Small sample alloy design method and system
By constructing multi-process alloy data sets, dimensionality reduction clustering and sampling, and combining multiple models to predict alloy performance, the problem of low data volume of complex processes is solved, and efficient and low-cost alloy composition design is achieved.
Patent Information
- Application Number
- CN202510403021.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-01
- Publication Date
- 2025-07-04
AI Technical Summary
In the design of alloy material that deals with complex processes, the small amount of data makes it difficult for machine learning models to accurately predict alloy performance and cannot obtain high-performance alloy components.
By creating original data sets containing multiple processing processes and alloy components, using machine learning models for training, dimensionality reduction to two-dimensional space for clustering and sampling, new data points are generated, and mapped back to the original space, combining multiple models for performance prediction and preferred alloy components.
Accurate prediction of the performance of complex process alloys is achieved, design costs are reduced, and high-performance alloy components that meet expectations can be quickly found.
Smart Images
Figure CN120260755A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of alloy material design, and particularly to a small-sample alloy design method and system. Background Art
[0002] Traditional alloy materials are mainly designed by the "trial and error method" combined with relevant theoretical calculations (phase field calculation and molecular dynamics). This method consumes huge human, financial and time costs, and it is difficult to explore a relatively comprehensive composition space.
[0003] In recent years, with the development of computer artificial intelligence, machine learning has also been widely applied in the field of alloy material design. Therefore, with the development of machine learning technology, artificial neural networks can quickly learn the relationship between the four elements of materials "composition - structure - process - performance", and can better predict material properties. In addition, this method can also explore the composition space more fully and comprehensively, redesign the alloy composition under the same process, obtain higher performance values, and greatly reduce human, financial and time costs.
[0004] In alloy material design schemes, different service environments have different requirements for the properties of alloy materials, and the required processing technologies for materials are also different, resulting in a large gap in relevant process data. Most of the existing alloy design methods at present are to train machine learning models by collecting single-process data, realize the prediction of the properties of alloys in a single process, and design high-performance alloys under this process based on the trained machine learning model. However, some processes are relatively complex, and their time, human and financial costs are very high, so the alloys made from these processes are relatively few, and thus the data volume is small, resulting in it being difficult for machine learning models to accurately predict the properties of alloys in small-sample processes, and it is impossible to obtain alloy compositions with higher performance values based on this model. Summary of the Invention
[0005] Based on the defects existing in the above-mentioned prior art, the present invention provides a small-sample alloy design method and system, which solves the problem that some processes are relatively complex, their time, human and financial costs are very high, so the alloys made from these processes are relatively few, and thus the data volume is small, resulting in it being difficult for machine learning models to accurately predict the properties of alloys in small-sample processes, and it is impossible to obtain high-performance alloys based on this model.
[0006] The present invention adopts the following technical solutions:
[0007] In a first aspect, the present invention provides a small-sample alloy design method, including the following steps:
[0008] Establish an original data set of alloy materials, where the original data set includes various processing technologies, original alloy compositions and multiple performance values;
[0009] Taking the original alloy composition data of the original dataset and multiple processing technologies as inputs, and taking the corresponding multiple performance values as outputs, training a machine learning model to obtain a performance prediction model;
[0010] Reducing the dimension of the original dataset from the original space to a two-dimensional space, obtaining multiple regions formed by different processing technologies, and each region includes multiple original data points formed by different original alloy composition data; wherein, different original data points correspond to different performance values;
[0011] Clustering multiple original data points in each region to obtain multiple data clusters; sampling the multiple data clusters obtained by clustering each region to obtain multiple new data points; mapping each new data point to the original space to obtain a set of new data; wherein, each set of new data includes new alloy compositions and corresponding processing technologies;
[0012] Inputting multiple sets of new data into the performance prediction model to obtain corresponding performance values; comparing the corresponding performance values with the performance values of the corresponding processing technologies in the original dataset, and obtaining the preferred alloy composition according to the comparison results.
[0013] Preferably, the reducing the dimension of the original dataset from the original space to a two-dimensional space specifically includes the following steps:
[0014] Forming multi-dimensional data from multiple alloy compositions and multiple processing technologies in the original dataset and inputting the multi-dimensional data into an encoder;
[0015] The encoder performs multiple dimension mappings on the multi-dimensional data and compresses it into a two-dimensional vector.
