Machine learning based additive manufacturing metal material reverse design method

Abstract

The application relates to a kind of additive manufacturing metal material reverse design methods based on machine learning, comprising: according to historical data, source domain, target domain dataset is established;Source domain performance prediction model is established and trained;Migration model parameter, target domain performance prediction model is established and trained;Component design model is established and trained;Input target performance and requirement, and obtain initial material composition;Initial material composition is input into target domain performance prediction model to predict performance;Error of predicted performance and target performance is judged, if not in the allowable range, re-establish model, otherwise, material composition design is completed;According to the designed component, the required form material is prepared, and the performance is verified by metal additive manufacturing technology shaping.The present application mines the implicit complex relationship between material "composition-performance" by machine learning technology when existing metal composition and performance data are less, realizes the additive manufacturing metal material design oriented to performance requirement, and improves the research and development efficiency of new material.

Description

Technical Field

[0001] This invention relates to the field of metal material design technology, and in particular to a reverse design method for additive manufacturing metal materials based on machine learning. Background Technology

[0002] With the development of society and science and technology, additive manufacturing technology, as an advanced forming method, has attracted much attention in high-end manufacturing fields such as medical devices, electronic information, communication technology, aerospace, transportation, and energy equipment due to its ability to achieve personalized precision manufacturing. However, the development of additive manufacturing processes typically faces two key challenges. The first is the design of the forming material composition. In high-end manufacturing, there are often not only structural requirements for the forming process but also performance requirements for the materials. Furthermore, the market's demands for the performance of key metal materials such as aluminum alloys, titanium alloys, magnesium alloys, copper alloys, and high-temperature alloys are constantly increasing. The second challenge is exploring forming processes when faced with new materials or performance requirements. Solving both of these problems requires costly and time-consuming trial-and-error experiments, which significantly limits the development of metal additive manufacturing technology.

[0003] Therefore, a key step in the development of metal additive manufacturing technology is to find a new method for designing the composition of metal materials that is different from traditional material design methods and is geared towards high-performance requirements, so as to significantly improve the efficiency of new material research and development and engineering applications.

[0004] To date, the composition design of metallic materials has primarily relied on experimental trial-and-error methods, or empirical methods. Researchers roughly determine the metal composition based on their experience, then spend considerable time and resources conducting various experiments to finalize the composition. This trial-and-error approach lacks a model that can quantitatively describe the relationship between material composition and properties, making it even more difficult to design metal compositions quickly and accurately according to given performance requirements. This results in low efficiency, high cost, and long development cycles for new materials. Developing methods for quickly and accurately designing the composition of metallic materials according to given performance requirements has become a pressing problem in the field of metallic materials design.

[0005] Machine learning, as an emerging data-driven computing method, learns knowledge from the model building process of a material with a large amount of data (source domain) and transfers the relevant knowledge to the model building process of a target material with a small amount of data (target domain), so that the target domain can achieve better learning results. At the same time, the use of ensemble learning algorithms can reduce model variance and improve model generalization ability, stability and accuracy by having multiple base learners cooperate in various ways.

[0006] Therefore, it is necessary to design a reverse design method for additive manufacturing metallic materials based on ensemble learning and transfer learning to solve the above problems. Summary of the Invention

[0007] Based on this, a reverse design method for additive manufacturing metal materials based on ensemble learning and transfer learning is proposed. The method transfers the knowledge obtained in solving the source domain and source task to the target domain, realizes the mining of the implicit complex relationship between metal "composition-performance" under small data conditions, and uses optimization algorithms to achieve fast and accurate reverse design of metal composition according to performance requirements.

[0008] A machine learning-based reverse design method for additive manufacturing metallic materials, characterized by the following steps: S1: Establishing a source domain dataset and a target domain dataset based on historical data; S2: Establishing and training a source domain material performance prediction model from metal composition to properties based on the source domain dataset; S3: Transferring the parameters of the source domain material performance prediction model, establishing and training a target domain performance prediction model from metal composition to properties based on the source domain material performance prediction model parameters and the target domain dataset; S4: Establishing and training a reverse design model from metal properties to composition based on the target domain dataset using K-fold cross-validation; S5: S6: Input the target metal properties into the reverse design model from metal properties to composition to obtain the initial metal composition; S7: Input the initial metal composition into the target domain performance prediction model to obtain the predicted performance of the metal composition; S8: Compare the predicted performance with the actual performance requirements, calculate the relative error. If the error is not within the allowable range, adjust the parameters of the reverse design model in S4, retrain, and repeat S5-S6; if the error is within the allowable range, the metal composition design is completed, and the final metal composition is obtained; S9: Prepare the metal material required for forming according to the final metal composition, use a metal additive manufacturing method for forming, and verify the metal properties through experiments.

