Metal matrix composite material performance prediction method based on transfer learning

Abstract

The invention discloses a metal-based composite material performance prediction method based on transfer learning. The training method comprises the following steps: providing an alloy data set and a metal-based composite material data set; pre-training the initial model by adopting the alloy data set to obtain a pre-trained model; the pre-training model is trained through the metal-based composite material data set by means of transfer learning, a metal-based composite material performance prediction model is obtained, and weight parameters of neurons related to the components of the basic alloy and the preparation process in the pre-training model are fixed. According to the technical scheme provided by the invention, a transfer learning framework is adopted, the relationship among alloy matrix components, a heat treatment process and strong correlation performance is more fully constructed, accurate prediction of the performance of the metal-based composite material can be realized under the condition of less data volume, and research, development and design of the novel high-performance metal-based composite material are remarkably accelerated.

Description

Technical Field

[0001] This invention relates to the field of materials science and technology, and in particular to a method for predicting the performance of metal matrix composites based on transfer learning. Background Technology

[0002] Metal matrix composites typically possess high specific strength, dimensional stability, and excellent heat resistance, making them key materials for lightweight and high-performance development in fields such as aerospace and marine engineering. Introducing reinforcing phases such as ceramic particles, carbon nanotubes, and whiskers into an alloy matrix to form heterogeneous material systems holds promise for overcoming the bottleneck of synergistic control of key properties such as strength, plasticity, thermal conductivity, and electrical conductivity in traditional alloys, thus meeting the stringent requirements of extreme service environments. However, the research and development of novel high-performance metal matrix composites requires precise control of the complex coupling effects between material properties and matrix composition and preparation process parameters. This often necessitates time-consuming and costly experimental trial and error, hindering the efficient iterative development of novel high-performance metal matrix composites.

[0003] In recent years, the rapid development of machine learning has opened up new avenues for materials research, demonstrating enormous potential in material composition control and process parameter optimization. Patent application number 202410934256.3 discloses a method for predicting the mechanical properties of A356 aluminum alloy using machine learning, providing highly accurate prediction models for tensile strength and elongation. However, this model is only applicable to performance prediction of metal systems with sufficient data and standardized processes, exhibiting poor generalizability for material systems with scarce data. Patent application number 202310651663.9 discloses a method for predicting the properties of metallic materials using transfer learning, enabling reverse design of novel multi-component metallic materials even with limited data. However, metal matrix composites differ from metallic materials, being significantly influenced not only by matrix composition but also closely related to process parameters and interface microstructure. Traditional machine learning methods struggle to achieve multi-objective performance prediction, necessitating the development of suitable performance prediction methods for metal matrix composites. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide a method for predicting the performance of metal matrix composites based on transfer learning.

[0005] To achieve the aforementioned objectives, the technical solution adopted by this invention includes:

[0006] In a first aspect, the present invention provides a training method for a performance prediction model of metal matrix composites based on transfer learning, comprising:

[0007] Provides alloy datasets and metal matrix composite datasets. The alloy datasets include the composition, preparation process, and performance data of various base alloys of a selected metal. The metal matrix composite datasets include the metal composition, preparation process, physicochemical properties of the reinforcing phase, interface properties, and performance data of metal matrix composites formed by combining the selected metal or its alloy with a reinforcing phase.

[0008] The initial model is pre-trained using the alloy dataset to obtain a pre-trained model, which predicts a strong correlation between the composition, preparation process, and performance data of the basic alloy.

[0009] The pre-trained model is trained using the metal matrix composite data set by transfer learning to obtain a metal matrix composite performance prediction model, wherein the weight parameters of neurons related to the composition and preparation process of the base alloy in the pre-trained model are fixed.

[0010] Secondly, the present invention also provides a method for preparing a metal-based composite material, comprising:

[0011] The optimal performance prediction model for metal matrix composites was obtained using the above training method.

[0012] Input the set of relevant parameters of the metal matrix composite material into the performance prediction model of the metal matrix composite material to obtain the set of predicted performance.

[0013] Select the optimal relevant parameters from the predicted performance set;

[0014] Metal matrix composites were prepared based on the optimal correlation parameters.

[0015] Thirdly, the present invention also provides a training system for a performance prediction model of metal matrix composites based on transfer learning, comprising:

[0016] The basic data module is used to provide alloy datasets and metal matrix composite datasets. The alloy datasets include the composition, preparation process, and performance data of various base alloys of the selected metal. The metal matrix composite datasets include the metal composition, preparation process, physicochemical properties of the reinforcing phase, interface properties, and performance data of the metal matrix composites formed by combining the selected metal or its alloy with a reinforcing phase.