[0016] Preferably, before mapping each new data point to the original space, data processing needs to be performed on each new data point, and the data processing process includes:
[0017] Training a binary classifier with multiple original data points to obtain a discriminator;
[0018] Inputting each new data point into the discriminator in sequence to obtain the corresponding log probability;
[0019] Adding the log-likelihood value of the current new data point to the corresponding log probability to obtain a sum value; comparing the sum value of the current data point with the sum value of the previous new data point;
[0020] If the sum value of the current new data point is greater than the sum value of the previous new data point, the current new data point is retained, otherwise it is eliminated.
[0021] Preferably, the training of the machine learning model includes the following steps:
[0022] The machine learning model includes an NN model and an XGBDT model;
[0023] Taking the original alloy composition and various processing techniques as inputs and the corresponding multiple performance values as outputs, the NN model and the XGBDT model are respectively trained to obtain a first performance prediction model and a second performance prediction model. Both the first performance prediction model and the second performance prediction model include multiple sets of different hyperparameters.
[0024] Preferably, the step of inputting multiple sets of new data into the performance prediction model to obtain the corresponding performance values includes the following steps:
[0025] Inputting multiple sets of new data into the first performance prediction model and the second performance prediction model to obtain multiple performance values;
[0026] Averaging the multiple performance values of each set of new data to obtain the corresponding performance value.
[0027] Preferably, the step of comparing the corresponding performance value with the performance value of the corresponding processing technique in the original dataset and obtaining the preferred alloy composition according to the comparison result includes the following steps:
[0028] Obtaining the scores corresponding to multiple sets of new data based on the multiple performance values, and sorting the multiple scores;
[0029] Taking the top three sets of data in the sorting as the alloy compositions to be preferred;
[0030] Comparing the performance values of the alloy compositions to be preferred with the performance values of the corresponding processing techniques in the original dataset. If it is greater, then taking the alloy composition to be preferred as the preferred alloy composition; otherwise, substituting it into the original dataset for an iterative process of dimensionality reduction, clustering, sampling, and performance value acquisition.
[0031] Preferably, the scores corresponding to multiple sets of new data are obtained based on the multiple performance values, and the scores corresponding to multiple sets of new data are obtained through the following formula:
[0032]
[0033] In the formula, Rs represents the fractional value, pre_mean represents the average performance value predicted by multiple sets of hyperparameters, pre_std represents the standard deviation, elements i represents the element composition in the i-th set of new data, elements j represents the element composition in the j-th set of new data.
[0034] Preferably, the alloys in the original alloy composition include Al, Si, Mg, Cu, Ni, Zn, Fe, Mn, and Cr. The multiple processing techniques include gravity casting, gravity casting + heat treatment, gravity casting + hot extrusion, and gravity casting + hot extrusion + heat treatment. The performance value is the tensile strength value.
[0035] Preferably, before reducing the original data set from the original space to a two-dimensional space, one-hot encoding is required to encode the multiple processing techniques. Among them, gravity casting is encoded as 1000, gravity casting + heat treatment is encoded as 0100, gravity casting + hot extrusion is encoded as 0010, and gravity casting + hot extrusion + heat treatment is encoded as 0001.
[0036] In a second aspect, the present invention provides a small-sample alloy design system, including:
[0037] A construction module for establishing an original data set of alloy materials, where the original data set includes multiple processing techniques, original alloy compositions, and multiple performance values;
[0038] A training module for training a machine learning model with the original alloy composition data and multiple processing techniques of the original data set as inputs and the corresponding multiple performance values as outputs to obtain a performance prediction model;
[0039] A dimensionality reduction module for reducing the original data set from the original space to a two-dimensional space to obtain multiple regions formed by different processing techniques. Each region includes multiple original data points formed by different original alloy composition data; among them, different original data points correspond to different performance values;
[0040] A sampling module for clustering multiple original data points in each region to obtain multiple data clusters; sampling the multiple data clusters obtained by clustering each region to obtain multiple new data points; mapping each new data point to the original space to obtain a set of new data; among them, each set of new data includes a new alloy composition and the corresponding processing technique;
[0041] An acquisition module for inputting multiple sets of new data into the performance prediction model to obtain the corresponding performance values; comparing the corresponding performance values with the performance values of the corresponding processing techniques in the original data set, and obtaining the preferred alloy composition according to the comparison results.
