Metal matrix composite parameter design method, device and medium based on deep learning

Abstract

The application discloses a kind of based on deep learning's metal matrix composite parameter design method, equipment and medium, belong to metal matrix composite field.The application obtains the process parameter of metal matrix composite, constructs process parameter space, all parameters in process parameter space are transported to the machine learning model constructed in advance, the mechanical property prediction result is obtained by machine learning model, further according to first preset threshold, performance standard result is selected from mechanical property prediction result, then the qualified process parameter corresponding to performance standard result is obtained, and the qualified parameter space is obtained by combination.The application filters out the parameter combination that the mechanical property is not qualified by prediction, reduces the range of industrial parameter set, and the parameter is selected from the qualified parameter space by research and development personnel for metal matrix composite research and development, so that the metal matrix composite satisfying performance requirement can be obtained faster, to shorten research and development time.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of metal matrix composites, in particular to a metal matrix composite parameter design method, device and medium based on deep learning. BACKGROUND

[0002] Metal matrix composites are composites artificially combined with one or more metal or non-metal reinforcing phases as the matrix. Through reasonable composition, performance design and process, the metal matrix and the reinforcing phase can be optimally combined to obtain a composite material that has good plasticity and toughness of the metal matrix, easy processability and electrical and thermal conductivity, and also has high hardness of the reinforcing phase, good thermal stability, low expansion coefficient and other advantages to meet the needs of advanced equipment manufacturing industry, and has been widely used in aerospace, energy engineering, automobile manufacturing, shipbuilding and marine engineering equipment fields.

[0003] However, the preparation process of metal matrix composites is complex, including powder mixing, ingot loading, sintering, forming, heat treatment and other links, involving powder particle size, reinforcing phase content, sintering pressure, sintering temperature, forming process parameters, heat treatment process parameters and other aspects. In practical applications, metal matrix composites usually need to meet both strength and toughness requirements, and exploring a suitable process window in such a complex parameter space often occupies a large amount of time for enterprises and research institutions, and how to shorten the process exploration process is a major problem in the industry. SUMMARY

[0004] The main purpose of the present application is to provide a metal matrix composite parameter design method, device and medium based on deep learning, which aims to solve the problem of long research and development time of metal matrix composites and shorten the process parameter exploration process of metal matrix composites.

[0005] To achieve the above-mentioned purpose, the present application provides a metal matrix composite parameter design method based on deep learning, which comprises the following steps:

[0006] Obtain the process parameters of the metal matrix composite, and construct a process parameter space;

[0007] Transport all parameters in the process parameter space to a pre-constructed machine learning model to obtain a mechanical property prediction result;

[0008] According to a first preset threshold, select a performance meeting result from the mechanical property prediction result;

[0009] Obtain the qualified process parameters corresponding to the performance meeting result, and combine to obtain a qualified parameter space.

[0010] Optionally, before the step of delivering all parameters in the process parameter space to a pre-constructed machine learning model to obtain a mechanical property prediction result, the method further comprises:

[0011] selecting a preset number of process parameter combinations from the process parameter space, and obtaining experimental mechanical properties corresponding to the process parameter combinations;

[0012] constructing a training set and a test set through the experimental mechanical properties and the process parameter combinations;

[0013] training a first initial model through the training set, and verifying the accuracy of the first initial model through the test set, and obtaining the machine learning model when the accuracy meets a preset condition.

[0014] Optionally, the step of training a first initial model through the training set comprises:

[0015] inputting the training set into a gated residual network, and training a second initial model in combination with a parameter selection network;

[0016] calculating the accuracy and loss function of the second initial model, adjusting the hyperparameters of the second initial model according to the accuracy and the loss function, and obtaining a first initial model after parameter adjustment.

[0017] Optionally, the step of adjusting the hyperparameters of the second initial model according to the accuracy and the loss function to obtain a first initial model after parameter adjustment comprises:

[0018] iteratively calculating the loss function and the accuracy of the second initial model;

[0019] recursively performing the step of reducing the learning rate to a preset multiple when the loss function converges until the accuracy reaches a second preset threshold, stopping iteration and obtaining a first initial model after parameter adjustment.

[0020] Optionally, the step of delivering parameters in the process parameter space to a pre-constructed machine learning model comprises:

[0021] constructing a completely orthogonal parameter space based on all parameters of the process parameter space;

[0022] inputting the completely orthogonal parameter space into a pre-constructed machine learning model.

[0023] Optionally, the mechanical properties include fracture strength and elongation at break, and the step of selecting performance meeting results from the mechanical property prediction results according to a first preset threshold comprises:

[0024] acquire a first predicted performance meeting probability corresponding to the fracture strength in the mechanical property prediction result and a second predicted performance meeting probability corresponding to the elongation at break, and compare the first predicted performance meeting probability and the second predicted performance meeting probability with the first preset threshold value respectively;

[0025] The mechanical property prediction result, in which both the first predicted performance meeting probability and the second predicted performance meeting probability are greater than the first preset threshold value, is screened out as the performance meeting result.

[0026] Optionally, after the step of acquiring the qualified process parameters corresponding to the performance meeting result and combining to obtain the qualified parameter space, the method further comprises:

[0027] According to a preset condition, process parameters meeting the preset condition are screened out from the qualified process parameter space to construct an application process parameter space, and the application process parameter space is used for metal matrix composite product research and development.

[0028] In addition, to achieve the above-mentioned purpose, the application further provides a metal matrix composite parameter design device based on deep learning, which comprises:

[0029] A process parameter module is configured to acquire process parameters of a metal matrix composite and construct a process parameter space.