[0009] In one embodiment, the specific steps of S1 include: S11, collecting historical metal data to establish a basic dataset, wherein the historical metal data includes metal composition data and metal performance data, and the basic dataset includes real experimental data and simulation data; S12, organizing and classifying the basic dataset, selecting material data with sufficient data volume that are close to the target metal composition or performance to establish a source domain dataset, and establishing a target domain dataset, and normalizing all initial datasets to obtain the dataset.

[0010] In one embodiment, the metal property data in S11 includes at least one or more combinations of mechanical properties such as tensile strength, compressive yield strength, elongation at break, reduction of area, and hardness, as well as spectral absorption capacity, biodegradability, and biocompatibility.

[0011] In one embodiment, the metal composition data in S11 includes the types of elements contained in the metal and their corresponding contents.

[0012] In one embodiment, the metal in S11 includes one or more of aluminum alloy, titanium alloy, zinc alloy, magnesium alloy, iron alloy, nickel alloy, and copper alloy.

[0013] In one embodiment, the specific steps of S2 include: S21, taking the metal composition data in the source domain dataset as input and the metal performance data as output, selecting an ensemble learning model, using a nonlinear algorithm as the base learner, setting the ensemble learning model parameters, and establishing the source domain material performance prediction model from metal composition to performance; S22, selecting a transfer function and the corresponding training method, training the source domain material performance prediction model until the model accuracy reaches a preset value.

[0014] In one embodiment, the ensemble learning model includes, but is not limited to, the Stacking method.

[0015] In one embodiment, the training method includes one or more of the following: steepest descent, adaptive learning rate, momentum-adaptive learning rate adjustment, additional momentum, and quantized conjugate gradient.

[0016] In one embodiment, the accuracy of the source domain material performance prediction model is not less than 0.9.

[0017] In one embodiment, before training the source domain material performance prediction model, the source domain dataset is divided into a training set, a validation set, and a test set, with the sum of the proportions of the training set, validation set, and test set in the source domain dataset being 1. The aforementioned source domain dataset should be randomly divided into a training set (60%–90%), a validation set (5%–20%), and a test set (5%–20%) before model training begins, with the sum of the proportions of the training set, validation set, and test set always being 1. The training set is used to train and establish a source domain material performance prediction model from metal composition to properties; the validation set is used to prevent model overfitting; these two datasets are used during the training of the source domain material performance prediction model; and the test set is used after model training is completed to test the model's accuracy.

[0018] Preferably, the training set accounts for 60% to 90% of the source domain dataset, the validation set accounts for 5% to 20% of the source domain dataset, and the test set accounts for 5% to 20% of the source domain dataset.

[0019] In one embodiment, the specific steps of S3 include: S31: inputting the source domain dataset into the source domain material performance prediction model to obtain predicted performance data; S32: calculating the relative error between the predicted performance data and the actual performance; if the error is not within the allowable range, adjusting the parameters of the source domain material performance prediction model and re-entering step S2; if the error is within the allowable range, migrating the parameters of the source domain material performance prediction model to the target domain performance prediction model; S33: establishing a target domain performance prediction model from metal composition to performance based on the target domain dataset and the migrated parameters of the source domain material performance prediction model, and training it, making certain adjustments to the parameters.

[0020] In one embodiment, the specific steps of S32 include: S321: comparing the predicted performance data with the actual performance curve, calculating and determining whether the relative error is within the allowable range; S322: if the error is not within the allowable range, readjusting the source domain material performance prediction model of S2 and retraining; if the error is within the allowable range, establishing a feature distribution transformation matching network to transfer the effective parameters of the source domain material performance prediction model to the target domain performance prediction model.

[0021] In one embodiment, the relative error of S32 is calculated as: (target performance - predicted performance) / target performance; the allowable error range in S32 is 0-10%.

[0022] In one embodiment, the specific steps of S33 include: S331: using the transferred source domain material performance prediction model parameters to establish an ensemble learning model of the same type as the source domain material performance prediction model parameters, as an initial target domain performance prediction model; S332: taking the metal composition data in the processed target domain dataset as input and the metal performance data as output, and appropriately adjusting the target domain performance prediction model parameters.

[0023] In one embodiment, the accuracy of the target domain performance prediction model is not less than 0.9.

[0024] In one embodiment, the target domain performance prediction model parameters include, but are not limited to, the number of iterations, learning rate, base learner parameters, random sampling or feature selection parameters.

[0025] In one embodiment, the specific steps of S7 include: S71: comparing the predicted performance with the actual performance requirements, calculating and determining whether the relative error is within the allowable range; S72: if the error is within the allowable range, performing denormalization on the metal composition of the inverse design model from metal properties to composition to complete the composition design of the metal material; if the error is not within the allowable range, adjusting the parameters of the inverse design model from metal properties to composition and retraining. The above-mentioned normalization and denormalization are common data processing methods and commonly used algorithms in mathematics. The specific process will not be elaborated, but its purpose is to transform the data in terms of scale without changing the original data distribution pattern.