[0042] Compared with the prior art, the above at least one technical solution adopted by the present invention can achieve the following beneficial effects:
[0043] The present invention constructs an original dataset including various processing techniques, original alloy compositions, and multiple performance values, integrates alloy data with various process complexities and large sample size differences, trains a machine learning model, and obtains a performance prediction model. The multi-process collaboration greatly reduces the prediction error of small-sample processes by capturing the relationships between different processes and alloy properties, and can achieve accurate prediction of process performance with complex processes and small data volumes. Then, the original dataset is reduced to a two-dimensional latent space to visualize the distribution of alloy data points of different processes. The alloy data of different processes in the latent space are clustered to obtain the cluster distribution of different process data points in the clustering space, and multiple new data points are obtained by sampling multiple data clusters obtained by clustering each region; each new data point is mapped to the original space to obtain a set of new data. Finally, multiple sets of new data are input into the performance prediction model to obtain the corresponding performance values, and the corresponding performance values are compared with the performance values of the corresponding processing techniques in the original dataset, and the preferred alloy compositions with higher performance values are obtained according to the comparison results. The entire design process of the present invention requires relatively low costs, can quickly and accurately predict alloy properties, and can design high-performance alloys that meet expectations. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings without creative efforts based on these drawings.
[0045] Figure 1 It is a design framework diagram of a small-sample alloy design method of the present invention;
[0046] Figure 2 It is a collaborative improvement diagram of multiple processes in an embodiment of the present invention;
[0047] Among them, Figure 2 (a) of it: Performance comparison and analysis of the NN model and the XGBDT model on a single-process dataset and a multi-process collaborative dataset, Figure 2 (b) of it: Comparison diagram of the UTS value predicted by the NN model and the experimental UTS value, Figure 2 (c) of it: Comparison diagram of the UTS value predicted by the XGBDT model and the experimental UTS value;
[0048] Figure 3 It is a clustering sampling diagram of the latent space in an embodiment of the present invention;
[0049] Among them, Figure 3 (a) of it: Distribution diagram of Al-Si alloy data points of different processing techniques in the two-dimensional latent space,Figure 3 (b) of this: Schematic diagram of the GMM clustering results of the gravity casting + hot extrusion process data points, Figure 3 (c) of this: 5-fold cross-validation results of the classifier classification accuracy under the gravity casting + hot extrusion process;
[0050] Figure 4 Training result diagrams of NN and XGBDT in the embodiments of the present invention;
[0051] Among them, Figure 4 (a) of this: Model training effect of the Bayesian optimization NN model under the first set of hyperparameters, Figure 4 (b) of this: Model training effect of the Bayesian optimization NN model under the second set of hyperparameters, Figure 4 (c) of this: Model training effect of the Bayesian optimization NN model under the third set of hyperparameters, Figure 4 (d) of this: Model training effect of the Bayesian optimization NN model under the fourth set of hyperparameters, Figure 4 (e) of this: Model training effect of the Bayesian optimization NN model under the fifth set of hyperparameters, Figure 4 (f) of this: Model training effect of the Bayesian optimization NN model under the sixth set of hyperparameters, Figure 4 (g) of this: Model training effect of the Bayesian optimization XGBDT model under the first set of hyperparameters, Figure 4 (h) of this: Model training effect of the Bayesian optimization XGBDT model under the second set of hyperparameters, Figure 4 (i) of this: Model training effect of the Bayesian optimization XGBDT model under the third set of hyperparameters, Figure 4 (j) of this: Model training effect of the Bayesian optimization XGBDT model under the fourth set of hyperparameters, Figure 4 (k) of this: Model training effect of the Bayesian optimization XGBDT model under the fifth set of hyperparameters, Figure 4 (l) of this: Model training effect of the Bayesian optimization XGBDT model under the sixth set of hyperparameters;
[0052] Figure 5 Experimental verification result diagrams in the present invention;
[0053] Among them, Figure 5 (a) of this: Stress-strain curve of the high-performance Al-Si alloy under the gravity casting + hot extrusion process, Figure 5 (b) of this: Comparison of the prediction results of the present invention with the experimental results in the existing literature. Specific implementation manners
[0054] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.
[0055] Based on the problems existing in the prior art, the present invention proposes a small-sample alloy design method, taking Al-Si as an example. Refer to Figure 1 , including the following steps:
[0056] S1: Establish an original data set of alloy materials.
[0057] The original data set mainly includes three parts: original alloy composition (mass percentage of each component), performance value, and multiple processing technologies.