[0030] A machine learning module is configured to input parameters in the process parameter space into a pre-constructed machine learning model to obtain a mechanical property prediction result.

[0031] A screening module is configured to select a performance meeting result from the mechanical property prediction result according to a first preset threshold value.

[0032] A qualified parameter module is configured to acquire qualified process parameters corresponding to the performance meeting result and combine to obtain a qualified parameter space.

[0033] Optionally, the machine learning module is further configured to:

[0034] select a preset number of process parameter combinations from the process parameter space and acquire experimental mechanical properties corresponding to the process parameter combinations.

[0035] construct a training set and a test set through the experimental mechanical properties and the process parameter combinations.

[0036] train a first initial model through the training set and verify the accuracy of the first initial model through the test set, and when the accuracy meets a preset condition, the machine learning model is obtained.

[0037] Optionally, the machine learning module is further configured to:

[0038] input the training set into the gated residual network and combine the parameter selection network to obtain a second initial model through training;

[0039] calculate the accuracy and loss function of the second initial model, adjust the hyperparameters of the second initial model according to the accuracy and the loss function, and obtain a first initial model after parameter adjustment.

[0040] Optionally, the machine learning module is further configured to:

[0041] iteratively calculate the loss function and the accuracy of the second initial model;

[0042] perform the step of reducing the learning rate to a preset multiple when the loss function converges until the accuracy reaches a second preset threshold, stop iteration and obtain a first initial model after parameter adjustment.

[0043] Optionally, the machine learning module is further configured to:

[0044] based on all parameters of the process parameter space, construct a full orthogonal parameter space;

[0045] input the full orthogonal parameter space into a pre-constructed machine learning model.

[0046] Optionally, the screening module is further configured to:

[0047] obtain a first predicted performance compliance probability corresponding to the fracture strength in the mechanical property prediction result and a second predicted performance compliance probability corresponding to the elongation at break, and compare them with the first preset threshold, respectively;

[0048] screen out the mechanical property prediction result whose first predicted performance compliance probability and second predicted performance compliance probability are both greater than the first preset threshold as a performance compliance result.

[0049] Optionally, the qualified parameter module is further configured to:

[0050] According to a preset condition, screen out process parameters that meet the preset condition from the qualified process parameter space to construct an application process parameter space, and the application process parameter space is used for metal matrix composite product research and development.

[0051] In addition, to achieve the above-mentioned purpose, the application further provides a metal matrix composite parameter design device based on deep learning, which comprises a memory, a processor and a metal matrix composite parameter design program based on deep learning stored on the memory and executable on the processor, and the metal matrix composite parameter design program based on deep learning is configured to implement the steps of the metal matrix composite parameter design method based on deep learning.

[0052] Further, to achieve the above object, the application further provides a medium, characterized in that the medium stores a deep learning-based metal matrix composite parameter design program, and the deep learning-based metal matrix composite parameter design program, when executed by a processor, implements the steps of the deep learning-based metal matrix composite parameter design method.

[0053] The deep learning-based metal matrix composite parameter design method provided by the application comprises the following steps: obtaining process parameters of a metal matrix composite material; constructing a process parameter space; inputting parameters in the process parameter space to a pre-constructed machine learning model; obtaining a mechanical property prediction result from the machine learning model; selecting a performance meeting result from the mechanical property prediction result according to a first preset threshold; obtaining qualified process parameters corresponding to the performance meeting result; and combining the qualified process parameters to obtain a qualified parameter space. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 FIG. 1 is a structural schematic diagram of a deep learning-based metal matrix composite parameter design device according to an embodiment of the application;

[0055] Figure 2 FIG. 2 is a flowchart of a deep learning-based metal matrix composite parameter design method according to a first embodiment of the application;

[0056] Figure 3 FIG. 3 is a flowchart of constructing a machine learning model in the deep learning-based metal matrix composite parameter design method according to the application;

[0057] Figure 4 FIG. 4 is a closed residual network diagram according to an embodiment of the deep learning-based metal matrix composite parameter design method according to the application;

[0058] Figure 5 FIG. 5 is a parameter selection network diagram according to an embodiment of the deep learning-based metal matrix composite parameter design method according to the application;

[0059] Figure 6 FIG. 6 is a partial data table of a training set according to an embodiment of the deep learning-based metal matrix composite parameter design method according to the application;

[0060] Figure 7 FIG. 7 is a diagram showing the changes of loss rate, precision rate and learning rate of a second initial model with the training round number according to an embodiment of the deep learning-based metal matrix composite parameter design method according to the application.

[0061] Figure 8 A schematic diagram for predicting results of all parameter spaces by a machine learning model and screening qualified parameter spaces in an embodiment of the metal matrix composite parameter design method based on deep learning of the present application;

[0062] Figure 9 A process parameter list in a qualified parameter space in an embodiment of the metal matrix composite parameter design method based on deep learning of the present application which does not need subsequent heat treatment;

[0063] Figure 10 A functional module schematic diagram of an embodiment of the metal matrix composite parameter design method based on deep learning of the present application.

[0064] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0065] It should be understood that the specific embodiments described herein are merely intended to explain the present application and are not intended to limit the present application.

[0066] Reference Figure 1 , Figure 1 A hardware running environment based on deep learning of the metal matrix composite parameter design device structure schematic diagram involved in the embodiment scheme of the present application.

[0067] As Figure 1 shown, the metal matrix composite parameter design device based on deep learning can include a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to realize the connection and communication between the components. The user interface 1003 can include a display screen, an input unit such as a keyboard, and an optional user interface 1003 can also include a standard wired interface, a wireless interface. The network interface 1004 can optionally include a standard wired interface, a wireless interface (such as a wireless fidelity (WIreless-FIdelity, WI-FI) interface). The memory 1005 can be a high-speed random access memory (RAM) memory, and can also be a stable non-volatile memory (NVM), such as a magnetic disk memory. The memory 1005 can also be a storage device independent of the aforementioned processor 1001.