[0026] In one embodiment, the relative error calculation method in S71 is: (target performance - predicted performance) / target performance; the allowable error range in S9 is within 10%.

[0027] In one embodiment, the metal material required for forming includes powder or filament.

[0028] In one embodiment, the metal additive manufacturing method includes selective laser melting, laser-directed energy deposition, and electron beam melting.

[0029] In one embodiment, the experiment is designed according to the performance requirements of the metal, including density, mechanical properties and biological performance tests.

[0030] In general, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0031] 1. Enables rapid and accurate design of new multi-component metals based on data-driven machine learning models and performance requirements through additive manufacturing. In particular, it can complete the design of new materials with limited historical data, thus significantly improving the efficiency of new material research and development and meeting the needs of various sectors of society for new high-performance materials.

[0032] 2. To achieve in-depth mining of existing data, establish a model of the relationship between "composition and performance" of metallic materials, further clarify the connection between "composition and performance" of materials, so as to enable the rapid and accurate design of metallic compositions according to given performance requirements, promote the development of materials informatics and improve the level of materials research and development. Attached Figure Description

[0033] Figure 1 A flowchart illustrating a machine learning-based additive manufacturing metal material reverse design method according to an embodiment of the present invention;

[0034] Figure 2This is a schematic diagram of the specific process S1 provided in one embodiment of the present invention;

[0035] Figure 3 This is a schematic diagram of the specific process S2 provided in one embodiment of the present invention;

[0036] Figure 4 This is a schematic diagram of the specific process of S3 provided in one embodiment of the present invention;

[0037] Figure 5 This is a schematic diagram of the specific process of S32 provided in one embodiment of the present invention;

[0038] Figure 6 This is a schematic diagram of the specific process of S33 provided in one embodiment of the present invention;

[0039] Figure 7 This is a schematic diagram of the specific process of S7 provided in one embodiment of the present invention;

[0040] Figure 8 A flowchart illustrating a specific implementation of a machine learning-based additive manufacturing metal material reverse design method according to an embodiment of the present invention;

[0041] Figure 9 This is a model framework diagram provided for one embodiment of the present invention. Detailed Implementation

[0042] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0043] 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.

[0044] The following describes, with reference to the accompanying drawings, some embodiments of the present invention, a reverse design method for additive manufacturing metal materials based on machine learning.

[0045] like Figures 1 to 9As shown, this embodiment discloses a machine learning-based reverse design method for additive manufacturing metallic materials, characterized by the following steps: S1: Establishing a source domain dataset and a target domain dataset based on historical data; S2: Establishing and training a source domain material performance prediction model from metal composition to properties based on the source domain dataset; S3: Transferring the parameters of the source domain material performance prediction model, establishing and training a target domain performance prediction model from metal composition to properties based on the source domain material performance prediction model parameters and the target domain dataset; S4: Establishing and training a reverse design model from metal properties to composition based on the target domain dataset using K-fold cross-validation. S5: Input the target metal properties into the reverse design model from metal properties to composition to obtain the initial metal composition; S6: Input the initial metal composition into the target domain performance prediction model to obtain the predicted performance of the metal composition; S7: Compare the predicted performance with the actual performance requirements, calculate the relative error. If the error is not within the allowable range, adjust the parameters of the reverse design model in S4, retrain, and repeat S5-S6; if the error is within the allowable range, the metal composition design is completed, and the final metal composition is obtained; S8: Prepare the metal material required for forming according to the final metal composition, use the metal additive manufacturing method for forming, and verify the metal properties through experiments.

[0046] like Figure 1 and Figure 2 As shown, in addition to the features of the above embodiments, this embodiment further defines: the specific steps of S1 include S11: collecting historical metal data to establish a basic dataset, the historical metal data including metal composition data and metal performance data, the basic dataset including real experimental data and simulation data; S12: organizing and classifying the basic dataset, selecting material data that is close to the target metal composition or performance and has sufficient data volume to establish a source domain dataset, and establishing a target domain dataset, normalizing all initial datasets to obtain a dataset.

[0047] In addition to the features of the above embodiments, this embodiment further specifies that: the metal performance data in S11 includes at least one or a combination of mechanical properties such as tensile strength, compressive yield strength, elongation at break, reduction of area, and hardness, as well as spectral absorption capacity, biodegradability, and biocompatibility.

[0048] In addition to the features of the above embodiments, this embodiment further specifies that: the metal composition data in S11 includes the types of elements contained in the metal and their corresponding contents.

