Crystalline grain microstructure prediction method based on multi-scale fusion data model

Abstract

The invention provides a grain microstructure prediction method based on a multi-scale fusion data model, and the method comprises the steps: constructing a SiC bulk phase structure model according to experimental data, and calculating the surface energy data of each surface structure; constructing a phase field model for grain orientation evolution in the SiC vapor deposition process; simulating the morphology evolution process of the SiC coating under different growth conditions to obtain a simulation result; thermodynamically calculating reaction substances based on a Gibbs free energy minimum principle to obtain a volume fraction of a reaction intermediate substance; constructing a database based on the experimental result, the simulation result and the volume fraction of the reaction intermediate substance, and interpolating feature missing values in the database; building a prediction model, and training the prediction model by using the database to obtain a trained prediction model; and predicting the growth process of the grain microstructure by using the prediction model. The technical problems that a large amount of experimental data is needed, the cost is high and the efficiency is low in the prior art are solved.

Description

Technical Field

[0001] This invention relates to grain micromorphology prediction technology, specifically to a grain micromorphology prediction method based on a multi-scale fusion data model. Background Technology

[0002] Silicon carbide (SiC), as a new generation of wide-bandgap semiconductor material, has been widely used in wear-resistant materials, electrothermal materials, armor protection materials, high-temperature and high-strength composite materials, and power device materials due to its excellent properties such as high strength, high hardness, corrosion resistance, wide tunable bandgap, and low density. Meanwhile, ceramic matrix composites (CMCs) have attracted much attention, especially in high-temperature applications, due to their high strength, high modulus, low density, high temperature resistance, and wear and corrosion resistance. Whether it's a SiC epitaxial layer or a SiC coating on the surface of a ceramic matrix composite, the performance of the material is highly dependent on its microstructure and surface quality.

[0003] Chemical vapor deposition (CVD) is the preferred method for preparing high-quality SiC epitaxial layers and coatings. CVD technology introduces precursor compounds containing desired elements into the gas phase, causing a chemical reaction on the substrate or fiber surface to generate uniform and high-purity SiC thin films or coatings. Compared to other inorganic material preparation methods, CVD technology has the following advantages: 1. High quality and purity: It can prepare high-purity SiC materials with low defect density; 2. Adaptability to complex shapes: Interface deposition of complex-shaped components can be achieved through controlled process parameters; 3. Controllable composition and structure: The chemical composition and material distribution of the coating can be precisely controlled.

[0004] Precise control of the microstructure is crucial for achieving high-performance SiC epitaxial layers and coatings. The grain size and orientation of SiC significantly influence its mechanical properties and thermal conductivity. During CVD deposition, grain growth is affected by parameters such as temperature, pressure, and precursor concentration. Achieving precise control of grain size and orientation by adjusting these parameters is key to improving SiC material performance; however, related technologies require extensive experimental data and suffer from high costs and low efficiency. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a method for predicting grain micromorphology based on a multi-scale fusion data model. This method enables rapid and accurate prediction of micromorphology under different process and reactor conditions using only a small number of experimental results, and allows for rapid and low-cost optimization of process parameters in CVD-SiC manufacturing.

[0006] To achieve the above-mentioned technical objectives, in a first aspect, the present invention provides a method for predicting grain microstructure based on a multi-scale fusion data model, comprising the following steps:

[0007] A bulk structure model of SiC was constructed based on experimental data. The bulk structure model was then cross-sectionalized, and the surface energy data of each surface structure were calculated.

[0008] Based on the surface energy data of each surface structure, a phase field model is constructed for the grain orientation evolution during the SiC vapor deposition process;

[0009] The morphological evolution of SiC coatings under different growth conditions was simulated using a phase-field model, and simulation results were obtained.

[0010] Obtain literature data and experimental results under real conditions. Based on the experimental results, simulation results and literature data, perform thermodynamic calculations on the reactants according to the principle of minimum Gibbs free energy to obtain the volume fraction of intermediate substances in the reaction.

[0011] A database was constructed based on experimental results, simulation results, and the volume fraction of reaction intermediates, and machine learning algorithms were used to imput the missing feature values ​​in the database.

[0012] A predictive model is built by using machine learning algorithms, and the model is trained using a database to obtain the trained predictive model.

[0013] Predictive models are used to predict the growth process of grain microstructure.

[0014] Compared with the prior art, the beneficial effects of the present invention include:

[0015] 1. It can decouple and quantify the key synergistic effects of SiC grain formation during chemical vapor deposition (CVD). It not only achieves excellent prediction accuracy of grain size, but also quantitatively identifies the synergistic effects between temperature and precursor kinetics.

[0016] 2. A robust and transferable paradigm was established, demonstrating how to combine multiphysics simulation with interpretable machine learning to accelerate the decoding of complex process-structure relationships.