[0058] In this embodiment, the alloy components mainly consider Al, Si, Mg, Cu, Ni, Zn, Fe, Mn, and Cr, the performance is mainly the ultimate tensile strength (UTS), and the processing technologies are mainly gravity casting (GC), gravity casting + heat treatment (GC + T6), gravity casting + hot extrusion (GC + HE), and gravity casting + hot extrusion + heat treatment (GC + HE + T6). Due to the differences in complexity and applicability of these processes, there are also significant differences in the sample size. Among them, there are 38 groups of data for the GC process, 70 groups for the GC + T6 process, 13 groups for the GC + HE process, and 19 groups for the GC + HE + T6 process. Obviously, the sample size of the GC + HE process is the smallest. Therefore, it is difficult for machine learning to accurately predict the performance of a single GC + HE process alloy. In order to uniformly represent these different processing technologies, the present invention uses one-hot encoding to replace these four processes. Among them, the one-hot encoding of the GC process is 1000, the GC + T6 process is 0100, the GC + HE process is 0010, and the GC + HE + T6 process is 0001. The alloy performance values obtained by combining different alloy components and different processing technologies are different.
[0059] S2: Reduce the dimension of the original data set from the original space to a two-dimensional space.
[0060] The present invention uses C-WAE to perform dimensionality reduction processing on the high-dimensional data set. Specifically, the alloy composition (9-dimensional) is used as the high-dimensional input, and at the same time, the processing technology information after one-hot encoding (4-dimensional) is used as the conditional input and passed to the encoder. The encoder is composed of a multi-layer fully connected network, which maps the 13-dimensional input to 96, 72, and 40 dimensions in sequence, and finally compresses it into a 2-dimensional latent vector. During this process, the processing technology conditions are combined to learn a more targeted low-dimensional latent representation, thereby retaining the key features in the input data and presenting a more concise structure.
[0061] Figure 3 (a) of this invention uses C-WAE to reduce the high-dimensional alloy composition and processing technology to a two-dimensional latent space. Obviously, the 4 processing technologies in the dataset of this invention form 4 isolated regions with different shapes in the two-dimensional latent space. Among them, the color depth of the data points in different regions corresponds to the UTS value of the right color tone.
[0062] S3: Cluster the specified region in the two-dimensional latent space to obtain the clustering result.
[0063] Use GMM to cluster the latent space. Specifically, in the reduced latent space, the regions corresponding to different processing technologies are clustered by GMM. Assume that the data points come from multiple Gaussian distributions, and combine hyperparameters such as the appropriate number of Gaussian components, initialization method, and random seed (for example, the number of Gaussian components 'n_components' is 2 to 5, the initialization method 'init_params = 'kmeans'' is used for k-means initialization, and the random seeds 'random_state' are 0, 42, etc. to ensure the repeatability of the results). Through the iterative expectation-maximization (EM) algorithm, continuously adjust the mean, covariance, and mixing weights of each cluster, and finally complete the division of the data points. Among them, the n_components under each condition is preset according to different processing technologies to ensure that each clustering result represents different processing technology regions. For example, the n_components of the GC+T6 process in this invention is set to 6. Finally, the data points are divided into different clusters, and each cluster contains similar samples.
[0064] In this embodiment, use the GMM model to Figure 3 cluster the GC+HE process small island in (a) of Figure 3 The result is shown in (b) of
[0065] Among them, the color gradient represents the probability density size of the data points when performing Gaussian distribution. The dark orange area indicates a higher log-likelihood value (high probability density, log likelihood > -5), which means that the data points are more concentrated in these areas. While the light green or blue-green areas indicate a lower log-likelihood value (low probability density, log likelihood < -5), indicating that there are fewer or almost no data points distributed in these areas.
[0066] Sample from these clusters using the MCMC method to generate new data points, and decide whether to accept them through a discriminator (i.e., a binary classifier trained in the latent space). Specifically, first perform an initial sampling on the Gaussian distribution corresponding to each cluster, and generate candidate points according to the normal distribution (e.g., a multivariate normal with a variance of σ = 0.3) in each iteration. Subsequently, calculate the sum of the log-likelihood value of the candidate point in the GMM and the log probability output by the classifier. If it is better than the current point (or randomly meets the condition), then accept the update; otherwise, reject it and keep the original point unchanged. The discriminator here is a fully connected network of 2→12→1, the hidden layer contains Dropout(0.5) and uses Sigmoid as the output. During training, use the Adam optimizer (lr = 5*10 ^-4 , batch_size = 8) and use 5-fold cross-validation to improve the robustness of the model. In this process, if the classifier probability of the candidate point (indicating the matching degree to the target performance or process requirements) is low, it will be directly discarded; if it meets the requirements, it will be accepted and used as a valid data point. Through the combination of "GMM distribution constraint + classifier decision", not only can it ensure that the sampling points are close to the original latent distribution, but also the finally retained points can meet specific processing technology or performance requirements, so as to realize the controllable generation of alloy formulations in the two-dimensional latent space.