[0049] In addition to the features of the above embodiments, this embodiment further specifies that the metal in S11 includes one or more of aluminum alloy, titanium alloy, zinc alloy, magnesium alloy, iron alloy, nickel alloy, and copper alloy.

[0050] like Figure 1and Figure 3 As shown, in addition to the features of the above embodiments, this embodiment further defines: the specific steps of S2 include S21, taking the metal composition data in the source domain dataset as input and the metal performance data as output, selecting an ensemble learning model, selecting a nonlinear algorithm as the base learner, setting the parameters of the ensemble learning model, and establishing a source domain material performance prediction model from metal composition to performance; S22, selecting a transfer function and the corresponding training method, training the source domain material performance prediction model until the model accuracy reaches a preset value.

[0051] In addition to the features of the above embodiments, this embodiment further defines that the ensemble learning model includes, but is not limited to, the Stacking method.

[0052] In addition to the features of the above embodiments, this embodiment further specifies that the training method includes one or more of the following: steepest descent method, adaptive learning rate method, momentum-adaptive learning rate adjustment method, additional momentum method, and quantized conjugate gradient method.

[0053] In addition to the features of the above embodiments, this embodiment further specifies that the accuracy of the source domain material performance prediction model is not less than 0.9.

[0054] like Figure 1 , Figure 3 and 8 As shown, in addition to the features of the above embodiments, this embodiment further specifies that: before the source domain material performance prediction model begins training, the source domain dataset is divided into a training set, a validation set, and a test set, with the sum of the proportions of the training set, validation set, and test set in the source domain dataset being 1. The source domain dataset should be randomly divided into a training set (60%–90%), a validation set (5%–20%), and a test set (5%–20%) before model training begins, with the sum of the proportions of the training set, validation set, and test set always being 1. The training set is used to train and establish a source domain material performance prediction model from metal composition to performance; the validation set is used to prevent model overfitting; these two datasets are used when training the source domain material performance prediction model; and the test set is used after model training is completed to test model accuracy.

[0055] In addition to the features of the above embodiments, this embodiment further specifies that: the proportion of the training set to the source domain dataset ranges from 60% to 90%, the proportion of the validation set to the source domain dataset ranges from 5% to 20%, and the proportion of the test set to the source domain dataset ranges from 5% to 20%.

[0056] like Figure 1 and Figure 4As shown, in addition to the features of the above embodiments, this embodiment further defines: the specific steps of S3 include S31: inputting the source domain dataset into the source domain material performance prediction model to obtain predicted performance data; S32: calculating the relative error between the predicted performance data and the actual performance; if the error is not within the allowable range, adjusting the parameters of the source domain material performance prediction model and re-entering step S2; if the error is within the allowable range, migrating the parameters of the source domain material performance prediction model to the target domain performance prediction model; S33: establishing a target domain performance prediction model from metal composition to performance based on the target domain dataset and the migrated parameters of the source domain material performance prediction model, and training it, making certain adjustments to the parameters.

[0057] like Figure 5 As shown, in addition to the features of the above embodiments, this embodiment further defines: the specific steps of S32 include S321: comparing the predicted performance data with the actual performance curve, calculating and determining whether the relative error is within the allowable range; S322: if the error is not within the allowable range, then readjusting the source domain material performance prediction model of S2 and retraining; if the error is within the allowable range, then establishing a feature distribution transformation matching network to transfer the effective parameters of the source domain material performance prediction model to the target domain performance prediction model.

[0058] In addition to the features of the above embodiments, this embodiment further specifies that: the relative error calculation method for S32 is: (target performance - predicted performance) / target performance; the allowable error range in S32 is 0-10%.

[0059] like Figure 6 As shown, in addition to the features of the above embodiments, this embodiment further defines: the specific steps of S33 include S331: using the transferred source domain material performance prediction model parameters to establish an integrated learning model of the same type as the source domain material performance prediction model parameters, as the initial target domain performance prediction model; S332: taking the metal composition data in the processed target domain dataset as input and the metal performance data as output, and appropriately adjusting the target domain performance prediction model parameters.

[0060] In addition to the features of the above embodiments, this embodiment further specifies that the accuracy of the target domain performance prediction model is not less than 0.9.

[0061] In addition to the features of the above embodiments, this embodiment further defines the target domain performance prediction model parameters as including but not limited to the number of iterations, learning rate, base learner parameters, random sampling or feature selection parameters.