[0017] The pre-training module is used to pre-train the initial model using the alloy dataset to obtain a pre-trained model, which predicts the strong correlation between the composition and preparation process of the base alloy and the performance data.

[0018] The transfer learning module is used to train the pre-trained model using the metal matrix composite data set to obtain a metal matrix composite performance prediction model, wherein the weight parameters of neurons in the pre-trained model related to the composition and preparation process of the base alloy are fixed.

[0019] Fourthly, the present invention also provides an application of the above-mentioned training method, specifically a method for predicting the performance of metal matrix composites based on transfer learning, including the process of performing performance prediction of metal matrix composites using the metal matrix composite performance prediction model obtained by the above-mentioned training method.

[0020] Based on the above technical solution, compared with the prior art, the beneficial effects of the present invention include at least the following:

[0021] The technical solution provided by this invention adopts a transfer learning framework that integrates Transformer-based neural networks and fully connected neural networks, which more fully constructs the relationship between alloy matrix composition, heat treatment process and strongly correlated properties. It can achieve accurate prediction of the performance of metal matrix composites with a small amount of data, and significantly accelerate the research and development of new high-performance metal matrix composites.

[0022] On the other hand, the preferred embodiments of the present invention, through dual precision evaluation using test and validation sets, avoid making erroneous predictions under low accuracy conditions, thereby improving the reliability of model predictions in complex problems. This enables the accurate design of metal matrix composites based on given target performance.

[0023] The above description is merely an overview of the technical solution of the present invention. In order to enable those skilled in the art to better understand the technical means of this application and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described below in conjunction with detailed drawings. Attached Figure Description

[0024] Figure 1 A flowchart illustrating a typical embodiment of the present invention for predicting the performance of metal matrix composites based on transfer learning;

[0025] Figure 2a A schematic diagram illustrating the tensile strength prediction accuracy of a metal matrix composite performance prediction model provided in a typical embodiment of the present invention;

[0026] Figure 2b A schematic diagram illustrating the elongation prediction accuracy of a metal matrix composite material performance prediction model provided in a typical embodiment of the present invention.

[0027] Figure 3a A schematic diagram of tensile strength verification of a metal matrix composite performance prediction model provided in a typical embodiment of the present invention;

[0028] Figure 3b A schematic diagram illustrating the elongation verification of a metal matrix composite performance prediction model provided as a typical embodiment of the present invention;

[0029] Figure 4a A tensile strength prediction diagram of 2xxx series aluminum alloy heat treatment process provided as a typical embodiment of the present invention;

[0030] Figure 4b A heat treatment process-elongation prediction diagram for TiB2 / 2xxx series aluminum matrix composites provided as a typical embodiment of the present invention;

[0031] Figure 5a A prediction diagram of the reinforcing phase and tensile strength of 2xxx series aluminum alloys is provided as a typical embodiment of the present invention.

[0032] Figure 5b The image shows the predicted elongation of the reinforcing phase in TiB2 / 2xxx series aluminum matrix composites, provided as a typical embodiment of the present invention. Detailed Implementation

[0033] In view of the shortcomings of the prior art, the inventors of this invention, through long-term research and extensive practice, have proposed the technical solution of this invention. The following will further explain and illustrate this technical solution, its implementation process, and its principles.

[0034] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0035] Embedding data patterns related to the alloy matrix into the performance prediction model through transfer learning is a feasible approach to establishing a high-precision prediction model for the mechanical properties of metal matrix composites with small datasets.

[0036] This invention provides a training method for a performance prediction model of metal matrix composites based on transfer learning, which includes the following steps:

[0037] Provides alloy datasets and metal matrix composite datasets. The alloy datasets include the composition, preparation process, and performance data of various base alloys of a selected metal. The metal matrix composite datasets include the metal composition, preparation process, physicochemical properties of the reinforcing phase, interface properties, and performance data of metal matrix composites formed by combining the selected metal or its alloy with a reinforcing phase.

[0038] The initial model is pre-trained using the alloy dataset to obtain a pre-trained model, which predicts a strong correlation between the composition, preparation process, and performance data of the basic alloy.

[0039] The pre-trained model is trained using the metal matrix composite data set by transfer learning to obtain a metal matrix composite performance prediction model, wherein the weight parameters of neurons related to the composition and preparation process of the base alloy in the pre-trained model are fixed.