[0067] During the MCMC sampling process, new data points will be preferentially generated from the region with a higher log-likelihood value (orange region), and then use the discriminator to decide whether to accept, as shown in Figure 3 (c) of. Through 5-fold cross-validation, the discriminator's judgment of the MCMC sampling points is mainly based on the accuracy of the training and test sets. The average training accuracy is 0.78, and the test accuracy is 0.77, indicating that overall the sampling points conform to the target distribution and can be accepted by the discriminator. Since the GC+HE process data is scarce and relatively complex, in order to demonstrate the performance prediction of the present invention in complex process data and scarce data processes, the present invention conducts a design with this process as the object. Of course, the process here can also be the other three processes, and the design steps are the same. Table 1 shows some of the component data finally sampled in the clustering space of the GC+HE process.
[0068] Table 1 is some of the component data of the finally sampled GC+HE
[0069] Al Si Mg Cu Ni Zn Fe Mn Cr 0.981927 0.012504 0.002755 0.001918 7.59E-05 3.19E-05 0.000602 8.87E-05 9.77E-05 0.977577 0.017471 0.002737 0.001407 5.63E-05 3.26E-05 0.000557 7.96E-05 8.16E-05 0.890302 0.104043 0.004445 0.000623 2.44E-05 1.78E-05 0.000434 3.45E-05 4.63E-05 0.877823 0.116183 0.004919 0.000546 2.78E-05 3.83E-05 0.000398 2.52E-05 4.03E-05 0.880164 0.114157 0.004564 0.000571 2.86E-05 3.99E-05 0.000407 2.67E-05 4.22E-05 0.882848 0.110544 0.005524 0.000546 2.35E-05 4.14E-05 0.000404 2.83E-05 4.06E-05 0.882935 0.110388 0.005593 0.000546 2.33E-05 4.16E-05 0.000405 2.84E-05 4.06E-05 0.872074 0.120225 0.006725 0.000479 2.42E-05 3.54E-05 0.000379 2.25E-05 3.57E-05 0.866849 0.125634 0.006573 0.00046 2.73E-05 3.20E-05 0.00037 1.94E-05 3.42E-05 0.901075 0.093656 0.003885 0.000734 2.66E-05 5.22E-05 0.000477 4.10E-05 5.34E-05 0.857123 0.132098 0.009919 0.000407 2.45E-05 3.35E-05 0.000349 1.59E-05 3.12E-05
[0070] S5: Decode the new data points to obtain a new data set.
[0071] The decoder maps the sampled valid data points back to the original high-dimensional alloy composition and processing technology, thus completing the generation of the alloy formulation.
[0072] The corresponding decoder then concatenates the 2D latent vector with the 4D conditions again and maps them back to the 9D alloy composition through a multi-layer fully connected network (with dimensions of 40, 72, and 96 in sequence before the output layer), and adds a Wasserstein regularization based on the Maximum Mean Discrepancy (MMD) to the loss function to align the latent distribution with the prior distribution. During the training process, 350 training epochs are set, the batch size is 50, the learning rate (lr) is 5*10^ -4 , the weight decay is 0, the prior noise intensity (sigma) is 2.0, and the MMD regularization coefficient (MMD_lambda) is set to 8*10^ -4 to ensure the structure and generatability of the latent space.
[0073] S6: Train the NN model and the XGBDT model using the original dataset.
[0074] After completing the composition sampling, the original dataset is used to train the NN model and the XGBDT model, and the trained models are used to predict the performance of the sampled compositions. Specifically, when training the NN model, the alloy composition and processing technology information (such as one-hot encoded process conditions) are used as inputs, and regression prediction is performed through a multi-layer fully connected network, and finally the tensile strength of the alloy material is output. During training, MAPE (Mean Absolute Percentage Error) is used as the loss function, and the Adam optimizer is used to update the network weights; at the same time, hyperparameters including the learning rate Learning_rate, batch size Batch_size, the number of hidden layer units Module__n_hiodden, and the depth of the hidden layer Module_w are iteratively searched through Bayesian optimization. Each time, a probability model (such as a Gaussian process) is constructed and updated based on the completed training results, and new hyperparameter combinations are automatically proposed for verification in subsequent training, and finally converge to the best configuration that can minimize the test error (Target), so as to effectively model and learn the "composition-process-strength" relationship.