[0062] like Figure 1 and Figure 7As shown, in addition to the features of the above embodiments, this embodiment further defines: the specific steps of S7 include S71: comparing the predicted performance with the actual performance requirements, calculating and determining whether the relative error is within the allowable range; S72: if the error is within the allowable range, then performing inverse normalization on the metal composition of the inverse design model from metal properties to composition to complete the composition design of the metal material; if the error is not within the allowable range, then adjusting the parameters of the inverse design model from metal properties to composition and retraining. Normalization and inverse normalization are common data processing methods and are commonly used algorithms in mathematics. The specific process is not elaborated here, but their purpose is to transform the data in terms of scale without changing the original data distribution pattern.

[0063] In addition to the features of the above embodiments, this embodiment further defines: the relative error calculation method in S71 is: (target performance - predicted performance) / target performance; the allowable error range in S9 is within 10%.

[0064] In addition to the features of the above embodiments, this embodiment further specifies that the metal material required for forming includes powder or wire.

[0065] In addition to the features of the above embodiments, this embodiment further defines the metal additive manufacturing method as including laser selective melting technology, laser directional energy deposition technology, and electron beam melting technology.

[0066] In addition to the features of the above embodiments, this embodiment further specifies that the experiment is designed according to the performance requirements of the metal, including density, mechanical properties and biological performance tests.

[0067] Example 1

[0068] Using the method provided by this invention, with performance requirements of an elastic modulus less than 100 GPa and an elongation greater than 15%, and under the condition that the performance error is no greater than 10%, the composition of a low-modulus ternary Ti-Nb-Zr metal is designed. The specific design method is described as follows:

[0069] (1) Historical data collection: Collect and organize publicly available data on the composition, elastic modulus, and elongation of titanium to establish a basic dataset for the source domain; collect and organize publicly available data on the composition, elastic modulus, and elongation of Ti-Nb-Zr metals to establish a basic dataset for the target domain.

[0070] (2) Data processing: The collected datasets are classified and filtered according to their correlation with the target metal, and the initial datasets of the source domain and target domain are established. The metal composition data and performance data in the initial datasets are normalized respectively.

[0071] (3) Using Stacking ensemble learning method to establish source domain metal performance prediction model: Select Stacking ensemble learning algorithm, select KNN and SVR algorithms as base learners, use logistic regression as secondary learner, take the material composition data of the processed source domain dataset as input of base learner, take material performance as output, take the output result of base learner as input of secondary learner, balance the weight of base learner result, and obtain the final prediction performance as output to establish ensemble learning model for source domain material performance prediction;

[0072] KNN stands for K-Nearest Neighbors algorithm, and its algorithm flow is as follows:

[0073] 1) For a set of data X(i) of unknown category, calculate its distance to each set of data in the sample (the sample contains n sets of data, and each set of data corresponds to a category; the indicators of the sample are consistent with the indicators of the unknown data). The formula for the Euclidean distance between data X(i) and X(j) is:

[0074]

[0075] p represents the number of indicators (variables).

[0076] 2) Arrange the sample data in ascending order of distance;

[0077] 3) Select the k most recent sets of data;

[0078] 4) Count which category the k groups of data belong to most often. The category of the unknown data is the category with the most occurrences in the k groups of data.

[0079] SVR stands for Support Vector Regression, and its objective function is:

[0080]

[0081]

[0082] Where f(x) = ωx + b is the final objective model function. ∈ represents a slack variable, and ∈ represents the difference between functions.

[0083] (4) Training the source domain material property prediction model: The Sigmoid function is used as the hidden layer transfer function and the linear function is used as the output layer transfer function. The steepest descent method is used to train the prediction model so that the model accuracy is not less than 0.9.

[0084] (5) Parameter transfer: Establish a feature matching network and transfer the trained source domain prediction model parameters to the target domain prediction model in the form of transfer learning, as the initial parameters of the target domain prediction model.

[0085] (6) Using the Stacking ensemble learning method to establish a target domain metal performance prediction model: Select the Stacking ensemble learning model, use KNN and SVR algorithms as base learners, use logistic regression as a secondary learner, and use the transferred source domain prediction model parameters as parameters of the target domain initial performance prediction model. Then, use the processed target domain dataset material composition data as input to the base learner, and the material performance data as output. Use the output of the base learner as input to the secondary learner. After balancing the weights of the base learner results, the final prediction performance is obtained as output. Calculate the error between the prediction performance and the actual performance, adjust the transferred model parameters appropriately, and establish an ensemble learning model for predicting the material performance of the target domain.

[0086] (7) Training the target domain material property prediction model: The Sigmoid function is used as the hidden layer transfer function and the linear function is used as the output layer transfer function. The steepest descent method is used to train the prediction model so that the model accuracy is not less than 0.9.