[0017] 3. It enables rapid and accurate prediction of microstructure under different process and reactor conditions using only a small number of experimental results, and allows for rapid and low-cost optimization of process parameters for CVD-SiC manufacturing.

[0018] According to some embodiments of the present invention, the bulk structure model is cross-sectionalized, and the surface energy data of each surface structure is calculated, including the following steps:

[0019] The bulk structure model is cross-sectioned to obtain crystal surface models with different orientations, covering all potential grain orientations that the phase field model needs to simulate;

[0020] Surface energy data for various surface structures at different temperatures were calculated using LAMMPS software.

[0021] According to some embodiments of the present invention, a phase-field model is constructed based on the surface energy data of each surface structure to address the grain orientation evolution during the SiC vapor deposition process, including the following steps:

[0022] Constructing the phase-field model includes building the free energy functional, the phase-field evolution equation, and the concentration field equation:

[0023] The free energy functional incorporates gradient coefficients, double-well barrier height, and grain free energy density parameters.

[0024] The phase field evolution equation takes into account the phase field mobility and thermodynamic driving force.

[0025] The concentration field equation relates the total concentration of a single component to the concentrations of each phase.

[0026] According to some embodiments of the present invention, the morphological evolution process of SiC coatings under different growth conditions is simulated based on a phase-field model, including:

[0027] The microstructure and grain orientation evolution during SiC deposition at 900℃, 1200℃ and 1500℃ were simulated respectively.

[0028] According to some embodiments of the present invention, thermodynamic calculations of the reactants are performed based on the principle of minimum Gibbs free energy to obtain the volume fraction of the intermediate substances in the reaction, including the following steps:

[0029] Based on the principle of minimum Gibbs free energy, gas-phase chemical equilibrium calculations were performed on the MTS / H2 system during the CVD process using thermodynamic software to obtain the volume fraction of reaction intermediates, which include SiCl2, C2H2, and CH4.

[0030] According to some embodiments of the present invention, a machine learning algorithm is used to impute missing feature values ​​in a database. The machine learning algorithm includes:

[0031] Chain equation multiple interpolation (MICE), linear regression interpolation, K-nearest neighbor regression interpolation (KNN), and random forest regression interpolation.

[0032] According to some embodiments of the present invention, a prediction model is built using machine learning algorithms, including Ridge regression, Lasso regression, random forest regression, support vector machine regression (SVR), K-nearest neighbor regression, AdaBoost regression, and gradient boosting decision tree (GBDT). The hyperparameters of the prediction model are randomly searched and optimized to achieve prediction of SiC grain size and orientation under different CVD process parameters.

[0033] Secondly, the present invention provides a grain micromorphology prediction system based on a multi-scale fusion data model, comprising:

[0034] The surface energy calculation module constructs a bulk structure model of SiC based on experimental data, performs cross-sectional processing on the bulk structure model, and calculates the surface energy data of each surface structure.

[0035] The phase-field model construction module communicates with the surface energy calculation module and constructs a phase-field model based on the surface energy data of each surface structure for the grain orientation evolution during the SiC vapor deposition process.

[0036] The simulation module communicates with the phase field model construction module to simulate the morphological evolution of SiC coatings under different growth conditions based on the phase field model, and obtains simulation results.

[0037] The volume fraction calculation module communicates with the simulation module to obtain literature data and experimental results under real conditions. Based on the experimental results, simulation results and literature data, it performs thermodynamic calculations on the reactants based on the principle of minimum Gibbs free energy to obtain the volume fraction of the intermediate substances in the reaction.

[0038] The interpolation module constructs a database based on experimental results, simulation results, and the volume fraction of intermediate substances in the reaction, and uses machine learning algorithms to interpolate the missing feature values ​​in the database.

[0039] The model training module builds a prediction model using machine learning algorithms and trains the prediction model using a database to obtain a trained prediction model.

[0040] The prediction module uses a prediction model to predict the growth process of grain microstructure.

[0041] Thirdly, the present invention provides a computer-readable storage medium storing computer-executable instructions for causing a computer to perform the grain micromorphology prediction method based on a multi-scale fusion data model as described in any one of the first aspects.

[0042] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0043] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, wherein the abstract drawings are to be completely consistent with one of the drawings in the specification:

[0044] Figure 1 A flowchart illustrating a method for predicting grain microstructure based on a multi-scale fusion data model, provided in an embodiment of the present invention;

[0045] Figure 2 Surface model diagrams of different crystal planes of SiC;

[0046] Figure 3 The figure shows the simulation results of the microstructure and grain orientation evolution during the SiC deposition process at 1500℃.

[0047] Figure 4 The figure shows the simulation results of the microstructure and grain orientation evolution during the SiC deposition process at 1200℃.

