Shape memory alloy component prediction method and related device
By leveraging the synergistic operation of the WCGAN-GP network and the LSTM model, reverse design of Ti-Ni-based shape memory alloys was achieved, solving the problems of long R&D cycles and high costs in traditional methods, and improving R&D efficiency and accuracy.
Patent Information
- Application Number
- CN202511365516.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-23
- Publication Date
- 2026-01-13
AI Technical Summary
In the existing technology, the research and development cycle of Ti-Ni based shape memory alloys is long and the cost is high. Furthermore, traditional methods are difficult to reverse design, and computational materials science methods have the problems of high computational cost and inability to reverse derive design schemes.
By employing the collaborative operation of the WCGAN-GP network and the LSTM model, efficient prediction from target performance to composition is achieved through inverse design. The WCGAN-GP network generates phase field parameters, the LSTM model performs forward prediction, and the optimal composition is selected by combining similarity.
This approach enables reverse design from target performance to composition, significantly improving R&D efficiency and accuracy, reducing the time and resource consumption of high-throughput screening, and solving the computational bottleneck and reverse design challenges of traditional methods.
Smart Images
Figure CN121331298A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of material science, and particularly relates to a shape memory alloy component prediction method and related device. BACKGROUND
[0002] Ti-Ni-based shape memory alloys (SMAs) play an indispensable role in key fields such as aerospace and biomedical due to their unique shape memory effect and super-elasticity, especially in harsh scenarios with large temperature changes, where it is crucial for them to maintain stable super-elasticity in a wide temperature range. However, their performance is highly sensitive to the combination of composition and process, and the traditional "trial-and-error" research and development mode faces a wide design space, with long research and development cycles and high costs.
[0003] To address this challenge, computational materials science methods have emerged. Phase field simulation (based on Ginzburg-Landau theory) can accurately simulate the dynamic evolution of microstructure, but it has two major core defects: first, the extremely high computational cost hinders high-throughput screening; second, as a forward model, it cannot deduce the design scheme from the target performance. On the other hand, machine learning techniques represented by deep learning are good at mining non-linear relationships in data, but their application in materials science is still limited by the sparsity of high-quality data, and their "black box" nature may lead to a lack of physical consistency in prediction results, limiting the reliability and generalization ability of the model.
[0004] To solve the above problems, current research has begun to use machine learning to accelerate phase field simulation, but these works are mostly limited to "forward prediction" acceleration and cannot achieve the "reverse design" capability needed for designing new materials. SUMMARY
[0005] To solve the problems in the prior art, the purpose of the present application is to provide a shape memory alloy component prediction method and related device, which can predict the composition of shape memory alloys according to the performance requirements of shape memory alloys, and achieve the "reverse design" capability needed for designing new materials.
[0006] To achieve the above purpose, the following technical solutions are used in the present application: A shape memory alloy component prediction method, comprising the following processes: The target stress-strain curve of the shape memory alloy preset is processed through the pre-trained WCGAN-GP network to obtain a plurality of groups of phase field parameters corresponding to the target stress-strain curve; The plurality of groups of phase field parameters are processed through the trained LSTM model respectively to obtain a predicted stress-strain curve corresponding to each group of phase field parameters; calculate the similarity between the target stress-strain curve and each predicted stress-strain curve, screen the predicted stress-strain curve with the similarity meeting the requirement, take the predicted stress-strain curve with the similarity meeting the requirement as the optimal predicted stress-strain curve, and take the phase field parameter corresponding to the optimal predicted stress-strain curve as the final composition of the shape memory alloy.
[0007] Preferably, in the WCGAN-GP network, the critic input is a random parameter and a target stress-strain curve, the output is a real score, the loss function is a Wasserstein-1 distance plus a gradient penalty, and the Lipschitz continuity is forced.
[0008] Preferably, when training and evaluating the WCGAN-GP network, the critic is updated every 5 times, the generator is updated once, and the optimizer is Adam.