[0040] It should be noted that in typical embodiments of the present invention, mechanical properties are usually used as the prediction target, which is typical and representative, such as compressive / tensile strength, toughness, and resistance to deformation. However, this does not mean that the feasible implementation scope of the present invention is limited to mechanical properties. Other properties such as thermal conductivity, electrical properties, acoustic properties, magnetic properties, optical properties, etc., can all be included in the implementation scope of the present invention. Taking mechanical properties as an example, the technical solution of the present invention first predicts the strongly correlated properties of a selected metal or its alloy (e.g., aluminum alloy), such as the correlation between the composition, content, heat treatment process, and other conditions of the aluminum alloy on the tensile strength and elongation of the aluminum alloy. Then, based on the pre-training results, transfer learning is applied to predict the performance of composite materials combining the same selected metal with other reinforcements. Based on a deep understanding of the property laws of the base alloy of the same selected metal, the property laws of the alloy-reinforcing phase are fine-tuned through transfer learning, thereby enabling the rapid training of a prediction model that accurately predicts the relevant properties of metal matrix composites with a small number of samples.

[0041] Regarding specific technical details, in some implementations, the initial model includes a feature fusion layer, a multi-head self-attention layer, and a fully connected layer. The feature fusion layer uses the Transformer algorithm to fuse the composition features of the base alloy to obtain an element feature matrix. The multi-head self-attention layer maps the element feature matrix into a feature vector and obtains the interaction between components through normalization to form an alloy composition vector. The fully connected layer splices the alloy composition vector and the preparation process of the base alloy to output the predicted performance of the base alloy.

[0042] In some implementations, the feature fusion layer fuses the composition content of the base alloy with the number of valence electrons, electronegativity, atomic radius, and atomic mass of the constituent elements, transforming the composition content of the base alloy into an elemental feature matrix containing physical meaning. The multi-head self-attention layer uses a multi-head self-attention mechanism to map the elemental feature matrix into query vectors, bond vectors, and value vectors, and obtains the alloy composition vector by normalizing the interactions between the components of the base alloy through the SoftMax function. In a specific implementation, the alloy composition vector is concatenated with the process parameter input as the input of the fully connected layer.

[0043] In some implementations, the transfer learning process specifically includes:

[0044] Model transfer fixes the weights of neurons related to alloy composition and heat treatment process in the feature fusion layer;

[0045] Output layer reconstruction involves reconstructing the loss function to form a reconstructed loss function, removing non-critical relevant properties, and using the required key properties of the metal matrix composite material as the output layer of the fully connected layer.

[0046] Selective model fine-tuning, based on the metal matrix composite dataset, involves fine-tuning the weights of the fully connected layers while reducing the value of the reconstruction loss function.

[0047] In some implementations, the reconstruction loss function is a weighted average of the variances of the predicted and actual values ​​of multiple performance data.

[0048] In some implementations, the training method further includes:

[0049] Before performing the pre-training and transfer learning, the input parameters are preprocessed, including physical feature embedding and process encoding.

[0050] The preprocessed input parameters are then subjected to feature filtering and standardization.

[0051] The processed alloy dataset and metal matrix composite dataset are divided into training set and test set, respectively. The training set is used for model training, and the test set is used for evaluating the model prediction accuracy.

[0052] In some implementations, the standardization process is expressed as:

[0053]

[0054] Among them, X stdLet X be the input parameter after standardization, μ be the average value of the preprocessed input parameter, and δ be the standard deviation of the preprocessed input parameter.

[0055] In some implementations, during the pre-training and transfer learning processes, the coefficient of determination is used to evaluate the model's prediction accuracy; the coefficient of determination is expressed as:

[0056]

[0057] Among them, R 2 : represents the performance coefficient; y i Represents performance data; Indicates the predicted value; The value represents the average of the performance data; n represents the sample size.

[0058] For a typical example of the above technical solution, please refer to Figure 1 As shown, embodiments of the present invention provide a method for predicting the performance of metal matrix composites based on transfer learning, which is a further application of the above-mentioned training method. The prediction method includes:

[0059] S1: Collect several data on alloy composition, preparation process and strongly correlated properties based on publicly available experimental databases or literature, and construct a pre-training dataset using the aforementioned data;

[0060] S2: Collect several data on the matrix composition, preparation process, physicochemical properties of the reinforcing phase, interfacial properties and key properties of metal matrix composites based on publicly available experimental databases or literature, and construct a metal matrix composite dataset using the aforementioned data.

[0061] S3: Based on the raw data collected in steps S1 and S2, preprocess the data using methods such as physical feature embedding and process coding, and perform feature filtering and standardization of the input parameters. Divide the dataset into training and test sets using K-fold cross-validation or random partitioning methods.

[0062] S4: Based on the alloy dataset obtained in S3, the element feature fusion and multi-head self-attention method of the Transformer algorithm are used to perform matrix transformation on the alloy composition; the coefficient of determination is used as the model performance evaluation index, and the feature extraction of alloy composition and process parameters is completed through model training to obtain a pre-trained model with strong correlation performance of alloy matrix.