[0048] Figure 5 The figure shows the simulation results of the microstructure and grain orientation evolution during the SiC deposition process at 900℃;

[0049] Figure 6 The graph shows the change in the relative concentration of the intermediate substance with temperature when the pressure P = 2000 Pa and the H2 / MTS ratio α = 5.

[0050] Figure 7 A plot showing the R² values ​​of the model trained before adding simulation data;

[0051] Figure 8 A graph showing the fit and error between experimental and predicted data before adding simulation data;

[0052] Figure 9 A graph showing the R² values ​​of the trained model after incorporating simulation data;

[0053] Figure 10 This is a graph showing the fit between the experimental data and the predicted data after adding simulation data, along with the error. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0055] It should be noted that although functional modules are divided in the system diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the system or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0056] Reference Figures 1 to 10 , Figure 1 A flowchart illustrating a method for predicting grain microstructure based on a multi-scale fusion data model, provided in an embodiment of the present invention; Figure 2 Surface model diagrams of different crystal planes of SiC; Figure 3 The figure shows the simulation results of the microstructure and grain orientation evolution during the SiC deposition process at 1500℃. Figure 4 The figure shows the simulation results of the microstructure and grain orientation evolution during the SiC deposition process at 1200℃. Figure 5 The figure shows the simulation results of the microstructure and grain orientation evolution during the SiC deposition process at 900℃; Figure 6 The graph shows the change in the relative concentration of the intermediate substance with temperature when the pressure P = 2000 Pa and the H2 / MTS ratio α = 5. Figure 7 A plot showing the R² values ​​of the model trained before adding simulation data; Figure 8 A graph showing the fit and error between experimental and predicted data before adding simulation data; Figure 9 A graph showing the R² values ​​of the trained model after incorporating simulation data; Figure 10 This is a graph showing the fit between the experimental data and the predicted data after adding simulation data, along with the error.

[0057] In one embodiment, the grain micromorphology prediction method based on a multi-scale fusion data model includes the following steps: constructing a bulk structure model of SiC based on experimental data, performing cross-sectional processing on the bulk structure model, and calculating the surface energy data of each surface structure; constructing a phase field model based on the surface energy data of each surface structure for the grain orientation evolution during the SiC vapor deposition process; simulating the morphology evolution process of the SiC coating under different growth conditions according to the phase field model to obtain simulation results; acquiring literature data and experimental results under real conditions, and performing thermodynamic calculations of the reactants based on the principle of minimum Gibbs free energy according to the experimental results, simulation results, and literature data to obtain the volume fraction of the intermediate reactants; constructing a database based on the experimental results, simulation results, and volume fraction of the intermediate reactants, and imputing the missing feature values ​​in the database using a machine learning algorithm; building a prediction model using a machine learning algorithm, training the prediction model using the database to obtain the trained prediction model; and using the prediction model to predict the grain micromorphology growth process.

[0058] The specific steps are as follows:

[0059] Step S1: Construct a bulk structure model of SiC based on relevant experimental data. Then, perform cross-sectional processing on the bulk structure model to obtain surface models of different crystal planes of SiC (10-layer atomic structure). Calculate the surface energy of each surface structure using molecular dynamics (MD). For example... Figure 2 The image shows the surface structure models of different crystal planes of 3C-SiC.

[0060] Table 1. Surface energy of different 3C-SiC surfaces calculated based on LAMMPS (unit: J / m²) 2 )

[0061]

[0062]

[0063] Step S2 involves constructing a phase-field model for the grain orientation evolution during SiC vapor deposition. The free energy functional of the system is as follows:

[0064]

[0065] In the above formula, a αβ and W αβ These are the gradient coefficient between the α and β phases (grains) and the height of the double well potential, respectively. c Φ represents the grain free energy density. α As a field variable, Φ within the α grain α =1, Φ inside other grains α =0, at the grain boundary 0<Φ α Since <1, the following relationship holds:

[0066]

[0067] Then, a step function σ is defined. α When 0 < Φ < 1, σ α =1, σ in other cases α =0. Thus, this invention can define the current phase sequence number n:

[0068]

[0069] Therefore, in subsequent formulas, n will be used instead of N. In order to introduce anisotropy into the system, the gradient coefficients are defined as follows:

[0070]

[0071] In the above formula, It is a quantitative relationship related to grain boundary thickness and grain boundary energy, where θ0 is the crystal orientation of each grain, u is the intensity of anisotropy, and ζ is the anisotropic mode.

[0072] Therefore, based on the above description, the evolution equation of the phase field over time can be obtained as follows.

[0073] In the above formula, i, j, and k all represent the grain number at each point. The last term on the right-hand side of the formula is the thermodynamic driving force of the system. W ij , and The reference values ​​for barrier height, gradient coefficient, and phase mobility are respectively given, and their relationship is as follows.