[0087] (8) Use the NSGA-II algorithm to establish a material composition inverse design model: Use the NSGA-II algorithm to construct a material composition inverse design model, with the processed target domain performance data as input and the composition data as output;

[0088] The NSGA-II algorithm is as follows:

[0089]

[0090]

[0091] (9) Training the target domain inverse design model: Using the processed target domain dataset, select the Sigmoid function as the hidden layer transfer function and the linear function as the output layer transfer function, and use the steepest descent method to train the inverse design model so that the model accuracy is not less than 0.9;

[0092] The Sigmoid function can be represented as:

[0093]

[0094] Where e is the natural constant, z is the independent variable, and σ is the dependent variable.

[0095] (10) Input the performance requirements of elastic modulus less than 100 GPa and elongation greater than 15% into the reverse design model and restrict it to Ti-Nb-Zr metal to obtain the initial design composition;

[0096] (11) Use the initial material composition obtained in step (10) as the input of the material property prediction model trained in step (7) and output the predicted performance of the initial material composition.

[0097] (12) Compare the predicted performance obtained in step (11) with the given performance target, and calculate the relative error e. r = (Target performance - Predicted performance) / Target performance;

[0098] (13) If e r If e ≤10%, then the required metal composition is obtained; if e r >10%, readjust the parameters of the reverse design model in step (8), and train in step (9). Repeat steps (10) to (12) until the material composition that meets the requirements of error and other aspects is obtained.

[0099] (14) Perform the same reverse normalization process as in step (2) on the metal composition determined in step (13) to obtain the final metal composition and complete the metal design, as shown in Table 1 below:

[0100] Table 1

[0101] serial number Elastic modulus / GPa Elongation / % Predicted metal composition / wt% 1 82.2 17.4 Ti-22Zr-22Nb 2 77.3 15.5 Ti-22Zr-14.8Nb 3 74.7 15.9 Ti-24.6Zr-11.6Nb

[0102] (15) Experimental verification: Metal powder was prepared by gas atomization according to the component ratio, and formed by laser selective melting technology. The density of the sample was tested, and the elastic modulus and elongation of the formed specimen were tested.

[0103] Example 2

[0104] Using the method provided by this invention, with a yield strength of 200 MPa and an elongation of 10% as performance requirements, and under the condition that the performance error is no greater than 10%, the composition of a high-strength zinc alloy is designed. The specific design method is described as follows:

[0105] (1) Historical data collection: Collect and organize publicly available data on the composition, yield strength, and elongation of zinc metal to establish a basic dataset for the source domain; collect and organize publicly available data on the composition, yield strength, and elongation of multi-element zinc alloys to establish a basic dataset for the target domain.

[0106] (2) Data processing: The collected datasets are classified and filtered according to their correlation with the target metal, and the initial datasets of the source domain and target domain are established. The metal composition data and performance data in the initial datasets are normalized respectively.

[0107] (3) Using Stacking ensemble learning method to establish source domain metal performance prediction model: Select Stacking ensemble learning algorithm, select KNN and decision tree algorithm as base learner, use logistic regression as secondary learner, take the material composition data of the processed source domain dataset as input of base learner, take material performance as output, take the output result of base learner as input of secondary learner, balance the weight of base learner result, and get the final predicted performance as output to establish ensemble learning model for source domain material performance prediction;

[0108] KNN stands for K-Nearest Neighbors algorithm, and its algorithm flow is as follows:

[0109] 1) For a set of data X(i) of unknown category, calculate its distance to each set of data in the sample (the sample contains n sets of data, and each set of data corresponds to a category; the indicators of the sample are consistent with the indicators of the unknown data). The formula for the Euclidean distance between data X(i) and X(j) is:

[0110]

[0111] p represents the number of indicators (variables);

[0112] 2) Arrange the sample data in ascending order of distance;

[0113] 3) Select the k most recent sets of data;

[0114] 4) Count which category the k groups of data belong to most often. The category of the unknown data is the category with the most occurrences in the k groups of data.

[0115] The decision tree algorithm can be represented as:

[0116]

[0117] Where N is the total number of samples, N L N is the number of samples in the left subset. R Let y be the number of samples in the right subset. i This is the tag value.

[0118] (4) Training the source domain material property prediction model: The Sigmoid function is used as the hidden layer transfer function and the linear function is used as the output layer transfer function. The steepest descent method is used to train the prediction model so that the model accuracy is not less than 0.9.

[0119] The Sigmoid function can be represented as:

[0120]

[0121] Where e is the natural constant, z is the independent variable, and σ is the dependent variable.

[0122] (5) Parameter transfer: Establish a feature matching network and transfer the trained source domain prediction model parameters to the target domain prediction model in the form of transfer learning, as the initial parameters of the target domain prediction model.