[0063] S5: Based on the metal matrix composite dataset and pre-trained model obtained in S3, model construction is achieved through methods such as model transfer, selective model fine-tuning, and grid search. The coefficient of determination is used as the model performance evaluation index. When the prediction accuracy of each key performance indicator is greater than 90%, proceed to the next step; otherwise, continue to optimize hyperparameters and retrain the model.

[0064] S6: Collect independent literature experimental data to construct a validation set, compare the prediction results of the key performance prediction model of metal matrix composites with the experimental values ​​to verify the accuracy of the model; when the prediction deviation is less than the threshold, proceed to the next step; when the prediction deviation is greater than the threshold, include the experimental data in the corresponding dataset and return to step S1.

[0065] S7: Preset the input parameter set for the target metal matrix composite material, and use the prediction results to visualize and filter out the material composition and process parameters with the best performance of the metal matrix composite material.

[0066] S8: Prepare the required metal-based composite material according to the optimal material composition and process parameters, and verify the material properties through experiments.

[0067] In specific implementation, the alloy in step S1 may include one or more metallic materials selected from aluminum alloy, titanium alloy, magnesium alloy, copper alloy, nickel alloy, and steel, but is not limited to these; the preparation process in step S1 may include a standard process flow consisting of some or all of the processes in alloy smelting, homogenization treatment, hot deformation processing, and heat treatment, but is not limited to these; the strongly related properties in step S1 include at least mechanical properties such as tensile strength, elongation at break, and elastic modulus, as well as one or more related combinations of high-temperature creep resistance, thermal conductivity, and electrical conductivity, and is also not limited to these.

[0068] The metal matrix composite material in step S2 generally has an alloy matrix content ranging from 99.0 to 85.0 wt.% and a reinforcing phase content ranging from 1.0 to 15.0 wt.%, preferably from 1.0 to 8.0 wt.%, but not limited to these ranges. All metal matrix composite materials are prepared using similar process flows, and the set of preparation process parameters includes molding process, heat treatment process, and composite process parameters.

[0069] The physical and chemical properties of the reinforcing phase may include, for example, the physical and chemical properties of the reinforcing phase such as modulus and coefficient of thermal expansion obtained through literature collection; the interface properties include the reinforcing effect of the reinforcing phase, interface energy and other related properties obtained through empirical formulas or first-principles calculations.

[0070] The specific implementation method for physical feature embedding and process coding in step S3 can be as follows:

[0071] S3-1: Physical feature embedding. The augmentation phase content of the dataset obtained in S2 is embedded into physical features through the embedding layer and the physical properties of the augmentation phase, and then normalized.

[0072] S3-2: Process coding. The post-processing processes of the datasets obtained from S1 and S2 are classified and coded according to different heat treatment processes and deformation strengthening methods; and the forming and composite processes of the datasets obtained from S2 are classified and coded according to specific reinforcing phase composite methods.

[0073] Step S4, the process of building the pre-trained model, may include the following steps:

[0074] S41: Element feature fusion, which uses an element embedding layer to fuse the alloy composition content with the atomic properties of elements, and transforms the alloy composition content into an element feature matrix containing physical meaning;

[0075] S42: Self-attention mechanism, adopting a multi-head self-attention mechanism, maps the element feature matrix into query vector, key vector, and value vector, and obtains the interaction between the elements of the alloy matrix through normalization by the SoftMax function; the normalized alloy composition vector is concatenated with the process parameter input as the input of the fully connected layer neural network.

[0076] The transfer learning in step S5 can be achieved through the following methods:

[0077] Model transfer: Based on the pre-trained model obtained from S4, the weights of the alloy composition-related neurons and the heat treatment process-related neurons in the Transformer base neural network are fixed.

[0078] Output layer reconstruction involves reconstructing the loss function, removing non-critical, strongly correlated mechanical properties, and using the desired properties as the output layer of the fully connected metal matrix composite layer.

[0079] Selective model fine-tuning was performed using the Adam optimization algorithm on the metal matrix composite dataset obtained from S3. This process reduced the loss function value and fine-tuned the weights of the fully connected layers.

[0080] Regarding the specific threshold setting, in one possible implementation, the prediction deviation threshold in step S6 can be set to 15% of the experimental average value corresponding to each key performance, but it is not limited to the value shown in the example here.

[0081] This invention also provides a method for preparing a metal-based composite material, comprising:

[0082] The optimal performance prediction model for metal matrix composites can be obtained by using the training method provided in any of the above embodiments.

[0083] Input the set of relevant parameters of the metal matrix composite material into the performance prediction model of the metal matrix composite material to obtain the set of predicted performance.