[0074]

[0075] In the above formula, δ is the grain boundary thickness, and γ is... ij With M ij Let represent the energy and mobility of the grain boundary between grain i and grain j.

[0076] Furthermore, the concentration field equation is described as follows. The system contains only one component, whose total concentration is the sum of the concentrations in all phases. Here, b represents the total concentration of this component, and bi represents the concentration of this component in grain i.

[0077]

[0078] Therefore, the diffusion equation for this single component is as follows.

[0079]

[0080] In the above formula, D i Let be the diffusion coefficient of component i in phase i, and its variation with temperature follows the form of the Arrhenius equation. Therefore, by solving the above phase field governing equation and concentration field governing equation, the structural evolution of the system can be simulated.

[0081] Step S3: Simulate the morphological evolution of SiC coatings under different growth conditions.

[0082] Figure 3 The figures show the simulation results of microstructure and grain orientation evolution during SiC deposition at 1500℃. As can be seen from the figures, randomly distributed grains gradually grow outwards from the substrate and gradually differentiate, with the dominant grain orientation being (111). The grains with the (110) crystal plane and the later (111) oriented grains became dominant. The grains on the (110) crystal plane were gradually suppressed.

[0083] Figure 4 The results show the simulation of microstructure and grain orientation evolution during SiC deposition at 1200℃. As can be seen from the figure, in the initial stage, the grains of each phase are uniformly and randomly distributed (time t1). With increasing time, the grains of different orientations gradually differentiate on the surface, with grains having lower surface energy gaining an advantage in the competitive growth process. The dominant grains include (111) and... The grain size of the crystal plane, (100) (311) and Oriented grains are suppressed in the initial stage. (110) Oriented grains are also suppressed in the later stage. At high temperature, the gas phase chemical reaction is accelerated, the supersaturation of the reactants on the substrate surface is low, and the adsorbed atoms have enough energy and time to migrate on the substrate surface, thus forming a hemispherical cauliflower-like structure.

[0084] Figure 5 The figures show the simulation results of microstructure and grain orientation evolution during SiC deposition at 900℃. As can be seen from the figures, surface nucleation intensifies at this temperature, and even grains with a competitive growth potential fail to gain growth opportunities during the competitive growth process. At low temperatures, the deposition process is controlled by surface chemical kinetics. The supersaturation of adsorbed substances on the substrate surface is high, allowing adsorbed molecules to continuously form new nuclei. Furthermore, at low temperatures, adsorbed molecules have low energy and poor migration ability, resulting in less molecular exchange between nuclei and the growth of spherical particles in all directions. At high temperatures, the gas-phase chemical reaction accelerates, the supersaturation of reactants on the substrate surface is low, the nucleation rate decreases, and adsorbed atoms have sufficient energy and time to migrate on the substrate surface, thus forming larger grains.

[0085] Step S4: Perform thermodynamic calculations on the reactants based on the principle of minimum Gibbs free energy to obtain the volume fraction of the intermediate substances in the reaction.

[0086] Based on the principle of Gibbs free energy minimization, this invention utilizes independently developed thermodynamic software to systematically calculate the gas-phase chemical equilibrium of the MTS / H2 system under typical chemical vapor deposition (CVD) conditions. This step transforms the complex and difficult-to-measure chemical environment into a series of quantitative physical descriptors, providing important evidence for a deeper understanding of the CVD process. Through macroscopic process parameters (such as temperature, pressure, and H2 / MTS ratio), this invention calculates a series of variables reflecting ideal supersaturation conditions, including the equilibrium partial pressures of key growth-active species (such as SiCl2, C2H2, and CH4). The calculated concentrations of these gas-phase species are then integrated as additional input features into the final grain size prediction model. This method injects prior physicochemical knowledge into the machine learning model, ensuring a robust expression of the thermodynamic foundation within the model.

[0087] Figure 6 The study presents the trends in the concentrations of key gaseous components with temperature under typical process conditions (pressure P = 2000 Pa, H2 / MTS ratio α = 5). These trends reveal the chemical driving forces behind the surface reactions, particularly demonstrating the evolution of gas-phase supersaturation with temperature.

[0088] Step S5: Combine the experimental results with the data obtained in steps S3 and S4 to construct a database.

[0089] Supersaturation is a key factor determining the nucleation rate and grain growth kinetics during crystal growth, and its mathematical expression is as follows:

[0090]

[0091] Where C represents the solute concentration, C sat Indicates saturation concentration.

[0092] To improve the accuracy of grain size prediction, the concentrations of key intermediate species (such as SiCl3 and C2H2) obtained through thermodynamic calculations are incorporated into the input parameters of the machine learning model. This method enables the model to autonomously capture the complex relationship between supersaturation and grain size.