[0009] Preferably, when training the LSTM model, the mean square error loss function is used, the optimizer is Adam, the initial learning rate is 1e -4 , the learning rate decay and the early stopping mechanism are combined to prevent overfitting, the "teacher forcing" strategy is used in the training process to accelerate convergence, and the autoregressive generation is used in the inference stage.
[0010] Preferably, the process of screening the predicted stress-strain curve with the similarity meeting the requirement comprises: When the similarity between the target stress-strain curve and each predicted stress-strain curve is within a preset range of similarity, it is indicated that the similarity between the predicted stress-strain curve and the target stress-strain curve meets the requirement, otherwise, it is indicated that the similarity between the predicted stress-strain curve and the target stress-strain curve does not meet the requirement.
[0011] Preferably, when the similarity between the predicted stress-strain curve and the target stress-strain curve does not meet the requirement, the target stress-strain curve is adjusted to obtain a new target stress-strain curve, and the shape memory alloy composition prediction method is performed again using the new target stress-strain curve until the similarity between the predicted stress-strain curve and the target stress-strain curve meets the requirement.
[0012] Preferably, the phase field parameter comprises a temperature for generating a stress-strain curve, and atomic size parameters of each metal element in the shape memory alloy and doping concentrations of each metal element. The similarity between the target stress-strain curve and the predicted stress-strain curve is a mean square error of the horizontal coordinates of the target stress-strain curve and the predicted stress-strain curve.
[0013] The application also provides a shape memory alloy composition prediction system for implementing the shape memory alloy composition prediction method. Reverse design module: used to process the pre-set target stress-strain curve of the shape memory alloy through a pre-trained WCGAN-GP network to obtain multiple sets of phase field parameters corresponding to the target stress-strain curve; Forward prediction module: used to process the multiple sets of phase field parameters through the trained LSTM model to obtain the predicted stress-strain curves corresponding to each set of phase field parameters; The optimization module is used to calculate the similarity between the target stress-strain curve and each predicted stress-strain curve, filter the predicted stress-strain curves that meet the similarity requirements with the target stress-strain curve, take the predicted stress-strain curves that meet the similarity requirements as the optimal predicted stress-strain curves, and take the phase field parameters corresponding to the optimal predicted stress-strain curves as the final composition of the shape memory alloy.
[0014] The present invention also provides an electronic device, comprising: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the shape memory alloy composition prediction method of the present invention as described above.
[0015] The present invention also provides a storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the shape memory alloy composition prediction method of the present invention as described above.
[0016] The present invention has the following beneficial effects: This invention's method for predicting the composition of shape memory alloys effectively solves the dual dilemma of phase-field simulation and machine learning applications in existing technologies through the collaborative operation of the WCGAN-GP network and the LSTM model, achieving efficient reverse design of shape memory alloys from target performance to composition. Addressing the problems of long development cycles and high costs associated with traditional trial-and-error R&D, this invention first processes the target stress-strain curve using a pre-trained WCGAN-GP network, directly generating multiple sets of matching phase-field parameters. This overcomes the limitation that phase-field simulation, as a forward model, cannot be reverse-derived, allowing for rapid identification of potential composition design directions without relying on extensive physical experiments. Simultaneously, a trained LSTM model is used to forward predict the corresponding stress-strain curves from multiple sets of phase-field parameters, replacing computationally expensive phase-field simulation. This significantly reduces the time and resource consumption for high-throughput screening, and the LSTM model's ability to mine nonlinear relationships in the data compensates for prediction biases caused by data sparsity in pure machine learning. By calculating the similarity between the target curve and the predicted curve and selecting the optimal result, the matching accuracy between the final phase-field parameters and the target performance is further ensured, avoiding the physical inconsistency problems that may arise from the black box nature of machine learning. The entire process forms a complete closed loop of target performance input, phase field parameter generation, performance prediction and verification, and optimal parameter selection, which truly realizes the reverse design requirement of performance-based composition finding and significantly improves the efficiency, accuracy and reliability of shape memory alloy composition research and development. Attached Figure Description