[0068] Those skilled in the art can understand, Figure 1The structure shown in the figure does not constitute a limitation on the deep learning-based metal matrix composite parameter design device, and can include more or fewer components than the figure, or combine certain components, or different component arrangements.

[0069] As shown in Figure 1 The memory 1005 as a storage medium can include an operating system, a data storage module, a network communication module, a user interface module, and a deep learning-based metal matrix composite parameter design program.

[0070] In Figure 1 The network interface 1004 of the deep learning-based metal matrix composite parameter design device is mainly used for data communication with other devices; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the deep learning-based metal matrix composite parameter design device of the application can be arranged in the deep learning-based metal matrix composite parameter design device, and the deep learning-based metal matrix composite parameter design device calls the deep learning-based metal matrix composite parameter design program stored in the memory 1005 through the processor 1001, and executes the deep learning-based metal matrix composite parameter design method provided by the embodiment of the application.

[0071] The embodiment of the application provides a deep learning-based metal matrix composite parameter design method, which refers to Figure 2 , Figure 2 The flowchart of a first embodiment of a deep learning-based metal matrix composite parameter design method of the application.

[0072] In this embodiment, the deep learning-based metal matrix composite parameter design method comprises:

[0073] Step S10, obtaining the process parameters of the metal matrix composite material, and constructing a process parameter space;

[0074] Step S20, delivering all parameters in the process parameter space to a pre-constructed machine learning model to obtain a mechanical property prediction result;

[0075] Step S30, selecting a performance meeting result from the mechanical property prediction result according to a first preset threshold;

[0076] Step S40, obtaining qualified process parameters corresponding to the performance meeting result, and combining to obtain a qualified parameter space.

[0077] The embodiment of the metal matrix composite parameter design method based on deep learning is applied to the research and development of metal matrix composites. In recent years, with the rapid development of algorithms and computing power, machine learning has shown great application value in various disciplines. Combining machine learning with metal matrix composite preparation process, potential information is mined from the data obtained from the experiment to establish a relationship model between process parameters and strength toughness, and a suitable process window can be explored according to the target performance to realize the rapid development of metal matrix composites. Therefore, the embodiment obtains the process parameters of the preparation of the metal matrix composite, constructs a process space, and then predicts the mechanical properties corresponding to the process parameters through a machine learning model, and selects the process parameter set corresponding to the qualified mechanical properties from the process parameters, thereby constructing a qualified parameter space.

[0078] The following will be described in detail:

[0079] In step S10, the process parameters of the metal matrix composite are obtained, and a process parameter space is constructed.

[0080] In an embodiment, the process parameters of the metal matrix composite are obtained, and a process parameter space is constructed. Specifically, the process parameters used for preparing the metal matrix composite are obtained according to the existing research and production conditions, and the process parameter space is constructed by combining the process parameters. The commonly used methods for preparing metal matrix composites include stirring casting method, powder metallurgy method, in-situ generation method, etc., and the corresponding process parameters include composition, temperature, pressure, etc. Different preparation methods correspond to different process parameters. The preparation process and the corresponding process parameters can be confirmed according to the research needs, and the process parameter space is constructed.

[0081] In step S20, the parameters in the process parameter space are input into the machine learning model constructed in advance to obtain the mechanical property prediction result.

[0082] In an embodiment, the parameters in the process parameter space are input into the machine learning model constructed in advance to predict the mechanical property prediction result corresponding to the process parameters. Machine learning can be divided into classification problems and regression problems. For the specific problem in this scheme, the regression problem can predict specific performance data such as fracture strength and elongation at break, etc. However, the accuracy of the regression problem needs to rely on a large amount of data, and the data mainly comes from the experimental data in step S10. It is not realistic to generate a large amount of data through experiments, and it also goes against the original intention of shortening the development cycle. Therefore, this scheme proposes to use a classification problem, and the strength performance and toughness performance are classified into two categories: qualified (represented by floating point data 1.0) and unqualified (represented by floating point data 0.0) according to the expected performance. In this way, the accuracy of the model prediction can be greatly improved. Specifically, a machine learning model can be constructed by selecting a traditional neural network or a commonly used network for classification models.

[0083] Step S30, selecting performance meeting results from the mechanical property prediction results according to a first preset threshold value;

[0084] In an embodiment, the performance meeting results are selected from the mechanical property prediction results of the machine learning model according to a first preset threshold value. Specifically, since a classification problem is used to solve the problem, there are only two classifications of meeting and not meeting. The metal matrix composite material has high specific strength and high specific modulus, high strength in the transverse direction and shear, good toughness and compression resistance of the material, and other characteristics. Among them, the strength and toughness of the metal matrix composite material are the focus. Toughness refers to the resistance of a material to break when subjected to a force that causes it to deform. The better the toughness, the less likely it is to break in a brittle manner. Strength is one of the most basic mechanical properties of engineering materials. Common strength performance indicators include tensile strength and yield strength. Elongation at break refers to the percentage of the length of a test bar that elongates when it is subjected to an external force (tension) and can be used to characterize the tensile strength of a material. Breaking strength, also known as break load, is the ratio of the tensile force to the cross-sectional area at the time of fracture, i.e. stress. According to the characteristics of the metal matrix composite material, the model takes the breaking strength exceeding 200 MPa as the standard, and the elongation at break is not less than 8% as the standard. The prediction results include the performance meeting probability of the room temperature breaking strength and the room temperature elongation at break, and the results exist in floating point form. After obtaining the mechanical property prediction results, the results with a performance meeting probability greater than the first preset threshold value are selected as the performance meeting results from the prediction results. The first preset threshold value can be selected according to the actual situation, and preferably, the first preset threshold value of the present example is 0.9, and the breaking strength and elongation at break with a probability greater than 0.9 are considered as qualified performance. It should be noted that the threshold value of the model is set according to experience, and in the present embodiment, the breaking strength exceeding 200 MPa is taken as the standard, and the elongation at break is not less than 8% as the standard. Of course, other indicators and threshold values can also be selected to characterize the strength and toughness of the metal matrix composite material.