[0075] Similarly, the XGBDT model is trained. The alloy composition and processing technology information (such as one-hot encoded process conditions) are used as inputs, and regression prediction is performed through XGBRegressor. Finally, the tensile strength of the alloy material is output. During training, MAPE (Mean Absolute Percentage Error) is also used as the evaluation index, and hyperparameters including learning rate Learning_rate, maximum depth Max_depth, number of weak learners N_estimators, subsampling rate Subsample, feature sampling rate Colsample_bytree, regularization coefficients Reg_alpha and Reg_lambda, etc. are iteratively searched through Bayesian optimization. Each time, a probability model (such as Gaussian process) is constructed and updated based on the completed training results, and new hyperparameter combinations are automatically proposed for verification in subsequent training, finally converging to the best configuration that can minimize the test error (Target), thereby effectively modeling and learning the "composition - process - strength" relationship.
[0076] Figure 4 Figure for the training and learning of the dataset by the NN model and the XGBDT model. Obviously, after optimizing the six groups of hyperparameters of the NN model and the XGBDT model by the Bayesian method, the models have relatively high determination coefficient R 2 values on both the training set and the test set, indicating good prediction performance. The six groups of hyperparameters corresponding to the NN model and the XGBDT model are shown in Table 2 and Table 3 respectively. Among them, because the XGBDT model has more hyperparameters, the present invention lists the first few hyperparameters.
[0077] Table 2 Six groups of hyperparameters of the NN model
[0078]
[0079] Table 3 Six groups of hyperparameters of the XGBDT model
[0080]
[0081] S7: Input the new dataset into the two models.
[0082] Next, the present invention uses 6 sets of hyperparameters of the NN model and XGBDT respectively to predict the performance of the sampled components, and the average value of the predictions of these 12 sets of hyperparameters will be used as the final prediction value. By combining two different modeling methods, NN and XGBDT, the advantages of each can be exploited in the prediction: NN is good at learning high-dimensional non-linear mappings, while XGBDT is more robust to sparse features and heterogeneous data. The combination of the two can reduce the overfitting or bias that may occur in a single model, thereby improving the overall prediction accuracy and robustness. Therefore, taking the average (or weighted average) of the prediction results of the two is an effective ensemble learning strategy, which can achieve better comprehensive prediction performance in the modeling of the "composition - process - performance" relationship.
[0083] Next, the prediction data is ranked using a ranking criterion, mainly using experimental experience and numerical ranking methods, as follows:
[0084]
[0085] Among them, pre_mean represents the average value of the predictions of 12 sets of hyperparameters, pre_std represents the standard deviation, Rs represents the ranking score, and elements i represents the elemental composition in the i-th alloy, and elements j represents the elemental composition in the j-th alloy. Additionally, for the convenience of experimental verification, the elemental difference between the selected high-performance components is at least greater than 0.8%. According to this ranking criterion, the present invention finally conducts experimental verification on the top three groups of data with better performance and iterates them into the original dataset.
[0086] Figure 5 For the tensile experiment and performance comparison of the finally predicted high-performance alloys. After the ranking criterion, 3 alloy compositions with high performance are finally selected, as shown in Table 3. Through experiments, the actual performances of the first and second groups of components are 208.5 MPa and 220.5 MPa respectively, which have exceeded the existing performances in the dataset. Since the data of the GC + HE process is small and the performance is low. Therefore, only one iteration is required to find alloy components with superior performance. This also indicates that the present invention can achieve accurate prediction of complex and scarce data processes. Not only the GC + HE process, but also other processes can achieve accurate performance prediction. Therefore, only one iteration is required to find alloy components with superior performance.
[0087] Table 3 Composition - Performance Prediction of High-Performance Al-Si Alloys under the GC + HE Process
[0088]
[0089] Figure 2It is a collaborative improvement diagram for multiple processes. The present invention integrates the composition-process-performance data of Al-Si alloys containing the same elements, which includes multiple processes. Two models, NN and XGBDT, are used to compare the MAE error and error uncertainty of individual processes and the collaboration of multiple processes respectively. Figure 2 Figure (a) shows the performance comparison and analysis of the NN model and the XGBDT model on the individual process dataset and the dataset of the collaboration of multiple processes. It can be seen that the prediction error of the model after the collaboration of multiple processes is significantly reduced. For example, in the GC process, the prediction error and error uncertainty of the NN for individual processes are 15.8 and 10.5, and those of the XGBDT are 17.7 and 13.2. After the overall collaboration, the prediction errors and error uncertainties of the two models are, for NN: 8.1, 8.2, and for XGBDT: 9.3, 6.9. Figure 2 Figure (b) shows the comparison diagram of the predicted UTS value and the experimental UTS value. It can be seen that the training and learning capabilities of the model are significantly enhanced after the collaboration of multiple processes. Among them, the determination coefficients R of the NN model and the XGBDT model for predicting the dataset after the collaboration of multiple processes 2 both exceed 0.9, which are 0.903 and 0.912 respectively, far greater than the determination coefficient R of the model for predicting individual processes 2 value. The R 2 value quantifies the prediction accuracy of the NN model and the XGBDT model on the individual process and the dataset of the collaboration of multiple processes.