[0017] Figure 1 This is a diagram illustrating the overall research framework for material design that integrates phase-field simulation and machine learning in this embodiment of the invention. Figure 2 This is a schematic diagram of the LSTM model architecture in an embodiment of the present invention; Figure 3 The prediction curves generated by the LSTM model with shape memory effect behavior in this embodiment of the invention are compared with the theoretical simulations of Ginzburg-Landau. Figure 4 The prediction curves with hyperelastic behavior generated by the LSTM model in this embodiment of the invention are compared with the theoretical simulations of Ginzburg-Landau. Figure 5 The prediction curves with linear elastic behavior generated by the LSTM model in this embodiment of the invention are compared with the theoretical simulations of Ginzburg-Landau. Figure 6 Performance evaluation of the LSTM model in this embodiment of the invention: scatter plot of predicted strain and actual strain on the test set; Figure 7 Performance evaluation of the LSTM model in this embodiment of the invention: comparison of the probability distribution functions of real and generated data points; Figure 8 This invention provides a comparison of the computational efficiency of the LSTM model and the Ginzburg-Landau theoretical simulation in this embodiment. Figure 9 This is a schematic diagram of the WCGAN-GP network architecture in an embodiment of the present invention; Figure 10 For the WCGAN-GP network performance evaluation in this embodiment of the invention: a comparison of the distribution of real parameters (solid dots) and generated parameters (crosses) in the t-SNE dimensionality reduction space; Figure 11 The distribution of three types of data—real, reconstructed, and innovative—in the t-SNE space for WCGAN-GP network performance evaluation in this embodiment of the invention; Figure 12 For the standard hyperelastic stress-strain curves in this embodiment of the invention, 20 sets of phase field parameter screening and comparison charts are generated; Figure 13 Twenty sets of phase field parameter screening and comparison charts were generated for the non-standard hyperelastic stress-strain curves in this embodiment of the invention. Figure 14 The phase field parameters obtained from the standard hyperelastic stress-strain curve prediction in this embodiment of the invention are input into the phase field model to predict the corresponding stress-strain curve. Figure 15 The phase field parameters obtained from the prediction of the non-standard hyperelastic stress-strain curve in the embodiments of the present invention are input into the phase field model to predict the corresponding stress-strain curve. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0019] The core idea of this invention is the deep integration of physical models and data-driven methods: This invention utilizes physically complete phase-field simulations to "produce" massive amounts of high-quality training data for machine learning models; conversely, it uses machine learning to construct efficient surrogate models. This not only overcomes the computational bottleneck of phase-field simulations but also transforms the inverse design problem, which is difficult to solve using traditional methods, into a data-driven problem that can be solved efficiently. Ultimately, this invention aims to establish a new closed-loop R&D paradigm from "rapid prediction" to "on-demand design," with the goal of significantly accelerating the discovery process of shape memory alloys (such as wide-temperature-range superelastic Ti-Ni-based alloys).
[0020] The core of this invention's technical solution is a joint generation and selection framework that collaboratively utilizes a Conditional Generative Adversarial Network (WCGAN-GP) and a Long Short-Term Memory (LSTM) network. The entire process is as follows: First, an LSTM network is trained as an efficient forward prediction proxy model, capable of quickly and accurately predicting the corresponding stress-strain curve based on the input phase field parameters. Simultaneously, a WCGAN-GP network is trained as the core of the reverse design, tasked with receiving a target performance curve and generating a set of matching phase field parameters. Through the collaborative work of these two models, the optimal combination of phase field parameters can be efficiently explored and locked based on the given target performance, thus completing the reverse design closed loop from "finding components from performance."
[0021] Specifically, the shape memory alloy composition prediction method of this invention is a reverse design framework that integrates phase-field simulation and machine learning models, and its core logic is as follows: 1. Define target performance: Define a desired macroscopic performance of the shape memory alloy, such as the hyperelastic stress-strain curve (denoted as the target stress-strain curve) that exhibits good performance at low temperatures in this invention.