[0085] Step S40, obtaining qualified process parameters corresponding to the performance meeting results, and combining to obtain a qualified parameter space.

[0086] In one embodiment, qualified industrial parameters are combined to obtain a qualified parameter space by obtaining qualified process parameters corresponding to the performance compliance results. It can be understood that when the probability of performance compliance is greater than a first preset threshold, the corresponding preparation process parameters that can achieve these compliant mechanical properties are found, and these parameters are combined to obtain the qualified parameter space. In step S30, the model specifies the threshold for the predicted result compliance, and the probability of performance compliance is obtained through prediction. Therefore, the industrial parameters corresponding to these mechanical property prediction results with a performance compliance probability greater than the first preset threshold can be further studied and experimented with to prepare metal matrix composite materials that meet the requirements. Thus, we construct the qualified parameter space from these process parameters that can achieve compliant performance.

[0087] In this embodiment, by acquiring the process parameters of the metal matrix composite material, the mechanical properties of all process parameters are predicted by a pre-built machine learning model. The process parameter combination corresponding to the qualified mechanical properties is selected to obtain the qualified process parameter space. Without conducting experiments on all process parameters, the process parameters that meet the preset performance requirements of the metal matrix composite material are obtained, which greatly reduces the time spent on experiments and improves the R&D efficiency.

[0088] Furthermore, based on the first embodiment of the deep learning-based metal matrix composite parameter design method of the present invention, a second embodiment of the deep learning-based metal matrix composite parameter design method of the present invention is proposed.

[0089] Reference Figure 3 , Figure 3 This is a flowchart illustrating the construction of a machine learning model in the deep learning-based parameter design method for metal matrix composites of the present invention. In the second embodiment, before the step of feeding all parameters in the process parameter space to the pre-constructed machine learning model to obtain the mechanical property prediction results, the method further includes:

[0090] Step S11: Select a preset number of process parameter combinations from the process parameter space, and obtain the experimental mechanical properties corresponding to the process parameter combinations;

[0091] Step S12: Construct a training set and a test set by combining the experimental mechanical properties and the process parameters;

[0092] Step S13: Train the first initial model using the training set, and verify the accuracy of the first initial model using the test set. When the accuracy meets the preset conditions, obtain the machine learning model.

[0093] In this embodiment, representative combinations of process parameters are selected from the parameter space to conduct several sets of experiments, and the mechanical properties related to the strength and toughness of the material are tested. Based on the limited experimental data, training sets and test sets are constructed to train a machine learning model.

[0094] The following will provide a detailed explanation of each step:

[0095] Step S11: Select a preset number of process parameter combinations from the process parameter space, and obtain the experimental mechanical properties corresponding to the process parameter combinations;

[0096] In one embodiment, several sets of experiments are conducted using representative combinations of process parameters selected from the process parameter space to test the mechanical properties related to the material's strength and toughness. In practical applications, considering that machine learning is a data-driven technology with a strong reliance on data, the selection of parameters needs to be carefully considered when establishing the training set in the early stages of the experiment. Representative parameter combinations should be selected as evenly and evenly as possible within the parameter space to facilitate subsequent model analysis. Therefore, to improve the model's prediction accuracy and reduce experimental data preparation time, this scheme selects a predetermined number of process parameter combinations for experiments.

[0097] Step S12: Construct a training set and a test set by combining the experimental mechanical properties and the process parameters;

[0098] In one embodiment, experimental data—combinations of process parameters and their corresponding experimental mechanical properties—are used to construct training and testing sets. Specifically, experimental mechanical properties are obtained by selecting a subset of process parameters for experiments. Since training and testing sets are essential in machine learning, the simplest method to divide the training and testing sets is through the hold-out method, directly partitioning the dataset of experimental mechanical properties and corresponding combinations of process parameters into two mutually exclusive sets, one as the training set and the other as the testing set. After training the model on the training set, the testing set is used to evaluate its testing error. Alternatively, cross-validation can be used to divide the training and testing sets. For example, if an experimental dataset has 100 data points, it can be divided into training and testing sets in an 8:2 ratio. Then, the training and testing sets can be re-divided four times, each time using 80 different data points as the training set and 20 as the testing set. Finally, the results of the five experiments are averaged to obtain a 5-fold cross-validation.

[0099] Step S13: Train the first initial model using the training set, and verify the accuracy of the first initial model using the test set. When the accuracy meets the preset conditions, obtain the machine learning model.

[0100] In one embodiment, a first initial model is trained using a training set, and the accuracy of the initial model is verified using a test set. The first initial model is trained by inputting the training set data into a neural network, such as a multi-layer feedforward neural network, and its accuracy is verified using a test set. When the accuracy meets a preset condition, a machine learning model is obtained. The preset condition is a threshold for the model's evaluation metrics. Various evaluation metrics can be used to assess model accuracy, such as precision, recall, accuracy, F1 score, G-metric, and ROC curve. Different evaluation metrics correspond to different preset conditions; therefore, the preset condition can be determined based on the selected accuracy evaluation metric.