[0123] (6) Using the Stacking ensemble learning method to establish a target domain metal performance prediction model: Select the Stacking ensemble learning model, use KNN and decision tree algorithms as base learners, use logistic regression as a secondary learner, and use the transferred source domain prediction model parameters as parameters of the target domain initial performance prediction model. Then, use the processed target domain dataset material composition data as input to the base learner, and the material performance data as output. Use the output of the base learner as input to the secondary learner. After balancing the weights of the base learner results, the final prediction performance is obtained as output. Calculate the error between the prediction performance and the actual performance, adjust the transferred model parameters appropriately, and establish an ensemble learning model for predicting the material performance of the target domain.

[0124] (7) Training the target domain material property prediction model: The Sigmoid function is used as the hidden layer transfer function and the linear function is used as the output layer transfer function. The steepest descent method is used to train the prediction model so that the model accuracy is not less than 0.9.

[0125] (8) Use the NSGA-II algorithm to establish a material composition inverse design model: Use the NSGA-II algorithm to construct a material composition inverse design model, with the processed target domain performance data as input and the composition data as output;

[0126] (9) Training the target domain inverse design model: Using the processed target domain dataset, select the Sigmoid function as the hidden layer transfer function and the linear function as the output layer transfer function, and use the steepest descent method to train the inverse design model so that the model accuracy is not less than 0.9;

[0127] (10) Input the performance requirements of yield strength of 200MPa and elongation of 10% into the reverse design model to obtain the initial design composition;

[0128] (11) Use the initial material composition obtained in step (10) as the input of the material property prediction model trained in step (7) and output the predicted performance of the initial material composition.

[0129] (12) Compare the predicted performance obtained in step (11) with the given performance target, and calculate the relative error e. r = (Target performance - Predicted performance) / Target performance;

[0130] (13) If e r If e ≤10%, then the required metal composition is obtained; if e r>10%, readjust the parameters of the reverse design model in step (8), and train in step (9). Repeat steps (10) to (12) until the material composition that meets the requirements of error and other aspects is obtained.

[0131] (14) Perform the same reverse normalization process as step (2) on the metal composition determined in step (13) to obtain the final metal composition and complete the metal design, as shown in Table 2 below;

[0132] Table 2

[0133] serial number Yield strength / MPa Elongation / % Predicted metal composition / wt% 1 213 26.3 Zn-0.018Mg-0.023Cu 2 198.2 15.3 Zn-0.35Mn-0.41Cu 3 241.3 9.7 Zn-1.06Mg-0.4Zr

[0134] (15) Experimental verification: Metal powder was prepared by gas atomization according to the component ratio, and printed by laser selective melting technology. The density of the sample was tested, and the yield strength and elongation of the printed specimen were tested.

[0135] As can be seen from the above description, the embodiments of the present invention achieve the following technical effects:

[0136] 1. Enables rapid and accurate design of new multi-component metals based on data-driven machine learning models and performance requirements through additive manufacturing. In particular, it can complete the design of new materials with limited historical data, thus significantly improving the efficiency of new material research and development and meeting the needs of various sectors of society for new high-performance materials.

[0137] 2. To achieve in-depth mining of existing data, establish a model of the relationship between "composition and performance" of metallic materials, further clarify the connection between "composition and performance" of materials, so as to enable the rapid and accurate design of metallic compositions according to given performance requirements, promote the development of materials informatics and improve the level of materials research and development.

[0138] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0139] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A reverse design method for additive manufacturing metallic materials based on machine learning, characterized in that, Includes the following steps: S1: Establish source domain datasets and target domain datasets based on historical data; S2: Based on the source domain dataset, establish a source domain material performance prediction model from metal composition to performance, and train it; S3: Transfer the parameters of the source domain material performance prediction model, establish a target domain performance prediction model from metal composition to performance based on the source domain material performance prediction model parameters and the target domain dataset, and train it; S4: Based on the target domain dataset, establish an inverse design model from metal properties to composition using K-fold cross-validation, and train it. S5: Input the target metal properties into the inverse design model from metal properties to composition to obtain the initial metal composition; S6: Input the initial metal composition into the target domain performance prediction model to obtain the predicted performance of the metal composition; S7: Compare the predicted performance with the actual performance requirements, calculate the relative error. If the error is not within the allowable range, adjust the parameters of the reverse design model in S4, retrain, and repeat S5-S6. If the error is within the allowable range, the metal composition design is completed, and the final metal composition is obtained. S8: Prepare the metal material required for forming according to the final metal composition, form it using a metal additive manufacturing method, and verify the metal properties through experiments.

2. The machine learning-based additive manufacturing metal material reverse design method according to claim 1, characterized in that, The specific steps of S1 include: S11: Collect historical metal data to establish a basic dataset. The historical metal data includes metal composition data and metal property data. The basic dataset includes real experimental data and simulation data. S12: Organize and classify the basic dataset, select material data with sufficient data volume that are close to the target metal composition or performance to establish the source domain dataset and the target domain dataset, and normalize all the initial datasets to obtain the dataset.

