid search prevent overfitting of the model with small sample sizes and improve generalization ability. Leave-one-out cross-validation is used with only 30-50 training samples. 2 The accuracy can reach above 0.85, and the prediction mean square error is below 0.15, which is far superior to the pure concentration input model (R²). 2 (≈0.6), and the model's extrapolation ability is enhanced, still able to give reasonable predictions for extreme combinations that have not been trained.
[0016] 4. This invention incorporates sintering temperature as an optional input feature into the model and averages the results after multiple sampling and prediction of each candidate formulation within a reasonable temperature range during the virtual screening stage. This effectively absorbs sintering temperature fluctuations of ±15℃ in actual production, ensuring that the screened formulation maintains a stable high dielectric constant and low loss within the process window. Experimental verification shows that the optimal formulation exhibits K-value fluctuations of less than 8% and tanδ fluctuations of less than 0.3% within the 1330~1370℃ range, improving the industrial transferability and production yield of the formulation.
[0017] 5. This invention sets seven sets of boundary extreme formulations (e.g., BaCO3 + CaCO3 both at 1.5%, V2O5 + SiO2 both at the upper limit, etc.), enabling the model to learn the performance response in the doping space limit region. It can also automatically exclude high-risk combinations (e.g., V2O5 > 0.3% and SiO2 > 1.0%) during Monte Carlo generation. This boundary constraint mechanism effectively prevents the model from predicting false positive formulations that cannot be sintered in reality or exhibit drastic performance changes, reducing the failure rate of verification experiments after virtual screening compared to traditional methods.
[0018] 6. This invention utilizes feature importance ranking from random forests or XGBoost outputs and SHAP value analysis to quantitatively assess the contributions of nine dopants and four physical descriptors to dielectric constant and loss. For example, it can clearly identify the influence weights of the liquid-phase sintering effect of V₂O₅, the grain boundary insulation effect of SiO₂, and the amphoteric occupancy competition of Ho₂O₃. This mechanism analysis report can guide researchers to adjust doping strategies in a targeted manner, such as prioritizing the optimization of highly important elements or avoiding combinations that produce antagonistic effects, thereby forming a continuously optimized materials genome technology platform.
[0019] 7. This invention designs dopant combinations with known synergistic, antagonistic, or coupling effects (such as V2O5+SiO2 liquid-phase sintering pairs, CaCO3+SiO2 core-shell structure pairs, WO3+Ho2O3 amphoteric competitive pairs, etc.) and sets three concentration gradients (low, medium, and high). This enables the model to learn the interaction and nonlinear response when multiple elements coexist, avoiding prediction distortion caused by linear extrapolation based solely on the main effects of a single dopant. This accurately captures the synergistic enhancement or suppression effects in complex doped systems and improves the reliability of virtual screening.
[0020] 8. This invention enables the model to independently learn the main effects of each dopant on dielectric constant and loss (i.e., the pure effect without interference from other dopants) by adding each dopant individually to its upper limit or half of the upper limit concentration (with the remaining dopants at 0). This provides a benchmark for subsequent analysis of multi-element synergistic effects and allows for the identification of high-contribution dopants for feature importance ranking.
[0021] 9. This invention effectively avoids process failures such as abnormal grain growth, grain boundary insulation layer rupture, and sharp increase in dielectric loss caused by the formation of a eutectic liquid phase during sintering by excluding high-risk combinations of V2O5 > 0.3 mol% and SiO2 > 1.0 mol% in advance during the virtual screening stage. This improves the actual sinterability and performance stability of the model-recommended formulation, reduces the failure rate of verification experiments caused by excessive liquid phase, and guides the model to search for high-K, low-loss formulations within the safe doping region, ensuring a high conversion rate from virtual screening to experimental verification.
[0022] 10. The two-step sintering method of this invention suppresses abnormal grain growth caused by the liquid phase when V2O5 and SiO2 coexist, resulting in a more uniform microstructure of the ceramic and a more stable and reproducible relationship between properties (K and tanδ) and doping concentration. This reduces experimental data noise caused by process runaway. This means that the input labels (measured K / tanδ) during model training are less affected by accidental abnormal grain growth, improving the signal-to-noise ratio of the doping-property mapping relationship. Consequently, the machine learning model can more accurately learn the essential relationship between dopant concentration and dielectric properties, improving the reproducibility and predictive robustness of the virtually selected formulation under actual sintering conditions.
[0023] 11. The feature engineering of this invention (average ionic radius ratio, valence equilibrium factor, electronegativity difference, tolerance factor) is based on the general principles of perovskite crystal chemistry and does not depend on a specific matrix composition. Therefore, the trained model framework can be directly transferred to other A-site or B-site doped modified systems, such as lead-free piezoelectric ceramics like (Ba,Sr)TiO3 and (K,Na)NbO3. Only the physical descriptor needs to be recalculated and the hyperparameters fine-tuned according to the doping elements of the new system, providing a high-value starting point for the development of new systems.
[0024] 12. This invention requires only 30-50 sets of small sample data to complete model training and virtual screening, which is more than three orders of magnitude more efficient than traditional design spaces (4-5 levels for each of 9 dopants, requiring hundreds of thousands of sets for full factorial experiments). Through a combination of uniform design, single doping, compounding, and boundary extreme formulations, the amount of information provided by each experimental sample is maximized, avoiding redundant experiments. In practical applications, it reduces material consumption and shortens testing time, and is particularly suitable for low-cost, rapid exploration of systems sensitive to the content of noble metal oxides (such as Ho₂O₃). Attached Figure Description
[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments 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.
[0026] Figure 1 This is a flowchart illustrating a method for intelligent screening of high dielectric constant ceramic components based on deep learning, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation
[0027] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.
[0028] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.
[0029] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.
[0030] As used in the specification and appended claims of the present application, the term "if" may be interpreted as "when...", or "once", or "in response to determining" or "in response to detecting" depending on the context. Similarly, the phrase "if it is determined" or "if [the described condition or event] is detected" may be interpreted, depending on the context, to mean "once it is determined" or "in response to determining" or "once [the described condition or event] is detected" or "in response to detecting [the described condition or event]".
[0031] In addition, in the description of the specification and appended claims of the present application, the terms "first", "second", "third" and the like are only used for distinguishing the description, and shall not be construed as indicating or implying relative importance.
[0032] References to "one embodiment" or "some embodiments" and the like described in the specification of the present application mean that a particular feature, structure or characteristic described in connection with the embodiment is included in one or more embodiments of the present application. Thus, the appearances of the phrases "in one embodiment", "in some embodiments", "in some other embodiments", "in further some embodiments" and the like in various places in the present specification are not necessarily all referring to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in another manner. The terms "comprising", "including", "having" and their variants all mean "including but not limited to", unless otherwise specifically emphasized in another manner.
[0033] Referring to the attached draw Figure 1 , the present invention discloses an intelligent screening method for BT / BCT-based high dielectric constant ceramic components based on deep learning, which specifically comprises: Step 1: Construct a doping variable space and a matrix system.