[0101] Further, in one embodiment, the step of training the first initial model using the training set includes:

[0102] Step S131: Input the training set into the gated residual network and combine it with the parameter selection network to train and obtain the second initial model;

[0103] Step S132: Calculate the accuracy and loss function of the second initial model, and adjust the hyperparameters of the trained second initial model according to the accuracy and loss function to obtain the first initial model after parameter tuning.

[0104] In this embodiment, a gated residual network and a parameter selection network are selected to train the model. After the second initial model is trained, the hyperparameters of the second initial model are adjusted by the loss function and accuracy to optimize the second initial model and obtain the optimized first initial model.

[0105] The following will provide a detailed explanation of each step:

[0106] Step S131: Input the training set into the gated residual network and combine it with the parameter selection network to train and obtain the second initial model;

[0107] In one embodiment, considering the complexity of the metal matrix composite material preparation process, which includes both digital data such as cold pressing pressure, sintering temperature, and extrusion temperature, and textual data such as powder type, heat treatment process, and extrusion process, and given the coupling effect between these parameters, which influence each other and collectively affect the strength and toughness of the material, this solution proposes using a gated residual network (GRN) combined with a variable selection network (VSN) to train the parameters. See the schematic diagram of the specific model structure. Figure 4 and Figure 5 ,exist Figure 4The precise relationship between the source and target variables is often unknown, making it difficult to predict which variables are relevant. To allow the model to flexibly apply nonlinear processing only where needed, we propose... Figure 4 The GRN shown serves as a building block for VSN. The GRN accepts a primary input 'a' and an optional background input 'c' and produces:

[0108] GRN ω (a, c) = LayerNorm(a + GLU) ω (η1))

[0109] η1=W 1，ω η2+b 1，ω

[0110] η2=ELU(W 2，ω a+W 3，ω c+b 2，ω )

[0111]

[0112] In the above formula, LayerNorm is the layer normalization function;

[0113] ELU is an activation function;

[0114] W is the weight, and b is the bias.

[0115] While multiple variables may be available, their correlation and specific contribution to the output are often unknown. VSN aims to provide instance variable selection by using a variable selection network applied to both static and time-dependent covariates. Besides providing insight into the variables most important to the prediction problem, variable selection allows VSN to remove any unwanted noise input that might negatively impact performance. Most real-world datasets contain features with limited predictability, so VSN can filter parameters that have a greater impact on predictions during model training to reduce computation and improve predictive performance. See details... Figure 5 In the parameter selection network (VSN), where:

[0116] It is the i-th input. Weighting function

[0117] m χ That is, max is the last value.

[0118] Flattened input:

[0119] Normalized exponential function:

[0120] Closed residual networks provide flexibility to the model, allowing nonlinear processing to be applied only where necessary; parameter selection networks allow the model to smoothly remove any unnecessary noise inputs that might negatively impact performance. The combination of these two technologies helps improve the machine learning model's ability to understand the effects of multi-parameter coupling on results, thus enhancing its learning capacity. Without closed residual networks and parameter selection networks, such as a traditional 6-layer deep learning neural network, the trained model will produce poor predictive performance.

[0121] Step S132: Calculate the accuracy and loss function of the second initial model, and adjust the hyperparameters of the trained second initial model according to the accuracy and loss function to obtain the first initial model after parameter tuning.

[0122] In one embodiment, the hyperparameters of the model are adjusted by calculating the accuracy and loss function of the second initial model to obtain the hyperparameter-tuned first initial model. Typically, model performance can be adjusted by changing the number of hidden layer nodes, selecting different activation functions, and setting different learning rates. To achieve higher accuracy, this embodiment adjusts the hyperparameters of the second initial model by calculating its accuracy and loss function. Hyperparameters are parameters set before the learning process begins, such as the number of training epochs and the learning rate. Optimizing the hyperparameters can improve learning performance and effectiveness.

[0123] Further, in one embodiment, the step of adjusting the hyperparameters of the second initial model according to the accuracy and the loss function to obtain the hyperparameter-tuned first initial model includes:

[0124] Step S1321: Iteratively calculate the loss function and accuracy of the second initial model;

[0125] Step S1322: Repeatedly execute the step of reducing the learning rate to a preset multiple when the loss function converges, until the accuracy reaches a second preset threshold, then stop the iteration and obtain the first initial model after parameter tuning.

[0126] This embodiment iteratively calculates the loss function and accuracy of the second initial model to adjust the model's learning rate. When the loss function is detected to have converged, the learning rate is reduced, thereby achieving higher prediction accuracy of the model.

[0127] The following will provide a detailed explanation of each step:

[0128] Step S1321: Iteratively calculate the loss function and accuracy of the second initial model;

[0129] In one embodiment, the loss function and accuracy of the second initial model are calculated iteratively. To evaluate the model's performance, accuracy is chosen as the metric; accuracy refers to the proportion of correctly classified records out of the total number of records. The loss function is selected based on the classification model; for example, a logarithmic loss function is used for a binary classification problem based on logistic regression. The loss function and accuracy are calculated several times after each input training sample until the required number of training epochs are reached.

[0130] Step S1322: Repeatedly execute the step of reducing the learning rate to a preset multiple when the loss function converges, until the accuracy reaches a second preset threshold, then stop the iteration and obtain the first initial model after parameter tuning.