[0084] Select the optimal relevant parameters from the predicted performance set;

[0085] Metal matrix composites were prepared based on the optimal correlation parameters.

[0086] A third aspect of this invention also provides a training system for a performance prediction model of metal matrix composites based on transfer learning, comprising:

[0087] The basic data module is used to provide alloy datasets and metal matrix composite datasets. The alloy datasets include the composition, preparation process, and performance data of various base alloys of the selected metal. The metal matrix composite datasets include the metal composition, preparation process, physicochemical properties of the reinforcing phase, interface properties, and performance data of the metal matrix composites formed by combining the selected metal or its alloy with a reinforcing phase.

[0088] The pre-training module is used to pre-train the initial model using the alloy dataset to obtain a pre-trained model, which predicts the strong correlation between the composition and preparation process of the base alloy and the performance data.

[0089] The transfer learning module is used to train the pre-trained model using the metal matrix composite data set to obtain a metal matrix composite performance prediction model, wherein the weight parameters of neurons in the pre-trained model related to the composition and preparation process of the base alloy are fixed.

[0090] The technical solution of the present invention will be further described in detail below through several embodiments and in conjunction with the accompanying drawings. However, the selected embodiments are only for illustrating the present invention and do not limit the scope of the present invention.

[0091] Example 1

[0092] This embodiment provides a method for predicting the properties of metal matrix composites based on transfer learning. This embodiment has the following limitations:

[0093] S1: Collect several data on the composition, preparation process and mechanical properties of 2xxx, 4xxx, 5xxx, 6xxx and 7xxx aluminum alloys based on public databases or literature, and construct a pre-training dataset using the aforementioned data;

[0094] Specifically, in this implementation example, raw data is collected from public databases or literature to construct an alloy dataset. The dataset includes alloy composition (wt.%), heat treatment process parameters including heat treatment process type, solution temperature (°C), aging temperature (°C), and aging time (h), and mechanical properties including tensile strength (MPa), elongation (%), and yield strength (MPa). The alloy compositions in the pre-training dataset include Al, Si, Fe, Cu, B, Zn, Mn, Mg, Ti, V, Ni, Ce, Cr, Sc, Sr, Zr, and Li.

[0095] S2: Based on publicly available literature, collect data on the matrix composition, preparation process, physicochemical properties of reinforcing phases, interfacial properties, and mechanical properties of aluminum matrix composites reinforced with ceramic particles such as TiB2, ZrB2, SiC, and TiC. Construct an aluminum matrix composite data set using the collected data.

[0096] Specifically, in this implementation example, raw data is collected from public databases or literature to construct an aluminum matrix composite dataset. The aluminum alloy matrix composition and heat treatment process parameters are set consistent with the aluminum alloy dataset obtained in S1; the preparation process also includes the classification of reinforcing phase composite processes and melting temperature (°C), the physicochemical properties of the reinforcing phase include the reinforcing phase content (wt.%) and reinforcing phase particle size (nm), and the interface properties include the thermal mismatch strengthening effect Δσ. CTE (MPa) and the Orovan enhancement effect Δσ Or (MPa); mechanical properties are tensile strength (MPa) and elongation (%).

[0097] Specifically, in this embodiment, the thermal mismatch enhancement effect Δσ CTE Orovan's enhancement effect Δσ Or The calculation formula is:

[0098]

[0099] Where, ρ CTE Δα is the dislocation density formed by the mismatch of thermal expansion coefficients; ΔT is the difference between the thermal expansion coefficients of the reinforcing phase and the matrix; Vp is the volume fraction of the reinforcing phase; dp is the particle size; G is the shear modulus of the aluminum matrix (taken as 26 GPa); b is the Burger vector (taken as 0.286 nm); v is Poisson's ratio; λ is the average spacing between the reinforcing phase particles.

[0100] S3: Preprocess the raw data collected in steps S1 and S2, use Pearson coefficients for feature selection, and standardize the data. Divide the dataset into two sets: one set is randomly selected (80% of the sample data is used as the training set), and the other set is randomly selected (20% of the sample data is used as the test set).

[0101] Specifically, in this implementation example, step S3 includes the following steps:

[0102] S3-1: Physical feature embedding. The content of the reinforcing phase in the dataset obtained in S2 is embedded as a physical feature by embedding the hardness, elastic modulus, melting point, density, and thermal expansion coefficient difference between the embedding layer and the reinforcing phase and the aluminum matrix, and then normalized.

[0103] S3-2: Process coding. The post-processing processes of the datasets obtained from S1 and S2 are coded according to the classification methods of no heat treatment, peak aging, over-aging, work hardening, and artificial aging; the enhanced phase composite processes of the datasets obtained from S2 are coded according to the methods of external addition, molten salt endogenous method, ultrasonic dispersion, and self-propagating endogenous method.