[0034] The matrix material is BaTiO3 (BT) or Ba 1-y Ca y TiO3 (BCT, where 0<y≤0.15), and the stoichiometric ratio is controlled as (Ba+Ca) / Ti=1 (molar ratio). The doping variables are the following 9 raw materials, and their molar concentration ranges (mol%, relative to Ti or (Ba+Ca)) and action mechanisms are shown in Table 1 below: Table 1 The molar concentration of each dopant in the present invention is expressed in mole percentage (mol%), and the reference thereof is titanium (Ti) or the sum of barium (Ba) and calcium (Ca) (i.e., (Ba+Ca)) in the matrix material. Specifically: For the barium titanate (BaTiO3) matrix, the addition amount of the dopant is calculated by the amount of substance of the dopant added per 100 mol of Ti.
[0035] For barium calcium titanate (Ba 1-y Ca y The amount of dopant added is calculated as the amount of dopant added per 100 mol Ba+Ca matrix.
[0036] This definition ensures that all dopant molar concentrations are comparable under the same reference system and can be directly used as quantization inputs for feature variables in subsequent machine learning models.
[0037] Explanation of the working mechanism of each dopant: BaCO3 decomposes into BaO during sintering, providing additional Ba. 2+ Ions. In barium titanate-based ceramics, Ba at high temperatures 2+ Barium is easily volatile, leading to the loss of A-site cations and the formation of barium vacancies, which in turn reduces insulation resistance and increases dielectric loss. Adding an appropriate amount of BaCO3 compensates for A-site volatilization, bringing the (Ba+Ca) / Ti stoichiometric ratio closer to 1, reducing the molar concentration of barium vacancies, thereby stabilizing the perovskite structure and improving the temperature stability of the dielectric constant. Excessive addition (>1.5 mol%) introduces a barium-rich second phase, degrading dielectric properties.
[0038] W in WO3 6+ Ionic radius (approximately 0.60 Å, coordination number 6) and Ti 4+ (Approximately 0.605 Å) close, therefore W 6+ Ti that can replace the B position 4+ This belongs to donor doping (high-valence ions replacing low-valence ions). To maintain electroneutrality, the system generates electron carriers (Ti). 4+ Restore to Ti 3+ Alternatively, it can form cation vacancies to compensate. Electron carriers can reduce grain resistance, promote grain semiconductivity, and improve the dielectric constant. However, excessive WO3 (>0.5 mol%) can lead to over-semiconductivity, causing a sharp increase in dielectric loss and even a negative temperature coefficient.
[0039] V₂O₅ has a melting point of approximately 690℃, far below the ceramic sintering temperature (~1300℃). It forms a liquid phase early in sintering, encapsulating the grains and spreading along grain boundaries, significantly reducing the sintering activation energy and promoting densification. This liquid-phase sintering mechanism allows for the production of high-density ceramics at lower temperatures and also facilitates the rapid diffusion of dopant elements. Furthermore, V₂O₅… 5+ It can partially enter the B site and act as a donor to produce a semiconductor effect. However, the amount of V2O5 added must be strictly controlled (≤0.5mol%), otherwise excessive liquid phase will lead to abnormal grain growth and thinning of the grain boundary insulating layer, resulting in increased loss.
[0040] Zr 4+ The ionic radius (0.72 Å, coordination number 6) is greater than that of Ti.4+ (0.605 Å), the entry of ZrO2 into the B site causes lattice expansion, lowering the cubic-tetragonal phase transition temperature (Curie point) of barium titanate, thereby broadening the dielectric temperature spectrum near room temperature and improving temperature stability. Simultaneously, Zr... 4+ With Ti 4+ The heterovalent matching is good (isovalent substitution), and no additional charge carriers are generated, so no loss is increased. Appropriate Zr doping (≤0.5 mol%) can also suppress grain growth and obtain a fine-grained structure. The fine-grained effect itself helps to improve the dielectric constant.
[0041] Ho 3+ It is an amphoteric dopant with an ionic radius of approximately 1.02 Å (coordination number 12) and is similar to Ba. 2+ The difference (1.61 Å) is significant, but it is similar to Ca. 2+ (1.12 Å) is close, therefore it can occupy a small amount of A sites; at the same time, Ho 3+ The radius (0.90 Å, coordination number 6) is greater than that of Ti. 4+ (0.605 Å), can also occupy the B site. When Ho 3+ When occupying position A, a lower-priced replacement (Ho) 3+ Replace Ba 2+ This generates acceptor behavior, increases oxygen vacancies, inhibits grain semiconductivity, and improves insulation resistance; when a B site is occupied, high-valence substitution (Ho) occurs. 3+ Replace Ti 4+ This generates donor behavior, promoting semiconductorization. The actual site distribution depends on the composition and process. This dual-site competition mechanism allows Ho2O3 to simultaneously control grain resistance and grain boundary barrier, achieving a balance between dielectric constant and loss optimization.
[0042] CaCO3 decomposes into CaO, Ca 2+ The ionic radius (1.12 Å, coordination number 12) is smaller than that of Ba. 2+ (1.61 Å), mainly occupying A sites to replace Ba 2+ Due to Ca 2+ with Ba 2+ The difference in ion size causes partial shrinkage of the crystal structure, lowering the Curie temperature of barium titanate (approximately 3 °C / 1 mol% Ca). Appropriate Ca doping (<1.5 mol%) can shift the Curie point to near or slightly below room temperature, significantly improving the room temperature dielectric constant. Furthermore, Ca... 2+ It readily forms a Ca-Si-O insulating shell at the grain boundaries with SiO2, creating a "core-shell" structure. This structure can suppress excessive grain growth while improving breakdown strength and temperature stability.
[0043] Ni 2+ The ionic radius (0.69 Å, coordination number 6) is slightly larger than that of Ti. 4+(0.605 Å), occupying the B site to replace Ti 4+ Due to Ni 2+ The price level is lower than that of Ti 4+ This belongs to acceptor doping. To compensate for the negative charge, oxygen vacancies are generated in the system ( Oxygen vacancies can pin domain walls, reducing dielectric loss; simultaneously, acceptor doping can increase grain resistance, suppress semiconductivity, and improve insulation resistivity. However, the amount of NiO added should not exceed 0.5 mol%, otherwise the acceptor molar concentration will be too high, the grain boundary barrier will increase, and the dielectric constant will decrease significantly.
[0044] During sintering, SiO2 does not enter the perovskite lattice but segregates at grain boundaries, reacting with nearby BaO, CaO, etc., to form a glassy phase (such as BaSi2O5). This glassy phase has high resistivity, effectively isolating grains and forming a highly insulating grain boundary layer, thereby suppressing leakage current and reducing dielectric loss. Simultaneously, liquid-phase mass transfer contributes to densification. However, excessive SiO2 (>1.5 mol%) leads to excessively thick grain boundaries, and the capacitance series effect significantly reduces the apparent dielectric constant. When SiO2 and V2O5 coexist, they form a eutectic liquid phase at approximately 800°C, narrowing the sintering window and easily causing inhomogeneous grain growth; therefore, strict control of the simultaneous high molar concentration is necessary.