[0131] In one embodiment, the step of reducing the learning rate by a preset factor when the loss function converges is executed iteratively until the accuracy reaches a second preset threshold. The iteration then stops, and the first initial model after parameter tuning is obtained. It is understandable that in machine learning, if the learning rate is too large, the loss function may directly bypass the global optimum, resulting in an excessively large loss or a NaN value. If the learning rate is too small, the loss function changes very slowly, greatly increasing the convergence complexity of the network and making it prone to getting trapped in local minima or saddle points. Therefore, it is necessary to adjust the learning rate so that the model's accuracy reaches the second preset threshold. Since hyperparameters cannot be learned directly from the data during the standard model training process, they need to be predefined. Therefore, this embodiment detects the loss function and reduces the learning rate by a preset factor when the loss function reaches a plateau to obtain better accuracy. Typically, the preset factor ranges from 0.02 to 0.8, and the specific value can be determined according to the actual situation.

[0132] Further, in one embodiment, the step of feeding all parameters in the process parameter space to a pre-built machine learning model includes:

[0133] Step S21: Based on the parameters of the process parameter space, construct a completely orthogonal parameter space;

[0134] Step S22: Input the fully orthogonal parameter space into the pre-built machine learning model.

[0135] In this embodiment, a fully orthogonal parameter space is constructed by integrating all parameters in the process parameter space. This fully orthogonal parameter space is then input into a pre-built machine learning model. Typically, the preparation of metal matrix composites involves parameters such as powder particle size, reinforcing phase content, sintering pressure, sintering temperature, molding process parameters, and heat treatment process parameters. Specifically, research factors and index levels are determined based on the process parameters. Factors refer to parameters or indicators that are the research objects, and levels are the possible values ​​of a factor. In this embodiment, the orthogonal parameter space is constructed by arranging and combining process parameter types and their values. Furthermore, all parameter combinations in the fully orthogonal parameter space are input into the pre-built machine learning model to obtain the predicted mechanical properties of various parameter combinations.

[0136] In this embodiment, a fully orthogonal parameter space is constructed by constructing parameters from the process parameter space. A gated residual network and a parameter selection network are selected for model construction. The parameters in the fully orthogonal parameter space are input into the model, and the loss function and accuracy of the constructed second initial model are calculated. The learning rate of the second initial model is adjusted according to the loss function and accuracy until the accuracy of the second initial model reaches a second preset threshold, thus obtaining the parameter-tuned first initial model. Then, the accuracy of the first initial model is verified through a test set. When the accuracy of the first initial model meets the preset conditions, a machine learning model is obtained.

[0137] Furthermore, based on the previous embodiment of the deep learning-based metal matrix composite parameter design method of the present invention, a third embodiment of the deep learning-based metal matrix composite parameter design method of the present invention is proposed. In this embodiment, the mechanical properties include: fracture strength and elongation after fracture. The step of selecting the performance meeting the standard result from the predicted mechanical properties according to a first preset threshold includes:

[0138] Step S31: Obtain the first predicted performance compliance probability corresponding to fracture strength and the second predicted performance compliance probability corresponding to post-fracture elongation in the mechanical performance prediction results, and compare them with the first preset threshold respectively.

[0139] In one embodiment, the first predicted performance compliance probability corresponding to fracture strength in the mechanical property prediction results is compared with a first preset threshold, and the second predicted performance compliance probability corresponding to elongation after fracture in the mechanical property prediction results is also compared with the first preset threshold. Since both strength and toughness are important for metal matrix composites, and these two performance indicators are typically required for suitable applications, this embodiment uses a first preset threshold to screen the compliance probabilities of fracture strength and elongation after fracture.

[0140] Step S32: Select mechanical performance prediction results where both the first predicted performance compliance probability and the second predicted performance compliance probability are greater than the first preset threshold, and use them as performance compliance results.

[0141] In one embodiment, the predicted results where both the probability of achieving the first performance standard for fracture strength and the probability of achieving the second performance standard for elongation after fracture are greater than a first preset threshold are taken as the performance standard achievement results. In this example, both the strength and toughness of the metal matrix composite material were examined. In order to facilitate the selection of a better parameter combination, it is required that the probability of achieving the performance standard for fracture strength and the probability of achieving the performance standard for elongation after fracture are both greater than the first preset threshold, where the first preset threshold can be set according to the actual situation, for example, 0.8 or 0.9.

[0142] Furthermore, in one embodiment, after the step of obtaining the qualified process parameters corresponding to the performance compliance result and combining them to obtain the qualified parameter space, the method further includes:

[0143] Step S41: Based on preset conditions, select process parameters that meet the preset conditions from the qualified process parameter space to construct an application process parameter space, which is used for the research and development of metal matrix composite products.

[0144] In this embodiment, after obtaining a qualified process parameter space, process parameters that meet preset conditions are selected to construct an application process parameter space. It is understood that there are still many sets of process parameters within the qualified process parameter space. Based on the research and development direction and objectives, some process parameters can be further selected for the research and development of metal matrix composites. For example, process parameters with good economic benefits, i.e., those that do not require heat treatment, can be selected to prepare products, or process parameters with a particularly high probability of achieving performance targets can be chosen. Products prepared using these process parameters can reduce the probability of failure to a certain extent.

[0145] In this embodiment, the mechanical performance prediction results obtained by the machine learning model are filtered through a first preset threshold. Predictions with a first performance qualification probability of fracture strength and a second performance qualification probability of elongation after fracture both greater than the first preset threshold are taken as performance qualification results. After obtaining the performance qualification results, process parameters that meet different R&D needs are selected from them through preset conditions, and a corresponding process parameter space is constructed. Parameters in the process parameter space are selected for subsequent experiments and R&D, which greatly reduces the R&D time.