[0022] 2. Construction and training of the forward prediction model: Long short-term memory network (i.e., LSTM network or LSTM network model) is used to achieve high-precision prediction from 10-dimensional phase field parameters to 81-point stress-strain curve.
[0023] 3. Construction and training of the reverse design model: A Wasserstein conditional generative adversarial network with gradient penalty (i.e., WCGAN-GP network) is used to realize the reverse design from stress-strain curves to 10-dimensional phase field parameters.
[0024] 4. Reverse Design and Parameter Selection: The target stress-strain curve is used as a condition label and input into the pre-trained WCGAN-GP network. Multiple sets of phase field parameters are output to the LSTM reverse design network (i.e., the LSTM network mentioned above). Each set of phase field parameters will generate a corresponding stress-strain curve (denoted as the predicted stress-strain curve). In this way, multiple sets of phase field parameters can generate multiple predicted stress-strain curves. The generated multiple predicted stress-strain curves are compared with the target stress-strain curve, and the set of phase field parameters corresponding to the predicted stress-strain curve with the highest similarity to the target stress-strain curve is selected. This set of phase field parameters is used as the final composition of the predicted shape memory alloy of this invention.
[0025] The predicted shape memory alloy composition can also be verified: input the set of phase field parameters with the highest similarity into the phase field model, which can generate the stress-strain curve corresponding to the final composition. The predicted shape memory alloy composition can be verified based on the similarity between the stress-strain curve and the target stress-strain curve.
[0026] The above architecture includes two core parameters: 1. Phase-field parameters, consisting of a ten-dimensional vector of h1, s1, c1, h2, s2, c2, h3, s3, c3, and T, represent the alloy composition. (In Ginzburg-Landau theory, h1, s1, h2, s2, h3, and s3 represent the atomic sizes of different heterometallic materials, c1, c2, and c3 represent the doping concentration, and T represents the temperature at which the stress-strain curve is generated. Among them, h1, s1, and c1 correspond to one element, h2, s2, and c2 correspond to one element, and h3, s3, and c3 correspond to one element.)
[0027] 2. Stress-strain curves, which are 81-dimensional vectors, represent the macroscopic mechanical properties of the material. The overall research framework is as follows: Figure 1 As shown.
[0028] Construction and performance evaluation of LSTM models 1) Design an LSTM model that takes a 10-dimensional phase field parameter as input and outputs the predicted stress-strain curve. The LSTM model is trained using the mean squared error (MSE) loss function, with Adam as the optimizer and an initial learning rate of 1e. -4 This is combined with a learning rate decay and early stopping mechanism to prevent overfitting. During training, a "teacher-forced" strategy is used to accelerate convergence, while autoregressive generation is employed during the inference phase. Network construction follows... Figure 2 As shown 2) Input phase field parameters into the trained LSTM model, and the LSTM model outputs stress-strain curves, such as... Figure 3 The shape memory effect (De-twinning) with obvious residual strain is shown. The predicted stress-strain curve (orange) is similar to the actual stress-strain curve (green) simulated by Ginzburg-Landau theory at key feature points (such as maximum strain value and residual strain value).
[0029] 3) Input phase field parameters into the trained LSTM model, and the LSTM model outputs stress-strain curves, such as... Figure 4 The superelastic model shown has recoverable strain. The predicted stress-strain curve (orange) is similar to the actual stress-strain curve (green) simulated by Ginzburg-Landau theory at key feature points (such as maximum strain value and residual strain value).
[0030] 4) Input phase field parameters into the trained LSTM model and output stress-strain curves, such as... Figure 5The approximate linear elastic behavior is observed. The predicted stress-strain curve (orange) is similar to the actual stress-strain curve (green) simulated by Ginzburg-Landau theory at key characteristic points (such as maximum strain value and residual strain value).