[0045] Co2O3 decomposes into Co3O4 or CoO at high temperatures, with Co ions forming Co ions. 2+ or Co 3+ It exists in form. Co 2+ The radius (0.745 Å, coordination number 6) occupies the B site, acting as an acceptor dopant similarly to NiO, increasing grain resistivity and reducing losses. Furthermore, CoO can form a eutectic liquid phase with barium titanate, lowering the sintering temperature and aiding in grain refinement. Co2O3 is unique in its sensitivity to oxygen partial pressure; it readily changes valence under reducing atmospheres, allowing for dynamic adjustment of carrier molar concentration. Therefore, it is commonly used in the terminal electrode barrier layer material of multilayer ceramic capacitors. In this system, its concentration is limited to 0.5 mol% to avoid excessive acceptor effect leading to a sharp decrease in dielectric constant.
[0046] Step 2: Construct the core dataset.
[0047] Using orthogonal or uniform design methods, 30–50 representative formulations were designed, covering the molar concentration range of various dopants, including single-doped and multi-element complexes. Samples were prepared according to the following process: Ingredients and ball milling: Weigh the raw materials according to the formula (accuracy 0.0001g, purity ≥99.9%), use anhydrous ethanol as the medium, ball-to-material ratio 2:1, and ball mill for 12 hours.
[0048] Pre-calcination synthesis: Hold at 1000~1100℃ for 2 hours to form the perovskite main crystal phase.
[0049] Granulation and molding: Polyvinyl alcohol (PVA) is added for granulation, and the mixture is pressed into Φ10×1.5mm discs under a pressure of 200MPa.
[0050] Sintering: Hold at 1300~1400℃ (adjust according to V and Si content) for 2 hours, then cool with the furnace.
[0051] Electrode and Testing: Silver paste was applied, and the dielectric constant K and dielectric loss tanδ were measured at a frequency of 1 kHz. Each sample was repeated 3 times and the average value was taken.
[0052] The experimental data for K and tanδ are obtained to form the training dataset.
[0053] In practice, the single-step sintering process of "holding at 1300~1400℃ (adjusted according to V and Si content) for 2 hours and then cooling with the furnace" can be replaced by a more optimized two-step sintering method: the shaped sample is heated to 1350℃ at a heating rate of 5℃ / min and held for 10~15 minutes (to allow the grains to briefly reach an activated state without excessive growth), and then cooled to 1250℃ at a rapid cooling rate of 10~20℃ / min. It is then held at this temperature for 4~6 hours (to promote grain boundary diffusion and densification, while inhibiting abnormal grain growth), and finally cooled to room temperature with the furnace.
[0054] This two-step sintering method utilizes the principle that grain boundary migration is suppressed while grain boundary diffusion remains active during long-term holding at low temperatures. This effectively prevents excessive growth of some grains caused by the formation of a eutectic liquid phase between V2O5 and SiO2, thereby obtaining high-density ceramics with uniform grain size and clear grain boundaries, while maintaining a high dielectric constant and low loss.
[0055] Orthogonal design and uniform design are two statistical design methods used in multi-factor, multi-level experiments. Orthogonal design arranges experiments using a pre-constructed orthogonal array, ensuring that the number of different level combinations of any two factors occurs equally. This allows for independent estimation of the main effects of each factor and analysis of interactions with fewer experiments. Uniform design, based on uniform distribution in number theory, only guarantees that experimental points are uniformly distributed within the factor space. It requires even fewer experiments and is suitable for screening experiments with multiple factor levels, primarily considering the trends in the influence of each factor. This scheme employs both methods to efficiently cover the entire doping space with 30-50 representative formulations across a vast combination space of nine dopants, each with multiple molar concentration ranges. This ensures both the distinguishability of single doping effects and the nonlinear information from multi-element coupling, thereby providing a training dataset with maximum information for machine learning models.
[0056] "Covering the molar concentration range of various dopants, including single doping and multi-element complexes" means that when designing 30-50 experimental formulations, it is essential to first ensure that each dopant (such as BaCO3, WO3, V2O5, etc.) has at least 2-3 different molar concentration points (including endpoints) within its given molar concentration range (e.g., 0-1.5%) to comprehensively reflect the impact of individual dopant variations on performance. Secondly, the experimental formulations should include both "single-doped" samples with only one dopant (for analyzing the independent effects of each element) and "multi-element complex" samples with two or more dopants added simultaneously (for capturing synergistic, antagonistic, or nonlinear coupling effects between elements). This provides sufficiently rich and representative training data for the machine learning model, preventing the model from learning only the main effects while ignoring interactions. Priority should be given to dopants or dopant combinations known in current literature and already used in production.
[0057] Specifically, the molar concentration (mol%) ranges of the nine dopants (x1~x9) are equally divided into multiple levels (e.g., upper limit of 1 / 4, upper limit of 2 / 4, upper limit of 3 / 4, upper limit, etc.), and the 30~50 representative formulations include: Nine single-doped formulations (each dopant is individually measured for its upper molar concentration limit and 1 / 2 molar concentration limit), totaling 18 formulations; By adding each dopant individually to its upper limit or half of its upper limit concentration (with the remaining dopant concentrations at 0), the model can independently learn the main effects of each dopant on dielectric constant and loss (i.e., the pure effects without interference from other dopants). This provides a benchmark for subsequent analysis of multi-element synergistic effects and allows for the identification of high-contribution dopants for feature importance ranking.
[0058] 24 sets of two-element and three-element compound formulations (covering known synergistic combinations such as V2O5+SiO2, CaCO3+SiO2, WO3+Ho2O3, NiO+Co2O3, BaCO3+CaCO3, etc., with the molar concentration of each component within each combination set at 0.25 times, 0.5 times, 0.75 times, etc., upper limits), specifically including: Two-element combinations (15 groups): Groups 1-3 consist of V2O5 + SiO2, with molar concentrations of 0.1 mol% V2O5 + 0.3 mol% SiO2, 0.25 mol% V2O5 + 0.75 mol% SiO2, and 0.4 mol% V2O5 + 1.2 mol% SiO2, respectively. Groups 4-6 are CaCO3+SiO2, with concentrations of 0.3 mol% CaCO3+0.3 mol% SiO2, 0.75 mol% CaCO3+0.75 mol% SiO2, and 1.2 mol% CaCO3+1.2 mol% SiO2. Groups 7-9 consist of WO3 + Ho2O3, with concentrations of 0.1 mol%WO3 + 0.2 mol%Ho2O3, 0.25 mol%WO3 + 0.5 mol%Ho2O3, and 0.4 mol%WO3 + 0.8 mol%Ho2O3. Groups 10-12 are NiO+Co2O3, with concentrations of 0.1 mol% NiO+0.1 mol% Co2O3, 0.25 mol% NiO+0.25 mol% Co2O3, and 0.4 mol% NiO+0.4 mol% Co2O3. Groups 13 to 15 are BaCO3+CaCO3, take 0.3mol%BaCO3+0.3mol%CaCO3, 0.75mol%BaCO3+0.75mol%CaCO3, and 1.2mol%BaCO3+1.2mol% CaCO3.