[0146] Furthermore, based on the previous embodiments of the deep learning-based metal matrix composite parameter design method of the present invention, a fourth embodiment of the deep learning-based metal matrix composite parameter design method of the present invention is proposed. The deep learning-based metal matrix composite parameter design method of the present invention will be described according to this embodiment:

[0147] This example demonstrates the preparation of an aluminum-based silicon carbide composite material using an extrusion molding process, followed by heat treatment. One hundred sets of data were obtained and divided into a 70% training set and a 30% validation set. Figure 6 As shown, Figure 6 This is a partial data table of the training set for an embodiment of the parameter design method for metal matrix composites based on deep learning of the present invention. The main process parameters include powder composition (Comp, 3.0 is the particle size of aluminum powder of 3.0 μm, 10% is the mass percentage of silicon carbide, the same below), cold pressing pressure (300.0 is 300 MPa, the same below), sintering temperature (Sin_temp, 450.0 is the sintering temperature of 450℃, the same below), extrusion temperature (Ex_temp, 460 is the extrusion temperature of 460℃, the same below), extrusion shape (Ex_ratio, 200*20mm is the cross-sectional size, the same below), and heat treatment process (HT, 500C / 2h is the heat treatment at 500℃ for 2h, Nada is no heat treatment, the same below). The results are room temperature tensile strength (RT_UTS, NaN is unacceptable, 1.0 is acceptable) and elongation after fracture (RT_Elong).

[0148] After inputting the training set into the model, training is performed. This example problem has only two categories: compliant and non-compliant. Compliant is defined as a fracture strength exceeding 200 MPa and an elongation at break of at least 8%. Therefore, this example uses binary cross-entropy as the loss function and incorporates a dynamic learning rate adjustment feature, reducing the learning rate when the loss function reaches a plateau. (Refer to...) Figure 7 , Figure 7 This diagram illustrates the changes in loss rate, accuracy, and learning rate of the second initial model with the number of training rounds in one embodiment of the deep learning-based metal matrix composite parameter design method of the present invention. It can be seen that when the loss function reaches a plateau each time, the learning rate decreases to 0.4 times the original rate. After 500 rounds of training, the model's prediction accuracy can reach 93.26%, achieving high-precision prediction.

[0149] like Figure 8 As shown, Figure 8This diagram illustrates how a deep learning-based parameter design method for metal matrix composites of the present invention uses a machine learning model to predict the results of the entire parameter space and select qualified parameter spaces. After the model is trained, all 62,208 sets of data in the parameter space are input into the model to obtain the prediction results. The prediction results are divided into two columns, representing room temperature fracture strength and room temperature elongation after fracture, respectively. The results are in floating-point form, representing the probability that the performance meets the standard. In this example, 0.9 is taken as the first preset threshold. Performances with fracture strength and elongation after fracture that are approximately greater than 0.9 are considered qualified performances. 8,566 sets of qualified performances are selected from all predicted performances, and the corresponding qualified process parameters are found, forming a qualified parameter space consisting of 8,566 sets of processes.

[0150] In actual production, while subsequent heat treatment can effectively balance the conflict between strength and toughness, it typically requires significant time and cost. Therefore, this solution filters out 165 sets of process parameters from the qualified parameter space that do not require subsequent heat treatment, such as... Figure 9 , Figure 9 This is a list of process parameters that do not require subsequent heat treatment in the qualified parameter space of an embodiment of the deep learning-based metal matrix composite parameter design method of the present invention.

[0151] This embodiment acquires experimental data on aluminum-based silicon carbide composite materials, constructs training and testing sets using this data, inputs the training set into the model for training, and adjusts the learning rate to obtain a machine learning model. This model achieves high-precision prediction. After obtaining the model, all parameters in the parameter space are input into the model to predict the results. A first preset threshold is used to filter out qualified properties in the predicted results, such as fracture strength and elongation after fracture, which exceed the first preset threshold. The corresponding process parameters are then identified, forming a qualified parameter space. This invention uses deep learning to obtain a machine learning model and then uses this model to predict the qualified parameter space, shortening the R&D cycle of metal matrix composite materials and improving R&D efficiency.

[0152] This invention also provides a device for designing parameters of metal matrix composite materials based on deep learning. For example... Figure 10 As shown, Figure 10 This is a functional module diagram of an embodiment of the metal matrix composite parameter design method based on deep learning of the present invention.

[0153] The deep learning-based parameter design device for metal matrix composite materials of this invention includes:

[0154] The process parameter module 10 is used to obtain the process parameters of metal matrix composites and construct the process parameter space;

[0155] The machine learning module 20 is used to input the parameters in the process parameter space to a pre-built machine learning model to obtain mechanical performance prediction results;

[0156] The filtering module 30 is used to select performance-compliant results from the mechanical performance prediction results according to a first preset threshold.

[0157] The qualified parameter module 40 is used to obtain the qualified process parameters corresponding to the performance compliance results and combine them to obtain the qualified parameter space.

[0158] Optionally, the machine learning module is further used for:

[0159] Select a preset number of process parameter combinations from the process parameter space, and obtain the experimental mechanical properties corresponding to the process parameter combinations;

[0160] Training and testing sets are constructed by combining the experimental mechanical properties and the process parameters.

[0161] The first initial model is trained using the training set, and the accuracy of the first initial model is verified using the test set. When the accuracy meets a preset condition, the machine learning model is obtained.

[0162] Optionally, the machine learning module is further used for:

[0163] The training set is input into the gated residual network and combined with the parameter selection network to train a second initial model;

[0164] Calculate the accuracy and loss function of the second initial model, and adjust the hyperparameters of the trained second initial model based on the accuracy and loss function to obtain the first initial model after hyperparameter tuning.