[0104] S4: Based on the alloy dataset obtained in S3, the element feature fusion and multi-head self-attention method of the Transformer algorithm are used to perform matrix transformation on the alloy composition; the coefficient of determination is used as the model performance evaluation index, and the Adam optimizer is used as the loss function optimization algorithm. Through model training, the feature extraction of alloy composition and process parameters is completed, and a pre-trained model with strong correlation performance of alloy matrix is ​​obtained.

[0105] Specifically, in this implementation example, step S4 includes the following steps:

[0106] S4-1: Element feature fusion. The element embedding layer is used to fuse the alloy composition content with the outermost valence electron number, electronegativity, atomic radius and atomic mass of the elements, and transform the alloy composition content into an element feature matrix containing physical meaning.

[0107] S4-2: Self-attention mechanism, adopting a multi-head self-attention mechanism, maps the element feature matrix into query vector, key vector, and value vector, and obtains the interaction between the aluminum matrix alloy components through SoftMax function normalization; the normalized alloy component vector is concatenated with the process parameter input as the input of the fully connected layer neural network;

[0108] S4-3: Tensile strength, elongation, and yield strength are used as the output of the pre-trained model. The determination coefficients of each mechanical property are used as evaluation indicators to select the optimal pre-trained model.

[0109] Specifically, in this implementation example, the function expression of the self-attention mechanism implemented in step S42 is as follows:

[0110]

[0111] Where Q is the query vector, K is the key vector, and V is the value vector; dk is the dimension of the embedded element atomic features, and in this embodiment, dk = 6.

[0112] Specifically, in this implementation example, the pre-trained model algorithm is implemented using Python, with the overall features obtained in S4-2 as the input features; by executing step S4-3, the hyperparameters of the pre-trained model are filtered using grid search, and the optimal hyperparameter settings are obtained as shown in Table 1.

[0113] Table 1: Hyperparameter settings for Transformer-based pre-trained models

[0114]

[0115]

[0116] Specifically, in this embodiment, to prevent the pre-trained model from overfitting the aluminum alloy database and thus increasing the training difficulty of the mechanical property prediction model, an early stopping mechanism is added during the pre-training process. When the coefficient of determination of tensile strength, elongation and tensile strength in the training set is greater than 85%, the training of the pre-trained model is stopped.

[0117] Specifically, in this example, the pre-trained model's prediction accuracy for the three mechanical properties—tensile strength, elongation, and yield strength—is as follows: Figure 2a and Figure 2b As shown, in the training set R 2 The percentages were 96.70%, 92.90%, and 96.50% respectively; R in the test set 2 The percentages were 90.30%, 85.50%, and 89.40%, respectively.

[0118] S5: Based on the metal matrix composite data set and pre-trained model obtained in S3, model construction is achieved through methods such as model transfer, selective model fine-tuning, and grid search. The coefficient of determination is used as the model performance evaluation metric. When the prediction accuracy of each key performance indicator is greater than 90%, proceed to the next step; otherwise, continue to optimize hyperparameters and retrain the model.

[0119] Considering the strong correlation between the yield strength and tensile strength of aluminum matrix composites, the target output performance is reconstructed and a transfer learning process is completed during the construction of the mechanical property prediction model for aluminum matrix composites. This process specifically includes the following methods:

[0120] The type and content of the reinforcing phase are processed in parallel with the composition of the alloy matrix, and the composite process parameters and interface properties are spliced ​​into the input layer of the fully connected neural network.

[0121] Model transfer: Based on the mechanical performance pre-trained model obtained from S4, the weights of the alloy composition-related neurons and the heat treatment process neurons in the Transformer-based neural network are fixed.

[0122] Output layer reconstruction involves reconstructing the loss function, removing non-target strongly correlated mechanical properties, and using the desired mechanical properties as the output layer of the fully connected aluminum matrix composite layer.

[0123] Selective model fine-tuning was performed using the Adam optimization algorithm on the aluminum matrix composite dataset obtained from S3. This process reduced the loss function value and fine-tuned the weights of the fully connected layers.

[0124] Specifically, in this example, the target output performance is set as tensile strength and elongation, and the loss function formula is as follows:

[0125]

[0126] In this embodiment, the loss function is calculated using the mean squared error method, N is the batch size for a single training session, which is set to 64; ω1 and ω2 represent the weight settings for tensile strength and elongation, respectively, which are set to 1 and 1.5; y is the experimental value in the dataset. This represents the predicted value for the model.

[0127] Specifically, in this embodiment, the hyperparameters of the fully connected neural network layer are optimized by combining grid search, and the coefficient of determination is used to evaluate the model performance, thus completing the construction of a predictive model for the mechanical properties of aluminum-based composite materials.