[0059] Three-element combinations (9 groups): Groups 16 to 18 are V2O5+SiO2+CaCO3, take 0.1mol%V2O5+0.3mol%SiO2+0.3mol%CaCO3, 0.25mol%V2O5+0.75mol%SiO2+0.75mol% CaCO3, 0.4mol%V2O5+1.2mol%SiO2+1.2mol%CaCO3; Groups 19-21 consist of WO3 + Ho2O3 + NiO, with the following concentrations: 0.1 mol% WO3 + 0.2 mol% Ho2O3 + 0.1 mol% NiO, 0.25 mol% WO3 + 0.5 mol% Ho2O3 + 0.25 mol% NiO, and 0.4 mol% WO3 + 0.8 mol% Ho2O3 + 0.4 mol% NiO. Groups 22 to 24 are BaCO3+CaCO3+SiO2, take 0.3mol%BaCO3+0.3mol%CaCO3+0.3mol%SiO2, 0.75mol%BaCO3+0.75mol%CaCO3+0.75mol%SiO2, 1.2mol%BaCO3+1.2mol%CaCO3+1.2mol%SiO2; By designing dopant combinations known to have synergistic, antagonistic, or coupling effects (such as V2O5+SiO2 liquid-phase sintering pairs, CaCO3+SiO2 core-shell structure pairs, WO3+Ho2O3 amphoteric competitive pairs, etc.), and setting three concentration gradients (low, medium, and high), the model can learn the interaction and nonlinear response when multiple elements coexist. This avoids prediction distortion caused by linear extrapolation based solely on the main effects of a single dopant, thus accurately capturing the synergistic enhancement or suppression effects in complex doped systems and improving the reliability of virtual screening.
[0060] 7 sets of boundary extreme formulas (e.g., x1 and x6 take the upper limit and the rest take 0, x8 and x3 take the upper limit and the rest take 0, x2 and x5 take the upper limit and the rest take 0, etc.), specifically including: Group 1: x1 (BaCO3) = 1.5 mol%, x6 (CaCO3) = 1.5 mol%; Group 2: x3 (V2O5) = 0.5 mol%, x8 (SiO2) = 1.5 mol% Group 3: x2 (WO3) = 0.5 mol%, x5 (Ho2O3) = 1.0 mol% Group 4: x4 (ZrO2) = 0.5 mol%, x7 (NiO) = 0.5 mol% Group 5: x9 (Co2O3) = 0.5 mol%, x1 (BaCO3) = 1.5 mol% Group 6: x3 (V2O5) = 0.5 mol%, x6 (CaCO3) = 1.5 mol%, x8 (SiO2) = 1.5 mol% Group 7: x2 (WO3) = 0.5 mol%, x5 (Ho2O3) = 1.0 mol%, x9 (Co2O3) = 0.5 mol%.
[0061] By pushing specific dopant combinations to their concentration limits (e.g., BaCO3+CaCO3 both at 1.5%, V2O5+SiO2 both at their limits, WO3+Ho2O3 both at their limits, etc.), extreme conditions in the doping space are artificially created. This allows machine learning models to learn the extreme performance responses when certain elements are at their highest concentrations and other elements are at zero. This prevents the model from over-extrapolating to untrained extreme regions during virtual screening, thus preventing unreliable predictions. It also helps identify potential risks such as nonlinear effects, second-phase precipitation, or runaway liquid-phase sintering under high doping concentrations, enhancing the model's ability to constrain boundary regions and its predictive robustness.
[0062] Step 3: Physical enhancement feature engineering.
[0063] For each formulation, the following four types of physical descriptors are calculated and used together with the molar concentrations of nine dopants as model input features: 1) Average A / B site ion radius ratio r A / r B (Dimensionless).
[0064] , In the formula r AIt represents the weighted average radius of the A-site ion (usually 12-coordinate, where 12-coordinate means that the number of nearest-neighbor oxygen ions around the A-site ion is 12, which is the typical coordination number of the A-site ion in the perovskite structure).
[0065] r A This refers to the ions at the A site (including the matrix Ba ions). 2+ and doped ions such as Ca 2+ Ho 3+ The mole fraction of each ion in the total A site is the average radius obtained by weighted summation of the effective Shannon ionic radii (coordination number 12) of each ion.
[0066] c Ai For the i-th type of A-site doped ion (such as Ca) 2+ Ho 3+ The mole fraction of the i-th type of A-site ion relative to the total A-site ions (regardless of whether it is the matrix Ba) 2+ Still doped Ca 2+ Ho 3+ The proportion of A-site cations in the total amount of substance of all A-site cations, which is dimensionless and the sum of the mole fractions of all A-site ions is equal to 1.
[0067] Ba is the matrix 2+ The mole fraction of the total A-position.
[0068] r Ba Ba is the matrix 2+ The Shannon effective ionic radius of the ion (coordination number 12).
[0069] r i denoted as the radius of the corresponding dopant ion.
[0070] r B It represents the weighted average radius of the B-site ion (usually 6-coordinate, which means that the number of nearest-neighbor oxygen ions around the B-site ion is 6, which is the typical coordination number of the B-site ion in the perovskite structure).
[0071] r B This refers to the ions at the B site (including the matrix Ti). 4+ and doped ions such as W 6+ V 5+ Zr 4+ Ni 2+ Co 2+ / 3+ Si 4+ And Ho, who may take the B position 3+The mole fraction of each ion in the total B site is the average radius obtained by weighted summation of the effective Shannon ionic radii (coordination number 12) of each ion.
[0072] c Bj For the j-th type of B-site doped ion (such as W) 6+ V 5+ Zr 4+ Ni 2+ Co 2+ / 3+ Si 4+ And Ho, who may take the B position 3 + The mole fraction of B-site ions relative to the total number of B-site ions (the j-th type of B-site ion (W) 6+ V 5+ Zr 4+ Ni 2+ Co 2+ / 3+ Si 4+ And Ho, who may take the B position 3+ The proportion of B-site cations in the total amount of substance, which is dimensionless and the sum of the mole fractions of all B-site ions is equal to 1.
[0073] Ti as the matrix 4+ The mole fraction of ions in the total B sites.
[0074] r Ti Ti as the matrix 4+ The Shannon effective ionic radius (coordination number 6) of the ion.
[0075] r j denoted as the radius of the corresponding dopant ion.
[0076] Average A / B site ion radius ratio r A / r B for: ; This ratio reflects the degree of matching between the A-site and B-site ion sizes in the perovskite structure, and has a direct impact on the tolerance factor and lattice distortion.
[0077] 2) Valence equilibrium factor ΔV (dimensionless).