[0165] Optionally, the machine learning module is further used for:

[0166] Iteratively calculate the loss function and accuracy of the second initial model;

[0167] The process of reducing the learning rate to a preset factor when the loss function converges is repeated until the accuracy reaches a second preset threshold. Then, the iteration stops and the first initial model after parameter tuning is obtained.

[0168] Optionally, the machine learning module is further used for:

[0169] Based on all parameters in the process parameter space, a completely orthogonal parameter space is constructed;

[0170] The fully orthogonal parameter space is input into a pre-built machine learning model.

[0171] Optionally, the filtering module is further configured to:

[0172] The first predicted performance compliance probability corresponding to fracture strength and the second predicted performance compliance probability corresponding to post-fracture elongation are obtained from the mechanical performance prediction results and compared with the first preset threshold, respectively.

[0173] Mechanical performance prediction results in which both the first predicted performance compliance probability and the second predicted performance compliance probability are greater than the first preset threshold are selected as performance compliance results.

[0174] Optionally, the qualification parameter module is further used for:

[0175] Based on preset conditions, process parameters that meet the preset conditions are selected from the qualified process parameter space to construct an application process parameter space, which is used for the research and development of metal matrix composite products.

[0176] The present invention also provides a medium.

[0177] The present invention stores a deep learning-based metal matrix composite parameter design program on the medium. When the deep learning-based metal matrix composite parameter design program is executed by the processor, it implements the steps of the deep learning-based metal matrix composite parameter design method as described above.

[0178] The method implemented when the deep learning-based metal matrix composite parameter design program running on the processor is executed can be referred to in various embodiments of the deep learning-based metal matrix composite parameter design method of the present invention, and will not be repeated here.

[0179] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0180] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0181] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0182] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for parameter design of metal matrix composite materials based on deep learning, characterized in that, The method for designing parameters of the metal matrix composite material includes the following steps: Obtain the process parameters of metal matrix composites and construct the process parameter space; The parameters in the process parameter space are fed into a pre-built machine learning model to obtain mechanical property prediction results. The machine learning model is trained using a gated residual network and a parameter selection network. Based on a first preset threshold, select performance-compliant results from the predicted mechanical properties; Obtain the qualified process parameters corresponding to the performance compliance results, and combine them to obtain the qualified parameter space; The steps of feeding the parameters in the process parameter space into a pre-built machine learning model include: Based on all parameters in the process parameter space, a completely orthogonal parameter space is constructed; The fully orthogonal parameter space is input into a pre-built machine learning model; The construction of a completely orthogonal parameter space based on all parameters of the process parameter space includes: Based on the parameters in the process parameter space, the index level of the research factor is determined, where the research factor refers to the parameter as the research object, and the level is the possible value of a factor. The orthogonal parameter space is constructed by arranging and combining the types and values ​​of each parameter; After the step of obtaining the qualified process parameters corresponding to the performance compliance result and combining them to obtain the qualified parameter space, the method further includes: Based on preset conditions, process parameters that meet the preset conditions are selected from the qualified parameter space to construct an application process parameter space, which is used for the research and development of metal matrix composite products.

2. The method for parameter design of metal matrix composites based on deep learning as described in claim 1, characterized in that, Before the step of feeding the parameters in the process parameter space to a pre-built machine learning model to obtain the mechanical property prediction results, the method further includes: Select a preset number of process parameter combinations from the process parameter space, and obtain the experimental mechanical properties corresponding to the process parameter combinations; Training and testing sets are constructed by combining the experimental mechanical properties and the process parameters. The first initial model is trained using the training set, and the accuracy of the first initial model is verified using the test set. When the accuracy meets a preset condition, the machine learning model is obtained.

3. The method for designing parameters of metal matrix composites based on deep learning as described in claim 2, characterized in that, The step of training the first initial model using the training set includes: The training set is input into the gated residual network and combined with the parameter selection network to train a second initial model; Calculate the accuracy and loss function of the second initial model, and adjust the hyperparameters of the second initial model according to the accuracy and loss function to obtain the first initial model after hyperparameter tuning.

4. The method for parameter design of metal matrix composites based on deep learning as described in claim 3, characterized in that, The step of adjusting the hyperparameters of the second initial model according to the accuracy and the loss function to obtain the first initial model after hyperparameter tuning includes: Iteratively calculate the loss function and accuracy of the second initial model; The process of reducing the learning rate to a preset factor when the loss function converges is repeated until the accuracy reaches a second preset threshold. Then, the iteration stops and the first initial model after parameter tuning is obtained.

5. The method for parameter design of metal matrix composites based on deep learning as described in claim 1, characterized in that, The mechanical properties include: fracture strength and elongation after fracture. The step of selecting the performance compliance result from the predicted mechanical properties according to the first preset threshold includes: The first predicted performance compliance probability corresponding to fracture strength and the second predicted performance compliance probability corresponding to post-fracture elongation are obtained from the mechanical performance prediction results and compared with the first preset threshold, respectively. Mechanical performance prediction results in which both the first predicted performance compliance probability and the second predicted performance compliance probability are greater than the first preset threshold are selected as performance compliance results.

6. A deep learning-based parameter design device for metal matrix composite materials, characterized in that, The device includes: a memory, a processor, and a deep learning-based metal matrix composite parameter design program stored in the memory and executable on the processor, the deep learning-based metal matrix composite parameter design program being configured to implement the steps of the deep learning-based metal matrix composite parameter design method as described in any one of claims 1 to 5.

7. A medium, characterized in that, The medium stores a deep learning-based metal matrix composite parameter design program, which, when executed by a processor, implements the steps of the deep learning-based metal matrix composite parameter design method as described in any one of claims 1 to 5.

Images