[0031] 5) Randomly select 100 samples from the test set data (each sample represents a set of box-strain parameters), use the trained LSTM model to predict these samples, and then plot the true value (i.e., the true stress-strain curve) and the predicted value (i.e., the predicted stress-strain) of each sample on a scatter plot, fitting the ideal prediction diagonal (y=x) and 95% confidence interval, where the coefficient of determination R0 is... 2 =0.996, RMSE=0.0007, indicating high prediction accuracy of the LSTM model. The results are as follows: Figure 6 As shown.
[0032] 6) Randomly select 100 samples from the test set data (each sample represents a set of box-strain parameters), and use the trained LSTM model to predict these samples. Statistical analysis is performed on the true values (i.e., the true stress-strain curves) and the LSTM model predictions (i.e., the predicted stress-strain) of the entire test set. Probability density distribution plots (PDFs) of the true and predicted values are plotted. Histograms and kernel density estimation (KDE) methods are used to illustrate the distribution patterns of the two sets of data (i.e., the true and predicted values), and the analysis shows a high degree of consistency between the LSTM model predictions and the actual data in the overall distribution. Results are as follows: Figure 7 As shown.
[0033] 7) The time required to generate a complete curve using both the Ginzburg-Landau theoretical simulation and the LSTM model was statistically analyzed. Traditional Ginzburg-Landau theoretical simulation relies on the CPU for numerical calculations and takes 90 minutes. In contrast, the LSTM model-based method, performed on an NVIDIA GeForce RTX 3050 OEM graphics card (CUDA accelerated), fully utilizes the GPU's parallel computing capabilities, achieving a time of only 2 seconds, significantly improving the speed of generating a single stress-strain curve. The results are as follows... Figure 8 As shown.
[0034] Construction and Performance Evaluation of WCGAN-GP Network 1) A Wasserstein CGAN with gradient penalty (denoted as WCGAN-GP network) is used as the core model for inverse design. The generator conditional label input of the WCGAN-GP network is the target performance curve (81-dimensional) and random noise, and the output is a ten-dimensional vector (i.e., phase field parameters) of h1, s1, c1, h2, s2, c2, h3, s3, c3, T. The critic input of the WCGAN-GP network is random parameters and the target stress-strain curve, and the output is a real score. The loss function of the WCGAN-GP network is Wasserstein-1 distance plus gradient penalty, which enforces Lipschitz continuity. Training and evaluation methods: During training, the Critic is updated every 5 times, and the generator is updated once. The optimizer is Adam. The network construction is as follows. Figure 9 As shown.
[0035] 2) Dimensionality reduction analysis using t-SNE, such as... Figure 10 The data points generated by the WCGAN-GP network (Predicted, marked with "x") highly overlap with the distribution of the real data points (Real, marked with dots). Its clustering effect is significant: the data is clearly divided into three large clusters, corresponding to the "single-doped," "double-doped," and "triple-doped" cases, respectively. Within each cluster, the points generated by the model are perfectly blended with the real data points, almost indistinguishable, indicating that the WCGAN generator has accurately learned the inherent distribution patterns of the real parameters, and the parameters it generates possess the same diversity and realism as the real data.
[0036] 3) Using the training set, a new parameter space was constructed, and the data sources were labeled: Real, Reconstructed, and Novel. Single-doped, double-doped, and triple-doped data clusters each occupied independent and well-defined spatial regions, confirming that the WCGAN-GP network can clearly distinguish different tasks. "Novel" data points represent parameters generated by the model for target performance curves (i.e., target stress-strain curves) that did not appear in the training set. For example... Figure 11 The "Novel" point also falls perfectly within the corresponding category cluster, demonstrating the model's generalization ability. This WCGAN-GP network can not only reconstruct known data but also generalize to generate entirely new and effective parameters.