[0078] ; c i It is the molar concentration (mol%) of the i-th donor ion. v i It is the valence of the i-th donor ion; vhost It is the substituted matrix ion (such as Ti) 4+ Or Ba 2+ The valence of ) Difference (v) i -v host A positive value indicates the number of additional positive charges provided by donor doping. in, This means calculating and summing the additional positive charge contributed by all donor doped ions (i.e., concentration multiplied by valence difference), and the summation term represents the total additional positive charge contributed by all donor dopants. c j It is the molar concentration (mol%) of the j-th acceptor ion; v j It is the valence of the j-th acceptor ion; Difference (v) host -v j A positive value indicates the amount of positive charge missing due to acceptor doping (i.e., the equivalent positive charge compensation requirement). This indicates that the positive charge loss (i.e., concentration multiplied by valence difference) caused by all acceptor doped ions is calculated and summed. In other words, this summation term represents the total additional positive charge loss or total "charge compensation requirement" contributed by all acceptor doping. Its value is positive and is used to subtract the donor term to determine the net doping effect.
[0079] The valence equilibrium factor ΔV is defined as the total additional positive charge contributed by all donor doping minus the total positive charge loss caused by all acceptor doping, and is used to characterize the degree of net donor or net acceptor in a ceramic system.
[0080] 3) Electronegativity difference Δχ (dimensionless).
[0081] ; in, The average electronegativity of the A-site ions in the perovskite structure is calculated using the following formula: ; Where, n k To occupy position A (including Ba) 2+ Ca 2+ Ho 3+ The mole fraction of the kth ion when it occupies position A. The mole fraction refers to the proportion of the kth ion occupying position A in the total amount of substance of all ions at position A. χ k To occupy position A (including Ba) 2+ Ca 2+ Ho 3+Pauling electronegativity of the k-th ion when it occupies the A-site; The average electronegativity of the B-site ions in the perovskite structure is calculated using the following formula: n m To occupy position B (W) 6+ V 5+ Zr 4+ Ni 2+ Co 2+ / 3+ Si 4+ Ho 3+ The mole fraction of the m-th ion when occupying the B-site. The mole fraction refers to the proportion of the m-th ion occupying the B-site in the total amount of substance of all B-site ions. χ m To occupy position B (including W) 6+ V 5+ Zr 4+ Ni 2+ Co 2+ / 3+ Si 4+ Ho 3+ Pauling electronegativity (dimensionless) of the m-th ion when it occupies the B site. The electronegativity difference Δχ reflects the difference in ionicity between the AO and BO chemical bonds. The larger Δχ is, the greater the electronegativity gradient between the AO and BO bonds, which is beneficial to charge polarization. However, if it is too large, it may lead to lattice instability.
[0082] 4) Tolerance factor t (dimensionless).
[0083] Where, r O The radius of the oxygen ion (usually taken as 1.40 Å, coordination number 6). r A The weighted average radius of the A-site ions, according to the ions at the A-site (Ba² + Ca² + Ho³ + Weighted calculation of mole fractions (etc.); r B The weighted average radius of the B-site ions, according to each B-site ion (e.g., W). 6+ V 5+ Zr 4+ Ni 2+ Co 2+ / 3+ Si 4+ And Ho, who may take the B position3+ The mole fraction weighted calculation of ). All ionic radii are Shannon effective ionic radii, with coordination number 12 for A-site ions and coordination number 6 for B-site ions.
[0084] The tolerance factor t is used to measure the stability of the perovskite structure: when t is between 0.88 and 1.05, a stable cubic or pseudo-cubic perovskite phase can usually be formed; if it deviates from this range, distortion or impurity phases may occur.
[0085] This formula is common knowledge in solid-state chemistry and crystallography, and will not be elaborated further.
[0086] Step 4: Machine learning model training.
[0087] The Random Forest or XGBoost algorithm was chosen, with input features consisting of 9 doping concentrations and 4 physical descriptors (13 dimensions in total), and output targets being the dielectric constant K and dielectric loss tanδ. Leave-one-out cross-validation was employed, and evaluation metrics included R² score, mean squared error (MSE), and feature importance ranking. Regularization (e.g., max) was used during model training. depth Limitation, min samples_split Preventing overfitting The nine doping concentrations refer to the molar concentrations (mol%) of BaCO3 (x1), WO3 (x2), V2O5 (x3), ZrO2 (x4), Ho2O3 (x5), CaCO3 (x6), NiO (x7), SiO2 (x8), and Co2O3 (x9) relative to the total amount of Ti or Ba+Ca.
[0088] The four physical descriptors are: average A / B site ion radius ratio (r) A / r B The valence equilibrium factor (ΔV), electronegativity difference (Δχ), and tolerance factor (t) are calculated using the formulas defined above, and respectively reflect lattice distortion, net doping effect, chemical bond polarity, and perovskite structure stability.
[0089] In practice, the detailed machine learning model training process is as follows: The 13-dimensional input features (9 doping concentrations + 4 physical descriptors) of the dataset (30~50 recipes) are standardized (Z-score normalization), and the output targets are dielectric constant K and dielectric loss tanδ.
[0090] We employ either a RandomForestRegressor or an XGBoost Regressor, using leave-one-out cross-validation: N-1 samples are trained at a time, and the remaining sample is predicted; this process is repeated N times, and the average R² and MSE are calculated. The model hyperparameters are set as follows: For RandomForest, n is set... estimators =100~300, max depth =5~10, min samples_split =5, min samples_leaf =2, max features ='sqrt'; for XGBoost, setting n estimators =100~200, learning rate =0.01~0.1, max depth =4~8,subsample=0.8,colsample bytree =0.8, reg alpha =0.1, reg lambda =1.0.
[0091] The optimal hyperparameters are selected using grid search (GridSearchCV) combined with cross-validation. After training, the model is output with a ranking of feature importance (based on Gini importance or SHAP value) and saved as a pickle or joblib file. The above parameter settings and procedures ensure that those skilled in the art can directly use Python's scikit-learn and xgboost libraries to train the model.
[0092] Step 5: Virtual screening and Pareto optimal formulation acquisition.
[0093] Using the trained model, 10,000 candidate formulations satisfying the concentration range of step one are randomly generated using the Monte Carlo method. K and tanδ are predicted for each candidate formulation. The objective function is set as follows: K ≥ 3000 and tanδ ≤ 2%. Pareto front screening is used to obtain the top 10 formulations with the best performance.
[0094] The Monte Carlo method randomly generates 10,000 candidate formulations that meet the concentration range of step one. Specifically, for each of the nine dopants, a concentration value is randomly sampled in a uniform distribution within the concentration range given in step one (e.g., BaCO30~1.5mol%, WO30~0.5mol%) to form a candidate formulation. This process is repeated 10,000 times to obtain 10,000 virtual formulations covering the entire doping space.
[0095] For each candidate formulation, first calculate its four physical descriptors (rA / r B The model is then trained by inputting 13 features (9 doping concentrations and 4 physical descriptors) into the trained model to predict the dielectric constant K and dielectric loss tanδ.