3. The machine learning-based additive manufacturing metal material reverse design method according to claim 2, characterized in that, The metal performance data includes at least one or more combinations of mechanical properties such as tensile strength, compressive yield strength, elongation at break, reduction of area, and hardness; spectral absorption capacity; biodegradability; and / or biocompatibility; and / or The metal composition data includes the types and corresponding contents of the elements contained in the metal; and / or The metal includes one or more of the following: aluminum alloy, titanium alloy, zinc alloy, magnesium alloy, iron alloy, nickel alloy, and copper alloy.

4. The machine learning-based additive manufacturing metal material reverse design method according to claim 1, characterized in that, The specific steps of S2 include: S21: Take the metal composition data in the source domain dataset as input and the metal performance data as output, select an ensemble learning model, the ensemble learning model including but not limited to the Stacking method, select a nonlinear algorithm as the base learner, set the parameters of the ensemble learning model, and establish the source domain material performance prediction model from metal composition to performance. S22: Select a transfer function and a corresponding training method, wherein the training method includes one of the following: steepest descent method, adaptive learning rate method, momentum-adaptive learning rate adjustment method, additional momentum method, and quantized conjugate gradient method, and train the source domain material performance prediction model until the model accuracy is not less than 0.

9.

5. The machine learning-based additive manufacturing metal material reverse design method according to claim 1 or 4, characterized in that, Before training the source domain material performance prediction model, the source domain dataset is divided into a training set, a validation set, and a test set. The sum of the proportions of the training set, validation set, and test set in the source domain dataset is 1. The proportion of the training set in the source domain dataset ranges from 60% to 90%, the proportion of the validation set in the source domain dataset ranges from 5% to 20%, and the proportion of the test set in the source domain dataset ranges from 5% to 20%.

6. The machine learning-based additive manufacturing metal material reverse design method according to claim 1, characterized in that, The specific steps of S3 include: S31: Input the source domain dataset into the source domain material performance prediction model to obtain predicted performance data; S32: Calculate the relative error between the predicted performance data and the actual performance. If the error is not within the allowable range, adjust the parameters of the source domain material performance prediction model and re-enter step S2. If the error is within the allowable range, migrate the parameters of the source domain material performance prediction model to the target domain performance prediction model. S33: Based on the target domain dataset and the parameters of the source domain material performance prediction model transferred from the target domain dataset, establish a target domain performance prediction model from metal composition to performance, train it, and make certain adjustments to the parameters.

7. The machine learning-based additive manufacturing metal material reverse design method according to claim 6, characterized in that, The specific steps of S32 include: S321: Compare the predicted performance data with the actual performance curve, calculate and determine whether the relative error is within the allowable range. The relative error is calculated as: (target performance - predicted performance) / target performance. The allowable range of the relative error is 0-10%. S322: If the relative error is not within the allowable range, then the source domain material performance prediction model of S2 is readjusted and retrained; if the relative error is within the allowable range, then a feature distribution transformation matching network is established to transfer the effective parameters of the source domain material performance prediction model to the target domain performance prediction model.

8. The machine learning-based additive manufacturing metal material reverse design method according to any one of claims 6 to 7, characterized in that, The specific steps of S33 include: S331: Use the transferred source domain material performance prediction model parameters to establish an ensemble learning model of the same type as the source domain material performance prediction model parameters, and select the target domain performance prediction model with an accuracy of not less than 0.9 as the initial target domain performance prediction model. S332: Take the metal composition data in the processed target domain dataset as input and the metal performance data as output, and appropriately adjust the parameters of the target domain performance prediction model. The target domain performance prediction model parameters include, but are not limited to, the number of iterations, the learning rate, the base learner parameters, and random sampling or feature selection parameters.

9. The machine learning-based additive manufacturing metal material reverse design method according to claim 1, characterized in that, The specific steps of S7 include: S71: Compare the predicted performance with the actual performance requirements, calculate and determine whether the relative error is within the allowable range. The relative error is calculated as: (target performance - predicted performance) / target performance. The allowable range of the relative error is within 10%. S72: If the error is within the allowable range, the metal composition of the reverse design model from metal properties to composition is denormalized to complete the composition design of the metal material; if the error is not within the allowable range, the parameters of the reverse design model from metal properties to composition are adjusted and trained again.

10. The machine learning-based additive manufacturing metal material reverse design method according to claim 9, characterized in that, The metal material required for forming includes powder or wire; and / or The metal additive manufacturing method includes one of laser selective melting, laser directed energy deposition, and electron beam melting; and / or The experiment was designed according to the performance requirements of the metal, and included at least density, mechanical properties and biological properties tests.

Images