[0093] Step S6: Use machine learning algorithms to impute missing feature values ​​in the database.

[0094] Multiple imputation methods were employed to handle missing values, including chain equation multiple imputation (MICE), linear regression imputation, K-nearest neighbor regression imputation (KNN), and random forest regression imputation.

[0095] MICE is a Bayesian-based method that imputes missing values ​​using a series of iterative prediction models. In this method, each iteration uses the remaining variables in the dataset to estimate the variable for the missing value until convergence.

[0096] Linear regression imputation handles missing values ​​by creating a linear combination of features, aiming to minimize the error between the imputed value and the actual value by finding the optimal straight line or plane / hyperplane.

[0097] KNN regression imputation calculates the distance between the sample in the descriptor hyperspace and the training data, and determines the target value as the average of the k nearest neighbors.

[0098] Random forest regression imputation constructs multiple decision trees, which are then aggregated using a weighted average to determine the final target value. It's important to note that regression-based imputation methods are only effective when a complete feature dataset is available.

[0099] Step S7: Build a prediction model using machine learning algorithms.

[0100] In addition to the regression methods mentioned earlier—linear regression, KNN regression, and random forest regression—this study also employed several other commonly used regression algorithms.

[0101] One approach is Support Vector Machine Regression (SVR), which builds a minimally complex model by introducing an ε-insensitive loss function. This method ensures that most prediction errors fall within the ε range, focusing only on errors exceeding this threshold.

[0102] Ensemble learning methods such as AdaBoost and Gradient Boosting Decision Tree (GBDT) are widely used to improve the accuracy of regression tasks.

[0103] AdaBoost adjusts sample weights to make subsequent models focus more on samples that are difficult to predict, ultimately combining multiple weak learners into a strong learner.

[0104] In contrast, GBDT employs a gradient boosting strategy, where decision trees progressively fit the residuals of previous models, and the results of all trees are weighted and summed to obtain the final prediction. Both methods aim to improve regression accuracy by enhancing weak learners, thereby generating more powerful models.

[0105] The selection of hyperparameters, including maximum depth, learning rate, and network architecture, directly affects model performance. Hyperparameter optimization helps identify the optimal configuration that maximizes the model's learning potential. Systematic search and tuning of hyperparameters can effectively alleviate overfitting and underfitting, thereby enhancing the model's generalization ability and improving its predictive accuracy. In this study, hyperparameter optimization was performed through random search, i.e., constructing the model based on randomly sampled combinations of hyperparameters in the hyperparameter space. The model used for training and its hyperparameters are shown in Table 2.

[0106] Table 2 shows the models used for training and their hyperparameters.

[0107]

[0108]

[0109] Figure 7 The performance of different interpolation methods (MICE, KNN, linear, and random) across various training methods was compared. MICE interpolation performed well in most training methods, followed by random interpolation. Linear and KNN interpolation methods performed relatively poorly, with linear interpolation showing the worst performance. The best-performing model was the GBDT model using MICE interpolation, with an R² value of [missing value]. 2 The value was 0.863. To further test the accuracy of the model, the grain size was predicted under the same parameters and fitted with the experimental values. The fitting results are as follows: Figure 8 As shown, R 2The value is 0.8134, and the MAPE is 24.10%. The model fits the data well overall, with R... 2 The values ​​indicate this. However, the relatively high MAPE suggests that while the model captures the overall trend, there is still room for improvement in its predictive accuracy.

[0110] In addition to the original macroscopic parameters, this invention also incorporates the concentration percentage of key gaseous intermediates into the input parameters. This approach is based on two key considerations: First, incorporating the content of intermediate species can more accurately reflect the impact of supersaturation on grain size. Second, the multi-scale modeling framework, by integrating macroscopic and microscopic data features, significantly improves the model's expressive power, generalization ability, and robustness, while making decisions more accurate and enhancing adaptability under dynamically changing conditions. The performance of the trained model is as follows... Figure 9 As shown.

[0111] After introducing simulated data, the overall performance of the machine learning model significantly improved. The enhanced dataset provided more comprehensive and representative samples, enabling the model to capture complex relationships within the data and generalize better to unseen information. The GBDT model using random forest imputation performed best, providing robust and reliable results across various training techniques. Its fit to the experimental data is as follows: Figure 10 As shown, R 2 The value increased to 0.9658, while MAPE decreased to 9.02%. The performance of the machine learning model significantly improved after incorporating phase-field simulation data and thermodynamic calculation results. Phase-field simulation data not only enriched the database but also provided high-quality data for imputation of missing values, while the addition of key intermediate species concentrations obtained from thermodynamic calculations enhanced the model's ability to capture the relationship between supersaturation and grain size at the microscopic level. Furthermore, both phase-field simulation and thermodynamic calculations are based on physical principles, ensuring the physical consistency of the generated data. This consistency helps the machine learning model better understand the underlying physical mechanisms, thereby improving prediction accuracy and providing deeper insights.