[0096] Pareto front screening is then performed: all candidate formulations are plotted on a two-dimensional plane according to the predicted K value and tanδ. With the optimization objective of "maximizing K and minimizing tanδ", all formulations that are not simultaneously dominated by other formulations (i.e., there is no other formulation with a higher K and a lower tanδ) are identified to form the Pareto front. From the Pareto front, formulations that satisfy K≥3000 and tanδ≤2% are further screened, and sorted by K value from high to low. The top 10 groups are selected as the output formulations.
[0097] The significance of this step lies in the fact that by replacing a large number of physical experiments with computer-based virtual screening, it is possible to quickly locate the potential optimal formulation region in an extremely large multi-element doping combination space, significantly reduce the number of samples actually prepared and tested (from tens of thousands to single digits), and significantly shorten the research and development cycle. At the same time, Pareto screening ensures the best balance between the two mutually restrictive properties of dielectric constant and loss, avoiding the sacrifice of one indicator for the optimization of one indicator.
[0098] In practice, to prevent abnormal liquid-phase sintering caused by the simultaneous high concentrations of V2O5 and SiO2, combinations of V2O5 > 0.3 mol% (molar concentration) and SiO2 > 1.0 mol% (molar concentration) are excluded in Monte Carlo generation.
[0099] In practice, during the model training phase, if there are temperature fluctuations in the actual sintering process (for example, different batches or different formulations may use different sintering temperatures of 1320~1380℃ due to differences in V2O5 and SiO2 content), the sintering temperature T can be adjusted. s (Unit: °C) is added to the model as the 14th dimension input feature.
[0100] The specific procedure is as follows: During the sample preparation process in step two, the actual sintering temperature (accurate to ±5℃) used for each formulation (a total of 30-50 formulations) is recorded, and this temperature, along with 9 doping concentrations and 4 physical descriptors, forms an extended feature vector. During model training, the sintering temperature is used as a continuous numerical feature input, and the algorithm automatically learns its influence weights on K and tanδ.
[0101] In the subsequent virtual screening (step five), for each candidate formulation, a reasonable sintering temperature range needs to be set according to its doping composition (for example, based on the V2O5 and SiO2 content, an empirical formula or rule is used: if V+Si are both high, a lower temperature of 1330℃ is used, otherwise a higher temperature of 1370℃ is used), and multiple sampling predictions are performed within this range (for example, the average of 3 different temperature predictions for each candidate formulation is taken): In the virtual screening stage, for each candidate formulation, a reasonable sintering temperature range (e.g., 1330~1370℃) is determined according to its doping composition (especially the V2O5 and SiO2 content), and then 3 different temperature values (e.g., 1330℃, 1350℃, 1370℃) are randomly selected within this range. These 3 temperatures are used as input features (along with the doping concentration and physical descriptor) and input into the trained model to predict the corresponding K values and tanδ. Finally, the arithmetic mean of these 3 prediction results is taken as the final predicted dielectric constant and dielectric loss of the candidate formulation.
[0102] This approach effectively simulates the minute fluctuations in sintering temperature during actual processes, preventing the model from becoming overly sensitive to a single temperature point. This ensures that the selected formulations maintain stable high performance even with temperature variations of ±10~20℃ in real production, thereby improving the robustness and practicality of virtual screening.
[0103] If this optional feature is not used, all samples are sintered at the same temperature by default, and the model will be unable to distinguish whether the performance difference comes from doping or process deviation.
[0104] Step Six: Experimental Verification and Model Iteration.
[0105] Select 3-5 formulations from the top 10 selected formulations for experimental preparation and testing (process as in step two). Add the newly measured real data to the dataset from step two, retrain the model, and repeat step five to form a closed loop of "small sample training - model prediction - experimental verification - data expansion - model iteration" until the formulations with model prediction accuracy R²≥0.85, K≥3000, and tanδ≤2% are stably reproduced.
[0106] Stable reproducibility means that, under the same formulation and process conditions, at least three consecutive independent preparations and tests of samples have an average dielectric constant K of ≥3000 and a relative standard deviation of ≤5%, and an average dielectric loss tanδ of ≤2% and a relative standard deviation of ≤10%.
[0107] The relative standard deviation (RSD) is equal to the standard deviation (SD) divided by the arithmetic mean. The absolute value of ) multiplied by 100%, that is 。
[0108] The formula for calculating the standard deviation (sample standard deviation) is as follows: First, calculate the arithmetic mean of all data. Then, calculate the sum of the squares of the differences between each data point and the mean, divide by (sample size minus 1), and finally take the square root. , where n is the number of times the sample is prepared and tested independently, and n≥3.
[0109] Step Six: Construct a closed-loop feedback mechanism of "small sample training—model prediction—experimental validation—data augmentation—model iteration": Prepare and test the 3-5 optimal formulations obtained from virtual screening to obtain real K and tanδ data. Add this new data to the original dataset, retrain and optimize the model, and then perform virtual screening and validation again. This process can significantly correct potential prediction biases in small sample sizes, gradually improving the model's extrapolation ability and prediction accuracy (target R²). 2 ≥0.85), and finally stably reproduced the high-performance formulation with K≥3000 and tanδ≤2%, while avoiding false positive predictions caused by insufficient training data or omissions in the initial design, ensuring that the selected formulation is equally effective under real process conditions.
[0110] Throughout this patent, the units for all symbols are as follows: dopant molar concentration is expressed as a mole percentage (mol%), based on the total amount of Ti or (Ba+Ca) in the matrix material; ionic radius (e.g., r) A r B r Ti r j r O All values (e.g., oxygen) use Shannon effective ionic radii, measured in angstroms (Å) or picometers (pm), where the oxygen ion radius r is... O Typically taken as 1.40 Å; average A / B site ion radius ratio r A / r B Valence equilibrium factor ΔV, electronegativity difference Δχ, and tolerance factor t are all dimensionless quantities; dielectric constant K is the relative dielectric constant, which is dimensionless; dielectric loss tanδ is dimensionless; sintering temperature Ts is in degrees Celsius (°C); ball milling time, holding time, etc., are all in hours (h) or minutes (min); pressure is in megapascals (MPa); heating or cooling rate is in °C / min.
[0111] See attached document Figure 2 The present invention also provides a computer device, including: a memory and a processor, and a computer program stored in the memory. When the computer program is executed on the processor, it implements a deep learning-based intelligent screening method for the composition of BT / BCT-based high dielectric constant ceramics as described in any of the above methods.
[0112] The computer device may be a desktop computer, laptop, handheld computer, or cloud server, etc. This computer device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that... Figure 2 The examples of computer devices are merely examples and do not constitute a limitation on computer devices. They may include more or fewer components than shown in the illustration, or combinations of certain components, or different components. For example, they may also include input / output devices, network access devices, etc.
[0113] The processor referred to can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.
[0114] In some embodiments, the memory may be an internal storage unit of the computer device, such as a hard drive or RAM. In other embodiments, the memory may be an external storage device of the computer device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory may include both internal and external storage units of the computer device. The memory is used to store the operating system, applications, boot loader, data, and other programs, such as the program code of the computer program. The memory can also be used to temporarily store data that has been output or will be output.