[0128] Specifically, in this example embodiment, see Figure 3a and Figure 3b As shown, R represents the prediction accuracy of the pre-trained model for two mechanical properties: tensile strength and elongation. In the training set, R... 2 The R² values ​​were 99.10% and 98.10% respectively; the R² values ​​on the test set were 94.60% and 91.00% respectively.

[0129] This indicates that the prediction model for the mechanical properties of aluminum-based composite materials performs well on both the training and validation sets, with R... 2 All values ​​were greater than 90%, indicating no obvious overfitting.

[0130] S6: Construct a validation set by collecting independent literature experimental data, compare the prediction results of the final mechanical property prediction model with the experimental values ​​to verify the accuracy of the model; when the prediction deviation is less than the threshold, proceed to the next step; when the prediction deviation is greater than the threshold, include the experimental data in the aluminum-based composite material dataset and return to step S1.

[0131] Specifically, in this example, aluminum-based composite material data that is not in the aluminum-based composite material dataset constructed in S3 is collected; the aluminum-based composite material mechanical property prediction model constructed in S5 is used to predict the corresponding input parameter set, and the predicted values ​​are compared with the experimental values.

[0132] See Figure 4a and Figure 4b The figures show the comparison results for tensile strength and elongation, respectively. The prediction errors for TiB2, ZrB2, and SiC reinforced aluminum matrix composites are all within 15% of the experimental values ​​for tensile strength and elongation, which can further enable the prediction and screening of the mechanical properties of aluminum matrix composites.

[0133] S7: Preset the composition and process parameter set of the target aluminum matrix composite material, use grid search for multi-objective optimization, and use the prediction results to visualize and screen the optimal composition ratio, molding process and heat treatment process parameters of the aluminum matrix composite material.

[0134] Specifically, in this embodiment, an orthogonal design is performed on TiB2 ceramic particle-reinforced 2xxx series aluminum alloy composite materials. The screening method for novel high-performance aluminum-based composite materials includes the following process:

[0135] Specifically, in this embodiment, the screening method for novel high-performance aluminum-based composite materials includes the following steps:

[0136] S71: Optimization of heat treatment process for 2xxx series aluminum alloy matrix. The preset input parameter set composition combination is: w(Cu) = 3wt.%, w(Fe) = 0.1wt.%, w(Si) = 1.4wt.%, w(Mg) = 1.0wt.%, w(Mn) = 0.6wt.%, w(Al) = 93.9wt.%, and the remaining alloy composition is set to 0wt.%.

[0137] The process parameters of the preset input parameter set are set as follows: melting temperature is 750℃, solution temperature is 550℃, aging temperature is 150℃, 180℃, and 210℃; aging time is set to 2h to 40h respectively.

[0138] Multiple sets of predicted data for tensile strength and elongation of 2xxx series aluminum alloy matrices under different heat treatment processes were obtained. Combined with visualization methods of the predicted results, the heat treatment process window for the aluminum alloy matrix can be quickly determined.

[0139] S72: Optimization design of TiB2 ceramic particle reinforced aluminum matrix composite material. Further, based on the screening results of S71, the optimal design range of ceramic particles is: w(TiB2)=1-8wt.%, and the particle size is 200 to 1000nm.

[0140] See Figure 5a and Figure 5bAs shown, multiple sets of predicted data for the tensile strength and elongation of TiB2 ceramic particle-reinforced aluminum matrix composites were obtained. Combining the prediction results with visualization methods, the optimal component ratio and process parameters of this aluminum matrix composite system can be quickly determined.

[0141] S8: Samples are prepared using the molten salt endogenous method according to the material composition ratio and process parameters. Tensile tests are then conducted on the samples to test their tensile strength and elongation.

[0142] Based on the above embodiments, it is clear that the design method provided by the embodiments of the present invention relates to the field of performance prediction of metal matrix composites, and achieves the following technical effects:

[0143] On the one hand, the transfer learning framework integrating Transformer-based neural networks and fully connected neural networks more fully constructs the relationship between alloy matrix composition, heat treatment process and strongly correlated properties, enabling accurate prediction of the properties of metal matrix composites with limited data, and significantly accelerating the research and development of new high-performance metal matrix composites.

[0144] On the other hand, by using dual precision evaluation of the test set and validation set, erroneous predictions are avoided under conditions of low accuracy, thus improving the reliability of model predictions in complex problems. This enables the accurate design of metal matrix composites based on given target performance.