[0112] The above prediction results demonstrate that by establishing a CVD deposition phase-field model, expanding the data, and combining it with machine learning algorithms, CVD grain size can be accurately predicted. This invention is also an effective method for CVD processes in other multi-component systems.

[0113] Traditional methods for controlling SiC grain size and orientation require extensive experimental data to determine suitable parameters, resulting in high costs. This patent, however, simulates the morphological evolution of SiC coatings under different growth conditions using a phase-field model. Furthermore, it utilizes machine learning algorithms to build a database based on existing experimental results, simulation results, and literature data, and trains a predictive model. This eliminates the need for numerous repetitive experiments to explore the effects of different parameters on grains. For example, the phase-field model can simulate grain growth under varying parameters such as temperature, pressure, and precursor concentration in a virtual environment, significantly reducing the number of actual experiments and thus effectively lowering costs.

[0114] Previously, relying on numerous experiments to attempt parameter control was extremely time-consuming. This invention, utilizing computer simulations (such as phase-field model simulations) and machine learning algorithms to build and train predictive models, can quickly analyze the effects of different parameter combinations on grain size and orientation, obtaining multiple sets of corresponding results in a short time. Compared to simply relying on numerous experiments to try each parameter individually, this significantly improves the efficiency of research and control. During CVD deposition, grain growth is affected by multiple parameters such as temperature, pressure, and precursor concentration. Traditional methods struggle to comprehensively and accurately grasp how these parameters interact to influence grain size and orientation. This invention integrates experimental data, simulation results, and literature data, constructing a database based on multi-scale information. Furthermore, by using machine learning algorithms to mine the inherent patterns in the data, it fully considers the complex interrelationships between parameters and their combined impact on grain size and orientation. This allows for more precise determination of how to control these parameters, achieving accurate control over the grain size and orientation of SiC materials, which helps improve key performance indicators such as the mechanical properties and thermal conductivity of SiC materials.

[0115] This invention utilizes the principle of minimum Gibbs free energy to perform thermodynamic calculations on the reactants, obtaining the volume fraction of intermediate substances in the reaction. This allows for a deeper understanding of the material changes during the deposition process from a thermodynamic perspective. Furthermore, by combining phase-field models to simulate the evolution of grain orientation, this invention provides a solid theoretical and model basis for controlling the microstructure of grains through parameter adjustment. This makes parameter adjustment no longer blind, but rather based on scientific methods and precise directions, which is more conducive to achieving accurate control.

[0116] The bulk structure model is cross-sectioned to calculate the surface energy data of each surface structure. The steps include: cross-sectioning the bulk structure model to obtain crystal surface models with different orientations, covering all potential grain orientations that the phase field model needs to simulate; and calculating the surface energy data of each surface structure at different temperatures based on LAMMPS software.

[0117] Based on the surface energy data of each surface structure, a phase field model is constructed for the grain orientation evolution during the SiC vapor deposition process. The steps include: constructing the phase field model includes constructing a free energy functional, a phase field evolution equation, and a concentration field equation; the free energy functional introduces gradient coefficients, double-well barrier height, and grain free energy density parameters; the phase field evolution equation considers phase field mobility and thermodynamic driving force; and the concentration field equation relates the total concentration of a single component to the concentration of each phase.

[0118] The morphological evolution of SiC coatings under different growth conditions was simulated using a phase-field model, including the microstructure and grain orientation evolution during SiC deposition at 900℃, 1200℃, and 1500℃.

[0119] Thermodynamic calculations of the reactants based on the principle of minimum Gibbs free energy were performed to obtain the volume fraction of the intermediate substances in the reaction. The steps included: Based on the principle of minimum Gibbs free energy, gas-phase chemical equilibrium calculations were performed on the MTS / H2 system in the CVD process using thermodynamic software to obtain the volume fraction of the intermediate substances in the reaction. The intermediate substances in the reaction included SiCl2, C2H2 and CH4.

[0120] Machine learning algorithms are used to impute missing feature values ​​in a database. These algorithms include chain equation multiple imputation (MICE), linear regression imputation, K-nearest neighbor regression imputation (KNN), and random forest regression imputation.

[0121] A predictive model is built using machine learning algorithms, including Ridge regression, Lasso regression, random forest regression, support vector machine regression (SVR), K-nearest neighbor regression, AdaBoost regression, and gradient boosting decision tree (GBDT). The hyperparameters of the predictive model are then optimized through random search to achieve prediction of SiC grain size and orientation under different CVD process parameters.