[0115] This invention also provides a computer-readable storage medium storing a computer program thereon. When the computer program is run by a processor, it implements a deep learning-based intelligent screening method for high dielectric constant ceramic components, as described in any of the above methods.
[0116] In this embodiment, if the integrated control unit is implemented as a software functional unit and sold or used as an independent product for intelligent screening of the composition of BT / BCT-based high-dielectric-constant ceramics based on deep learning, it can be stored in a computer-readable storage medium specifically designed for intelligent screening of the composition of BT / BCT-based high-dielectric-constant ceramics based on deep learning. Based on this understanding, all or part of the processes in the methods of the above embodiments can be implemented by instructing relevant hardware for intelligent screening of the composition of BT / BCT-based high-dielectric-constant ceramics based on deep learning through a specific computer program. This computer program can be stored in a dedicated computer-readable storage medium for intelligent screening of the composition of BT / BCT-based high-dielectric-constant ceramics based on deep learning. When this computer program is executed by a processor for intelligent screening of the composition of BT / BCT-based high-dielectric-constant ceramics based on deep learning, it can implement the application steps of the above-described method embodiments in intelligent screening of the composition of BT / BCT-based high-dielectric-constant ceramics based on deep learning. The computer program includes computer program code for intelligent screening of BT / BCT-based high-dielectric-constant ceramic components based on deep learning. This computer program code can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium may include at least: any entity or device capable of carrying the computer program code for intelligent screening of BT / BCT-based high-dielectric-constant ceramic components based on deep learning into an intelligent screening device for BT / BCT-based high-dielectric-constant ceramic components; recording media; machine tool computer memory; read-only memory (ROM); random access memory (RAM); and other media suitable for distributing software for intelligent screening of BT / BCT-based high-dielectric-constant ceramic components based on deep learning, such as a dedicated intelligent screening card for BT / BCT-based high-dielectric-constant ceramic components based on deep learning, a data storage card, etc.
[0117] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.
[0118] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0119] In the embodiments disclosed in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.
[0120] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
Claims
1. A method for intelligent screening of high dielectric constant ceramic composition based on deep learning using BT / BCT, characterized in that, The method specifically includes the following steps: Step S1: Using barium titanate or barium calcium titanate as the matrix, determine the type of dopant, the range of molar concentration of the dopant, and the doping mechanism. Step S2: Design the formula and preparation process, prepare the ceramic product, and test the dielectric constant K and dielectric loss tanδ to form a training dataset; Step S3: Calculate the physical descriptor for each recipe group; Step S4: Use the physical descriptor and dopant molar concentration of each formulation as the model input features to form a training set. The output targets are dielectric constant K and dielectric loss tanδ. Then, train the machine learning model. Step S5: Use the trained model to randomly generate virtual recipes and predict the dielectric constant K and dielectric loss tanδ, and perform virtual screening and Pareto optimal recipe acquisition. Step S6: Conduct experimental verification and model iteration of the optimal formula.
2. The intelligent screening method for high dielectric constant ceramics based on deep learning according to claim 1, characterized in that, The types of dopants, the range of molar concentrations of the dopant, and the doping mechanism include: The dopant is BaCO3, with a molar concentration of 0~1.5%. The doping mechanism is A-site excess to compensate for barium volatilization. The dopant is WO3, with a molar concentration of 0-0.5%, and the doping mechanism is B-site donor doping. 6+ Replace Ti 4+ ; The dopant V2O5 has a molar concentration of 0~0.5% and acts as a liquid-phase sintering aid. The dopant ZrO2, with a molar concentration of 0~0.5%, is used for doping via B-site modification, which broadens the dielectric temperature spectrum. The dopant Ho2O3 has a molar concentration of 0-1.0% and its doping mechanism is amphoteric doping at the A / B sites, with acceptor behavior. The dopant is CaCO3, with a molar concentration of 0-1.5%, and the doping mechanism is A-site modification and moving Curie point. The dopant is NiO, with a molar concentration of 0~0.5%, and the doping mechanism is B-site acceptor doping. 2+ Replace Ti 4+ ; The dopant SiO2 has a molar concentration of 0-1.5%, and the doping mechanism is grain boundary modification to form an insulating layer. The dopant is Co2O3, with a molar concentration of 0~0.5%. The doping mechanism is B-site acceptor doping, and it is a sintering aid.
3. The intelligent screening method for high dielectric constant ceramic composition based on deep learning according to claim 1, characterized in that, In step S2, 30 to 50 formulations are designed using orthogonal design or uniform design methods, including single doping and multi-element complexes. Multi-element complexes include known synergistic combinations of different dopants.
4. The intelligent screening method for high dielectric constant ceramic composition based on deep learning according to claim 1, characterized in that, The formula includes: Nine single-doped formulations (each dopant is individually measured for its upper molar concentration limit and 1 / 2 molar concentration limit), totaling 18 formulations; There are 15 combinations of two elements, including: Groups 1-3 consist of V2O5 + SiO2, with molar concentrations of 0.1 mol% V2O5 + 0.3 mol% SiO2, 0.25 mol% V2O5 + 0.75 mol% SiO2, and 0.4 mol% V2O5 + 1.2 mol% SiO2, respectively. Groups 4-6 are CaCO3+SiO2, with concentrations of 0.3 mol% CaCO3+0.3 mol% SiO2, 0.75 mol% CaCO3+0.75 mol% SiO2, and 1.2 mol% CaCO3+1.2 mol% SiO2. Groups 7-9 consist of WO3 + Ho2O3, with concentrations of 0.1 mol%WO3 + 0.2 mol%Ho2O3, 0.25 mol%WO3 + 0.5 mol%Ho2O3, and 0.4 mol%WO3 + 0.8 mol%Ho2O3. Groups 10-12 are NiO+Co2O3, with concentrations of 0.1 mol% NiO+0.1 mol% Co2O3, 0.25 mol% NiO+0.25 mol% Co2O3, and 0.4 mol% NiO+0.4 mol% Co2O3. Groups 13~15 are BaCO3+CaCO3, take 0.3mol%BaCO3+0.3mol%CaCO3, 0.75mol%BaCO3+0.75mol%CaCO3, 1.2mol%BaCO3+1.2mol% CaCO3; There are 9 combinations of the three elements, including: Groups 16 to 18 are V2O5+SiO2+CaCO3, take 0.1mol%V2O5+0.3mol%SiO2+0.3mol%CaCO3, 0.25mol%V2O5+0.75mol%SiO2+0.75mol%CaCO3, 0.4mol%V2O5+1.2mol%SiO2+1.2mol%CaCO3; Groups 19-21 consist of WO3 + Ho2O3 + NiO, with the following concentrations: 0.1 mol% WO3 + 0.2 mol% Ho2O3 + 0.1 mol% NiO, 0.25 mol% WO3 + 0.5 mol% Ho2O3 + 0.25 mol% NiO, and 0.4 mol% WO3 + 0.8 mol% Ho2O3 + 0.4 mol% NiO. Groups 22 to 24 are BaCO3+CaCO3+SiO2, take 0.3mol%BaCO3+0.3mol%CaCO3+0.3mol%SiO2, 0.75mol%BaCO3+0.75mol%CaCO3+0.75mol%SiO2, 1.2mol%BaCO3+1.2mol%CaCO3+1.2mol%SiO2; Seven sets of boundary extreme formulations: Group 1: BaCO3 = 1.5 mol%, CaCO3 = 1.5 mol%. Group 2: V₂O₅ = 0.5 mol%, SiO₂ = 1.5 mol% Group 3: WO3 = 0.5 mol%, Ho2O3 = 1.0 mol% Group 4: ZrO2 = 0.5 mol%, NiO = 0.5 mol% Group 5: Co₂O₃ = 0.5 mol%, BaCO₃ = 1.5 mol%; Group 6: V₂O₅ = 0.5 mol%, CaCO₃ = 1.5 mol%, SiO₂ = 1.5 mol%; Group 7: WO3=0.5mol%, Ho2O3=1.0mol%, Co2O3=0.5mol%.