[0037] Reverse engineering and parameter selection: 1) Input the standard hyperelastic stress-strain curve into the trained WCGAN-GP network. Generate 20 sets of candidate parameters (i.e., phase field parameters) in batches for the target performance curve (i.e., the target stress-strain curve). The number of candidate parameter sets can be set in the WCGAN-GP network as needed. Input these candidate parameters into the LSTM model used as a surrogate model for fast forward prediction to obtain the predicted stress-strain curve corresponding to each set of phase field parameters. Calculate the similarity between the predicted stress-strain curve and the original target curve (i.e., the target stress-strain curve) (the similarity between the target stress-strain curve and the predicted stress-strain curve is the mean square error of the abscissa of the target stress-strain curve and the predicted stress-strain curve), and sort all 20 sets of candidate parameter schemes according to similarity. Select the set of phase field parameters with the highest similarity as the final phase field parameter (i.e., shape memory alloy composition) design result. Figure 12 The orange curve is the target curve, and the blue curve is the curve with the highest similarity. The parameters of the blue curve can be considered as the optimal phase field parameters. This set of parameters is selected for verification.
[0038] 2) Input the non-standard hyperelastic stress-strain curve into the trained WCGAN-GP network to generate 20 sets of candidate parameters for the target performance curve. Input these candidate parameters into an LSTM model (actually a surrogate model) for fast forward prediction, obtaining the predicted stress-strain curve for each set of parameters. Calculate the similarity between the predicted stress-strain curve and the original target stress-strain curve, rank all 20 candidate schemes, and select the parameter combination with the highest similarity as the final design result. Figure 13 The orange curve is the target curve, and the blue curve is the curve with the highest similarity. The parameters of the blue curve can be considered as the optimal phase field parameters. This set of parameters is selected for verification.
[0039] verify 1) Input the optimal phase field parameters obtained from the above standard hyperelastic stress-strain curve prediction into the phase field model to predict the corresponding stress-strain curve. The results are as follows: Figure 14 As shown, the predicted stress-strain curve is similar to the target stress-strain curve, which can be considered as a design of new phase field parameters.
[0040] 2) Input the optimal phase field parameters obtained from the above standard hyperelastic stress-strain curve prediction into the phase field model to predict the corresponding stress-strain curve. The results are as follows: Figure 15 As shown, the predicted stress-strain curve is roughly similar to the target stress-strain curve, which can be considered as a design of new phase field parameters.
[0041] As can be seen from the above scheme, this invention utilizes the WCGAN-GP network to achieve end-to-end reverse design of "target stress-strain curve → multi-component doping parameters," compressing the high-dimensional optimization that originally required several weeks to the second level. This framework not only fundamentally solves the two major problems of "slow calculation and inability to reverse" in SMA phase-field simulation, but also lays a solid foundation for the high-throughput development of novel multifunctional alloys. Furthermore, embodiments of the present invention also provide a system for implementing the above-described shape memory alloy composition prediction method of the present invention, the system comprising: Reverse design module: used to process the pre-set target stress-strain curve of the shape memory alloy through a pre-trained WCGAN-GP network to obtain multiple sets of phase field parameters corresponding to the target stress-strain curve; Forward prediction module: used to process the multiple sets of phase field parameters through the trained LSTM model to obtain the predicted stress-strain curves corresponding to each set of phase field parameters; The optimization module is used to calculate the similarity between the target stress-strain curve and each predicted stress-strain curve, filter the predicted stress-strain curves that meet the similarity requirements with the target stress-strain curve, take the predicted stress-strain curves that meet the similarity requirements as the optimal predicted stress-strain curves, and take the phase field parameters corresponding to the optimal predicted stress-strain curves as the final composition of the shape memory alloy.
[0042] The embodiments of the present invention also provide corresponding electronic devices and computer-readable storage media for implementing the solutions provided in the embodiments of the present invention.
[0043] The electronic device includes a storage device and one or more processors. The storage device stores instructions or code, and the processors execute the instructions or code to enable the device to perform the shape memory alloy composition prediction method described in any embodiment of this application.
[0044] The storage medium stores a computer program, which, when executed by a processor, implements the shape memory alloy composition prediction method described in any embodiment of this application.
[0045] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.