[0145] It should be understood that the above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A training method for a performance prediction model of metal matrix composites based on transfer learning, characterized in that, include: Provides alloy datasets and metal matrix composite datasets. The alloy datasets include the composition, preparation process, and performance data of various base alloys of a selected metal. The metal matrix composite datasets include the metal composition, preparation process, physicochemical properties of the reinforcing phase, interface properties, and performance data of metal matrix composites formed by combining the selected metal or its alloy with a reinforcing phase. The initial model is pre-trained using the alloy dataset to obtain a pre-trained model, which predicts a strong correlation between the composition, preparation process, and performance data of the basic alloy. The pre-trained model is trained using the metal matrix composite data set by transfer learning to obtain a metal matrix composite performance prediction model, wherein the weight parameters of neurons related to the composition and preparation process of the base alloy in the pre-trained model are fixed.

2. The training method according to claim 1, characterized in that, The initial model includes a feature fusion layer, a multi-head self-attention layer, and a fully connected layer. The feature fusion layer uses the Transformer algorithm to fuse the composition features of the base alloy to obtain an element feature matrix. The multi-head self-attention layer maps the element feature matrix into a feature vector and obtains the interaction between components through normalization to form an alloy composition vector. The fully connected layer splices the alloy composition vector and the preparation process of the base alloy to output the predicted performance of the base alloy.

3. The training method according to claim 2, characterized in that, The feature fusion layer fuses the composition content of the base alloy with the number of valence electrons, electronegativity, atomic radius, and atomic mass of the constituent elements, transforming the composition content of the base alloy into an element feature matrix containing physical meaning. The multi-head self-attention layer uses a multi-head self-attention mechanism to map the element feature matrix into a query vector, a key vector, and a value vector. The interaction between the components of the basic alloy is obtained by normalization using the SoftMax function to obtain the alloy composition vector. The alloy composition vector and process parameter input are concatenated as the input for the fully connected layer.

4. The training method according to claim 2, characterized in that, The transfer learning process specifically includes: Model transfer fixes the weights of neurons related to alloy composition and heat treatment process in the feature fusion layer; Output layer reconstruction involves reconstructing the loss function to form a reconstructed loss function, removing non-critical relevant properties, and using the required key properties of the metal matrix composite material as the output layer of the fully connected layer. Selective model fine-tuning, based on the metal matrix composite dataset, involves fine-tuning the weights of the fully connected layers while reducing the value of the reconstruction loss function.

5. The training method according to claim 4, characterized in that, The reconstruction loss function is the weighted average of the variances of the predicted and actual values ​​of multiple performance data.

6. The training method according to claim 1, characterized in that, Also includes: Before performing the pre-training and transfer learning, the input parameters are preprocessed, including physical feature embedding and process encoding. The preprocessed input parameters are then subjected to feature filtering and standardization. The processed alloy dataset and metal matrix composite dataset are divided into training set and test set, respectively. The training set is used for model training, and the test set is used for evaluating the model prediction accuracy.

7. The training method according to claim 6, characterized in that, The standardization process is expressed as follows: Among them, X std Let X be the input parameter after standardization, μ be the preprocessed input parameter, and 6 be the standard deviation of the preprocessed input parameter.

8. The training method according to claim 1, characterized in that, During the pre-training and transfer learning processes, the coefficient of determination is used to evaluate the prediction accuracy of the model; the coefficient of determination is expressed as: Among them, R 2 : represents the performance coefficient; y i Represents performance data; Indicates the predicted value; The value represents the average of the performance data; n represents the sample size.

9. A method for preparing a metal matrix composite material, characterized in that, include: The optimal performance prediction model for metal matrix composites is obtained by using the training method described in any one of claims 1-8; Input the set of relevant parameters of the metal matrix composite material into the performance prediction model of the metal matrix composite material to obtain the set of predicted performance. Select the optimal relevant parameters from the predicted performance set; Metal matrix composites were prepared based on the optimal correlation parameters.

10. A training system for a performance prediction model of metal matrix composites based on transfer learning, characterized in that, include: The basic data module is used to provide alloy datasets and metal matrix composite datasets. The alloy datasets include the composition, preparation process, and performance data of various base alloys of the selected metal. The metal matrix composite datasets include the metal composition, preparation process, physicochemical properties of the reinforcing phase, interface properties, and performance data of the metal matrix composites formed by combining the selected metal or its alloy with a reinforcing phase. The pre-training module is used to pre-train the initial model using the alloy dataset to obtain a pre-trained model, which predicts the strong correlation between the composition and preparation process of the base alloy and the performance data. The transfer learning module is used to train the pre-trained model using the metal matrix composite data set to obtain a metal matrix composite performance prediction model, wherein the weight parameters of neurons in the pre-trained model related to the composition and preparation process of the base alloy are fixed.

Images