[0090] Referring to Figure 1 , first, an Al-Si alloy dataset is constructed based on existing experimental data and literature data. Then, the conditional Wasserstein autoencoder (C-WAE) is used to reduce the dimension of the high-dimensional dataset, and the data points are distinguished according to different processes. Next, the Gaussian mixture model (GMM) and the Markov chain Monte Carlo (MCMC) sampling method are used to cluster the data points in the two-dimensional latent space for different processes and sample the components of the specified process. In addition, the present invention trains and learns the dataset through a neural network (NN) and an XBOOST gradient boosting decision tree (XGBDT), and uses the trained model to predict the performance of the sampling points. Finally, based on the ranking criterion, the components with the top rankings are experimentally verified, and the experimental data are iteratively re-entered into the dataset to finally find alloy components with high performance. The experimental results in the present invention also prove the rationality of the design scheme of the present invention and the accuracy of model prediction, which has guiding significance for actual industrial production.
[0091] Based on the same concept, the present invention also provides a small-sample alloy design system, including a construction module, a training module, a dimension reduction module, a sampling module, and an acquisition module.
[0092] The construction module is used to establish the original data set of alloy materials, and the original data set includes various processing technologies, original alloy compositions, and multiple performance values.
[0093] The training module is used to take the original alloy composition data and various processing technologies of the original data set as inputs, and the corresponding multiple performance values as outputs, to train a machine learning model to obtain a performance prediction model.
[0094] The dimensionality reduction module is used to reduce the dimensionality of the original data set from the original space to a two-dimensional space, obtaining multiple regions formed by different processing technologies, and each region includes multiple original data points formed by different original alloy composition data; wherein, different original data points correspond to different performance values.
[0095] The sampling module is used to cluster multiple original data points in each region to obtain multiple data clusters; sample the multiple data clusters obtained by clustering each region to obtain multiple new data points; map each new data point to the original space to obtain a set of new data; wherein, each set of new data includes new alloy compositions and corresponding processing technologies.
[0096] The acquisition module is used to input multiple sets of new data into the performance prediction model to obtain corresponding performance values; compare the corresponding performance values with the performance values of the corresponding processing technologies in the original data set, and obtain the preferred alloy compositions according to the comparison results.
[0097] Although the preferred embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications once they learn the basic creative concepts. Therefore, the appended claims are intended to be construed to include the preferred embodiments as well as all changes and modifications falling within the scope of the present invention.
[0098] Obviously, those skilled in the art can make various changes and deformations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and deformations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention also intends to include these changes and deformations.
Claims
1. A small-sample alloy design method, characterized in that, Including the following steps: Establish an original data set of alloy materials, where the original data set includes various processing techniques, original alloy compositions, and multiple performance values; Using the original alloy composition data and various processing techniques in the original data set as inputs, and the corresponding multiple performance values as outputs, train a machine learning model to obtain a performance prediction model; Reduce the dimensionality of the original data set from the original space to a two-dimensional space, obtaining multiple regions formed by different processing techniques, where each region includes multiple original data points formed by different original alloy composition data; among them, different original data points correspond to different performance values; Cluster the multiple original data points in each region to obtain multiple data clusters; sample the multiple data clusters obtained by clustering each region to obtain multiple new data points; map each new data point to the original space to obtain a set of new data; where each set of new data includes a new alloy composition and the corresponding processing technique; Input multiple sets of new data into the performance prediction model to obtain the corresponding performance values; compare the corresponding performance values with the performance values of the corresponding processing techniques in the original data set, and obtain the preferred alloy composition according to the comparison results.
2. The small-sample alloy design method according to claim 1, wherein The reduction of the dimensionality of the original data set from the original space to a two-dimensional space specifically includes the following steps: Input the multiple alloy compositions and various processing techniques in the original data set as multi-dimensional data into the encoder; The encoder performs multiple dimensionality mappings on the multi-dimensional data and compresses it into a two-dimensional vector.