5. The intelligent screening method for high dielectric constant ceramic composition based on deep learning according to claim 1, characterized in that, The preparation process and subsequent testing include: Ingredients and ball milling: Weigh the raw materials according to the formula, use anhydrous ethanol as the medium, ball-to-material ratio of 2:1, and ball mill for 12 hours. The raw material accuracy is 0.0001g, and the raw material purity is ≥99.9%. Pre-sintering synthesis: Hold at 1000~1100℃ for 2 hours to form the perovskite main crystal phase; Granulation and molding: Polyvinyl alcohol is added for granulation, and the mixture is pressed into Φ10×1.5mm discs under a pressure of 200MPa; Sintering: Hold at 1300~1400℃ for 2 hours, then cool to room temperature in the furnace; Electrode and Testing: Silver paste was applied, and the dielectric constant K and dielectric loss tanδ were measured at a frequency of 1 kHz. Each sample was repeated 3 times and the average value was taken.
6. The intelligent screening method for high dielectric constant ceramic composition based on deep learning according to claim 5, characterized in that, Step S3, calculating the physical descriptor, includes: 1) Average A / B site ion radius ratio r A / r B : ; Where, r A Represents the weighted average radius of A-site ions; c Ai The mole fraction of the i-th type of A-site dopant ion relative to the total number of A-site ions; Ba is the matrix 2+ The mole fraction of the total A-position; r Ba Ba is the matrix 2+ The Shannon effective ionic radius of the ion; r i The radius of the i-th type of A-site doped ion r B Represents the weighted average radius of the B-site ion; c Bj denoted as the mole fraction of the j-th type of B-site dopant ion relative to the total number of B-site ions; Ti as the matrix 4+ The mole fraction of ions in the total B sites; r Ti Ti as the matrix 4+ The Shannon effective ionic radius of the ion; r j Let be the radius of the j-th B-site doped ion; 2) Valence equilibrium factor ΔV: ; Among them, c i It is the molar concentration of the i-th donor ion; v i It is the valence of the i-th donor ion; v host It is the valence of the matrix ion that has been substituted; c j It is the molar concentration of the j-th acceptor ion; v j It is the valence of the j-th acceptor ion; 3) Electronegativity difference Δχ: ; in The average electronegativity of the A-site ions in the perovskite structure; The average electronegativity of the B-site ion in the perovskite structure; The calculation formula is: ; Where n k The mole fraction of the kth ion occupying the A site refers to the proportion of the kth ion occupying the A site in the total amount of substance of all A-site ions. χ k Pauling electronegativity of the k-th ion occupying the A site; The calculation formula is: ; Where n m The mole fraction of the m-th ion occupying the B-site refers to the proportion of the m-th ion occupying the B-site in the total amount of substance of all B-site ions. χ m Pauling electronegativity of the m-th ion occupying the B site; 4) Tolerance factor t: ; Where, r O The radius of the oxygen ion; r A The weighted average radius of the A-site ion; r B is the weighted average radius of the B-site ion.
7. The intelligent screening method for high dielectric constant ceramic composition based on deep learning according to claim 6, characterized in that, Step S4, machine learning model training, includes: Choose either Random Forest or XGBoost algorithm. The input features include 9 doping concentrations + 4 physical descriptors (13 dimensions in total), and the output targets are dielectric constant K and dielectric loss tanδ. Leave-one-out cross-validation was used, and the evaluation metrics included R² score, mean squared error (MSE), and feature importance ranking. Regularization is used during model training to prevent overfitting; The nine doping concentrations refer to the molar concentrations of BaCO3, WO3, V2O5, ZrO2, Ho2O3, CaCO3, NiO, SiO2, and Co2O3, respectively. The four physical descriptors are: average A / B site ion radius ratio (r) A / r B ), valence equilibrium factor (ΔV), electronegativity difference (Δχ), and tolerance factor (t).
8. The intelligent screening method for high dielectric constant ceramic composition based on deep learning according to claim 7, characterized in that, Step S5 includes: Using the trained model, 10,000 candidate formulations that meet the range of doping molar concentrations are randomly generated using the Monte Carlo method, and K and tanδ are predicted for each candidate formulation. The objective function is set as K≥3000 and tanδ≤2%. Pareto front screening is used to obtain the top 10 formulations with the best performance.
9. The intelligent screening method for high dielectric constant ceramics based on deep learning according to claim 4, characterized in that, Step S6 includes: Select 3 to 5 of the top 10 formulas for experimental preparation and testing; Add the newly measured real data to the model's training set, retrain the model, and repeat step S5 to form a closed loop of "small sample training - model prediction - experimental verification - data expansion - model iteration" until the formula with model prediction accuracy R²≥0.85 and K≥3000 and tanδ≤2% is stably reproduced. Stable reproducibility refers to the following: under the same formula and process conditions, samples prepared and tested independently for three consecutive times have an average dielectric constant K of ≥3000 and a relative standard deviation of ≤5%, and an average dielectric loss tanδ of ≤2% and a relative standard deviation of ≤10%.
10. The intelligent screening method for high dielectric constant ceramic composition based on deep learning according to claim 7, characterized in that, The sintering process is replaced by a two-step sintering method: the shaped sample is heated to 1350℃ at a heating rate of 5℃ / min and held for 10~15 minutes, then cooled to 1250℃ at a rapid cooling rate of 10~20℃ / min and held at this temperature for 4~6 hours, and finally cooled to room temperature in the furnace. To prevent abnormal liquid-phase sintering caused by high concentrations of both V2O5 and SiO2, combinations of V2O5 > 0.3 mol% (molar concentration) and SiO2 > 1.0 mol% (molar concentration) are excluded in the Monte Carlo generation process. During the model training phase, the sintering temperature T is set. s It is added to the model as the 14th dimension input feature.