[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A method for predicting the composition of shape memory alloys, characterized in that, The process includes the following: The pre-defined target stress-strain curve of the shape memory alloy is processed by a pre-trained WCGAN-GP network to obtain multiple sets of phase field parameters corresponding to the target stress-strain curve. The multiple sets of phase field parameters are processed by the trained LSTM model to obtain the predicted stress-strain curves corresponding to each set of phase field parameters. The similarity between the target stress-strain curve and each predicted stress-strain curve is calculated. Predicted stress-strain curves that meet the similarity requirements with the target stress-strain curve are selected. The predicted stress-strain curves that meet the similarity requirements are taken as the optimal predicted stress-strain curves. The phase field parameters corresponding to the optimal predicted stress-strain curves are taken as the final composition of the shape memory alloy.
2. The method for predicting the composition of shape memory alloys according to claim 1, characterized in that, In the WCGAN-GP network, the critic input consists of random parameters and the target stress-strain curve, and the output is a real number score. The loss function is Wasserstein-1 distance plus gradient penalty, which enforces Lipschitz continuity.
3. The method for predicting the composition of shape memory alloys according to claim 2, characterized in that, During training and evaluation of the WCGAN-GP network, the generator is updated once for every 5 updates by the critic, and the optimizer is Adam.
4. The method for predicting the composition of shape memory alloys according to claim 1, characterized in that, The LSTM model is trained using the mean squared error loss function, with Adam as the optimizer and an initial learning rate of 1e. -4 To prevent overfitting, a "teacher-forced" strategy is used during training to accelerate convergence, while autoregressive generation is used during the inference phase.
5. The method for predicting the composition of shape memory alloys according to claim 1, characterized in that, The process of selecting predicted stress-strain curves that meet the similarity requirements with the target stress-strain curve includes: When the similarity between the target stress-strain curve and each predicted stress-strain curve is within the preset similarity range, it indicates that the similarity between the predicted stress-strain curve and the target stress-strain curve meets the requirements; otherwise, it indicates that the similarity between the predicted stress-strain curve and the target stress-strain curve does not meet the requirements.
6. The method for predicting the composition of a shape memory alloy according to claim 5, characterized in that, When the similarity between the predicted stress-strain curve and the target stress-strain curve does not meet the requirements, the target stress-strain curve is adjusted to obtain a new target stress-strain curve. The shape memory alloy composition prediction method is then repeated using the new target stress-strain curve until the similarity between the predicted stress-strain curve and the target stress-strain curve meets the requirements.
7. The method for predicting the composition of shape memory alloys according to claim 1, characterized in that, The phase field parameters include the temperature at which the stress-strain curve is generated, as well as the atomic size parameters of each metal element in the shape memory alloy and the doping concentration of each metal element. The similarity between the target stress-strain curve and the predicted stress-strain curve is the mean square error of the abscissa of the target stress-strain curve and the predicted stress-strain curve.
8. A shape memory alloy composition prediction system, characterized in that, include: Reverse design module: used to process the pre-set target stress-strain curve of the shape memory alloy through a pre-trained WCGAN-GP network to obtain multiple sets of phase field parameters corresponding to the target stress-strain curve; Forward prediction module: used to process the multiple sets of phase field parameters through the trained LSTM model to obtain the predicted stress-strain curves corresponding to each set of phase field parameters; The optimization module is used to calculate the similarity between the target stress-strain curve and each predicted stress-strain curve, filter the predicted stress-strain curves that meet the similarity requirements with the target stress-strain curve, take the predicted stress-strain curves that meet the similarity requirements as the optimal predicted stress-strain curves, and take the phase field parameters corresponding to the optimal predicted stress-strain curves as the final composition of the shape memory alloy.
9. An electronic device, characterized in that, include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the shape memory alloy composition prediction method as described in any one of claims 1 to 7.
10. A storage medium, characterized in that, It stores a computer program, wherein the computer program, when executed by a processor, implements the shape memory alloy composition prediction method as described in any one of claims 1-7.