3. A small-sample alloy design method as claimed in claim 1, wherein Before mapping each new data point to the original space, each new data point needs to be processed, and the data processing process includes: Train a binary classifier through multiple original data points to obtain a discriminator; Input each new data point into the discriminator in turn to obtain the corresponding logarithmic probability; Add the logarithmic likelihood value of the current new data point to the corresponding logarithmic probability to obtain a sum value; compare the sum value of the current data point with the sum value of the previous new data point; If the sum value of the current new data point is greater than the sum value of the previous new data point, retain the current new data point, otherwise eliminate it.
4. A small-sample alloy design method as claimed in claim 1, characterized in that, The training of the machine learning model includes the following steps: The machine learning model includes an NN model and an XGBDT model; Using the original alloy composition and various processing techniques as inputs, and the corresponding multiple performance values as outputs, train the NN model and the XGBDT model respectively to obtain a first performance prediction model and a second performance prediction model, and both the first performance prediction model and the second performance prediction model include multiple groups of different hyperparameters.
5. The small-sample alloy design method according to claim 4, characterized in that The step of inputting multiple sets of new data into the performance prediction model to obtain the corresponding performance values includes the following steps: Input multiple sets of new data into the first performance prediction model and the second performance prediction model to obtain multiple performance values; Average the multiple performance values of each set of new data to obtain the corresponding performance value.
6. The small-sample alloy design method according to claim 1, characterized in that The step of comparing the corresponding performance values with the performance values of the corresponding processing techniques in the original data set and obtaining the preferred alloy composition according to the comparison results includes the following steps: Obtain the scores corresponding to multiple sets of new data based on multiple performance values, and sort the multiple scores; Take the top three sets of data in the sorting as the alloy compositions to be preferred; Compare the performance value of the alloy composition to be optimized with the performance value of the corresponding processing technology in the original dataset. If it is greater, the alloy composition to be optimized is regarded as the optimized alloy composition; otherwise, it is substituted into the original dataset for the iterative process of dimensionality reduction, clustering, sampling, and performance value acquisition.
7. The small-sample alloy design method according to claim 6, wherein Obtain the scores corresponding to multiple sets of new data based on multiple performance values. The scores corresponding to multiple sets of new data are obtained through the following formula: where Rs represents a fractional value, pre_mean represents the average performance value predicted by multiple sets of hyperparameters, pre_std represents the standard deviation value, elements i represents the element composition in the i-th group of new data, elements j represents the element composition in the j-th group of new data.
8. A small-sample alloy design method according to claim 1, characterized in that, The alloys in the original alloy composition include Al, Si, Mg, Cu, Ni, Zn, Fe, Mn, and Cr. The multiple processing technologies include gravity casting, gravity casting + heat treatment, gravity casting + hot extrusion, and gravity casting + hot extrusion + heat treatment. The performance value is the tensile strength value.
9. A small-sample alloy design method according to claim 8, characterized in that Before reducing the original dataset from the original space to a two-dimensional space, it is necessary to encode the multiple processing technologies through one-hot encoding. Among them, gravity casting is encoded as 1000, gravity casting + heat treatment is encoded as 0100, gravity casting + hot extrusion is encoded as 0010, and gravity casting + hot extrusion + heat treatment is encoded as 0001.
10. A small-sample alloy design system, characterized in that, It includes: A construction module for establishing the original dataset of alloy materials. The original dataset includes multiple processing technologies, original alloy compositions, and multiple performance values; A training module for training a machine learning model with the original alloy composition data and multiple processing technologies in the original dataset as inputs and the corresponding multiple performance values as outputs to obtain a performance prediction model; A dimensionality reduction module for reducing the original dataset from the original space to a two-dimensional space to obtain multiple regions formed by different processing technologies. Each region includes multiple original data points formed by different original alloy composition data; among them, different original data points correspond to different performance values; A sampling module for clustering multiple original data points in each region to obtain multiple data clusters; sampling the multiple data clusters obtained by clustering each region to obtain multiple new data points; mapping each new data point to the original space to obtain a set of new data; among them, each set of new data includes a new alloy composition and the corresponding processing technology; An acquisition module for inputting multiple sets of new data into the performance prediction model to obtain the corresponding performance values; comparing the corresponding performance values with the performance values of the corresponding processing technologies in the original dataset, and obtaining the optimized alloy composition according to the comparison results.