[0122] In one embodiment, the grain micromorphology prediction system based on a multi-scale fusion data model includes: a surface energy calculation module, which constructs a bulk structure model of SiC based on experimental data, performs cross-sectional processing on the bulk structure model, and calculates the surface energy data of each surface structure; a phase field model construction module, which is communicatively connected to the surface energy calculation module, and constructs a phase field model for the grain orientation evolution during the SiC vapor deposition process based on the surface energy data of each surface structure; a simulation module, which is communicatively connected to the phase field model construction module, and simulates the morphology evolution process of the SiC coating under different growth conditions according to the phase field model to obtain simulation results; and a volume fraction calculation module, which is communicatively connected to the phase field model construction module. The module establishes a communication connection to acquire literature data and experimental results under real-world conditions. Based on the experimental results, simulation results, and literature data, it performs thermodynamic calculations of the reactants using the principle of minimum Gibbs free energy to obtain the volume fraction of intermediate substances. The interpolation module constructs a database based on experimental results, simulation results, and the volume fraction of intermediate substances, and uses machine learning algorithms to interpolate missing feature values ​​in the database. The model training module builds a prediction model using machine learning algorithms and trains the prediction model using the database to obtain the trained prediction model. The prediction module uses the prediction model to predict the grain microstructure growth process.

[0123] This invention is based on real experimental data, using the input of the surface energy calculation module and the experimental results of the volume fraction calculation module to ensure that the model conforms to actual physical laws and avoids the problem of building castles in the air by purely theoretical modeling. This invention supplements with phase field simulation data, which can quickly generate a large number of morphological evolution results under different growth conditions, making up for the shortcomings of experimental data, such as high cost, long cycle and difficulty in covering multiple parameter combinations. This invention introduces literature data, integrates the research results verified in the field, reduces repeated experiments, and enriches the data dimensions to improve the representativeness of the data.

[0124] The volume fraction calculation module, based on the principle of minimum Gibbs free energy, transforms experimental, simulation, and literature data into the volume fraction of intermediate substances in the reaction. This key data not only reveals the essential laws of material transformation during SiC vapor deposition but also provides thermodynamic feature dimensions for the subsequent database. This upgrades the database from a simple morphology / parameter record to a multi-dimensional dataset containing reaction mechanisms, providing more fundamental feature support for prediction models.

[0125] The system incorporates machine learning algorithms into its imputation and model training modules, specifically addressing the core pain points of traditional research, such as incomplete data and difficulty in quantifying complex relationships among multiple parameters. Experimental or literature data often suffer from feature loss due to experimental errors, detection limits, or incomplete literature reports; traditional methods often directly remove missing data, leading to data waste. The imputation module uses machine learning algorithms to fill in missing values ​​based on potential correlations between data points. This preserves more data samples while avoiding insufficient sample size and model training bias caused by data removal, providing a high-quality dataset for subsequent prediction models.

[0126] The surface energy calculation module starts with the bulk structure (atomic arrangement) of SiC, calculating surface energy through facet processing—surface energy is a core microscopic parameter determining the grain growth direction (orientation) (low surface energy crystal planes are more likely to become growth termination surfaces, thus determining grain orientation); the phase field model construction module simulates grain orientation evolution based on surface energy data, essentially explaining grain growth laws from the microscopic mechanism of interatomic interactions. This modeling logic from microscopic mechanism to mesoscopic evolution ensures that subsequent predictions are not a black box of pure data fitting, but rather scientific predictions supported by clear physical meaning, avoiding erroneous results with high data fit but contradicting actual physical laws.

[0127] The system ultimately outputs the grain microstructure growth process through a prediction module. However, this prediction result can be inversely correlated with macroscopic growth parameters and microscopic mechanisms. For example, if the prediction finds that the grain size is too small at a certain temperature, the direction of parameter adjustment can be clarified by tracing back the surface energy calculation results or volume fraction calculation results. This reverse derivation capability of prediction result → mechanism analysis → parameter optimization makes the system not only a prediction tool, but also a guiding tool for SiC process optimization, directly serving the engineering goal of achieving precise control of grain size and orientation through parameter adjustment.

[0128] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0129] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0130] Furthermore, one embodiment of the present invention provides a computer-readable storage medium storing computer-executable instructions that are executed by a processor or controller, for example, by a processor in the above-described terminal embodiment, enabling the processor to execute the grain micromorphology prediction method based on a multi-scale fusion data model in the above-described embodiment.

[0131] Those skilled in the art will understand that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0132] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

[0133] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting grain microstructure based on a multi-scale fusion data model, characterized in that, Including the following steps: A bulk structure model of SiC was constructed based on experimental data. The bulk structure model was then cross-sectionalized, and the surface energy data of each surface structure were calculated. Based on the surface energy data of each surface structure, a phase field model is constructed for the grain orientation evolution during the SiC vapor deposition process; The morphological evolution of SiC coatings under different growth conditions was simulated using a phase-field model, and simulation results were obtained. Obtain literature data and experimental results under real conditions. Based on the experimental results, simulation results and literature data, perform thermodynamic calculations on the reactants according to the principle of minimum Gibbs free energy to obtain the volume fraction of intermediate substances in the reaction. A database was constructed based on experimental results, simulation results, and the volume fraction of reaction intermediates, and machine learning algorithms were used to imput the missing feature values ​​in the database. A predictive model is built by using machine learning algorithms, and the model is trained using a database to obtain the trained predictive model. Predictive models are used to predict the growth process of grain microstructure.

2. The method for predicting grain microstructure based on a multi-scale fusion data model according to claim 1, characterized in that, The bulk structure model is cross-sectionalized, and the surface energy data of each surface structure is calculated, including the following steps: The bulk structure model is cross-sectioned to obtain crystal surface models with different orientations, covering all potential grain orientations that the phase field model needs to simulate; Surface energy data for various surface structures at different temperatures were calculated using LAMMPS software.

3. The method for predicting grain microstructure based on a multi-scale fusion data model according to claim 1, characterized in that, Based on surface energy data of various surface structures, a phase-field model is constructed for the grain orientation evolution during SiC vapor deposition, including the following steps: Constructing the phase-field model includes building the free energy functional, the phase-field evolution equation, and the concentration field equation: The free energy functional incorporates gradient coefficients, double-well barrier height, and grain free energy density parameters. The phase field evolution equation takes into account the phase field mobility and thermodynamic driving force. The concentration field equation relates the total concentration of a single component to the concentrations of each phase.

4. The method for predicting grain microstructure based on a multi-scale fusion data model according to claim 1, characterized in that, The morphological evolution of SiC coatings under different growth conditions was simulated using a phase-field model, including: The microstructure and grain orientation evolution during SiC deposition at 900℃, 1200℃ and 1500℃ were simulated respectively.

5. The method for predicting grain microstructure based on a multi-scale fusion data model according to claim 1, characterized in that, Thermodynamic calculations of the reactants are performed based on the principle of minimum Gibbs free energy to obtain the volume fraction of intermediate substances in the reaction, including the following steps: Based on the principle of minimum Gibbs free energy, gas-phase chemical equilibrium calculations were performed on the MTS / H2 system during the CVD process using thermodynamic software to obtain the volume fraction of reaction intermediates, which include SiCl2, C2H2, and CH4.

6. The method for predicting grain microstructure based on a multi-scale fusion data model according to claim 1, characterized in that, Machine learning algorithms are used to impute missing feature values ​​in a database. These algorithms include: Chain equation multiple interpolation (MICE), linear regression interpolation, K-nearest neighbor regression interpolation (KNN), and random forest regression interpolation.

7. The method for predicting grain microstructure based on a multi-scale fusion data model according to claim 1, characterized in that, A predictive model is built using machine learning algorithms, including Ridge regression, Lasso regression, random forest regression, support vector machine regression (SVR), K-nearest neighbor regression, AdaBoost regression, and gradient boosting decision tree (GBDT). The hyperparameters of the predictive model are then optimized by random search to predict the SiC grain size and orientation under different CVD process parameters.

8. A grain microstructure prediction system based on a multi-scale fusion data model, characterized in that, include: The surface energy calculation module constructs a bulk structure model of SiC based on experimental data, performs cross-sectional processing on the bulk structure model, and calculates the surface energy data of each surface structure. The phase-field model construction module communicates with the surface energy calculation module and constructs a phase-field model based on the surface energy data of each surface structure for the grain orientation evolution during the SiC vapor deposition process. The simulation module communicates with the phase field model construction module to simulate the morphological evolution of SiC coatings under different growth conditions based on the phase field model, and obtains simulation results. The volume fraction calculation module communicates with the simulation module to obtain literature data and experimental results under real conditions. Based on the experimental results, simulation results and literature data, it performs thermodynamic calculations on the reactants based on the principle of minimum Gibbs free energy to obtain the volume fraction of the intermediate substances in the reaction. The interpolation module constructs a database based on experimental results, simulation results, and the volume fraction of intermediate substances in the reaction, and uses machine learning algorithms to interpolate the missing feature values ​​in the database. The model training module builds a prediction model using machine learning algorithms and trains the prediction model using a database to obtain a trained prediction model. The prediction module uses a prediction model to predict the growth process of grain microstructure.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform the grain micromorphology prediction method based on a multi-scale fusion data model as described in any one of claims 1 to 7.

Images