A method for predicting phase transition temperature of additive manufacturing NiTi-based shape memory alloy based on physical feature transfer learning

Abstract

This invention discloses a method for predicting the phase transition temperature of NiTi-based shape memory alloys in additive manufacturing based on physical feature transfer learning, belonging to the field of additive manufacturing technology. Addressing the scarcity of experimental data for NiTi-based alloys in additive manufacturing processes and the prediction failure caused by domain drift compared to traditional smelting processes, a hybrid optimization framework is proposed: First, a multi-source dataset is constructed containing a large amount of smelting data and a small amount of additive manufacturing experimental data (target domain); second, physical information feature engineering is implemented to map atomic components to physical features to enhance feature transferability; subsequently, an instance weight transfer learning algorithm is used to assign weight coefficients to the target domain support set to reconstruct the loss function, forcing the model to learn specific patterns; finally, transfer learning is used to optimize the model to achieve prediction of the target domain validation set. This invention achieves accurate prediction of the phase transition temperature with extremely small sample sizes, significantly shortening the R&D cycle of additive manufacturing shape memory alloys.

Description

Technical Field

[0001] This invention belongs to the field of additive manufacturing technology, and in particular relates to a method for predicting the phase transformation temperature of NiTi-based shape memory alloys based on physical feature transfer learning. Background Technology

[0002] NiTi-based shape memory alloys, with their superior shape memory effect, superelasticity, and good biocompatibility, have become indispensable functional materials in aerospace, biomedicine, and robotics. To meet the specific phase transformation temperature requirements of different applications, constructing a NiTiX ternary alloy system by introducing a third alloying element (X) is the mainstream strategy for controlling the phase transformation temperature. Additive manufacturing, with its high deposition efficiency, high material utilization, and cost advantages in complex components, has become an effective in-situ alloying technology for preparing NiTi-based alloy components. However, the complex thermal history during additive manufacturing leads to a non-equilibrium state in the alloy microstructure, causing a significant deviation in the phase transformation temperature of NiTi-based alloys with the same composition under additive manufacturing compared to alloys produced by traditional melting processes. This systematic shift in the composition-property mapping relationship caused by the manufacturing process is called domain drift.

[0003] Currently, the design of novel additive manufacturing NiTi-based alloys still primarily relies on empirical trial-and-error methods. This approach is time-consuming and costly. Although machine learning methods in materials genome engineering have shown great potential in accelerating materials discovery, they still face significant challenges in the field of additive manufacturing. On the one hand, the high cost and long cycle of additive manufacturing experiments result in a very small size of proprietary datasets that can be directly used for model training, making it difficult to support the training of deep learning or complex regression models. On the other hand, most existing material data originates from traditional smelting processes. Due to the aforementioned domain drift, when directly applying models trained on smelting data to predict the properties of additively manufactured alloys, the prediction accuracy often drops significantly, and predictions may even fail.

[0004] In summary, how to effectively utilize massive amounts of historical data from traditional smelting to help predict the phase transformation temperature of small-sample additive manufacturing alloys, while integrating physical mechanisms into data-driven models to address prediction biases caused by process differences, is a key technical challenge that urgently needs to be solved in the current field of additive manufacturing NiTi-based shape memory alloy design. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for predicting the phase transformation temperature of NiTi-based shape memory alloys in additive manufacturing based on physical feature transfer learning. The aim is to optimize a model based on a large amount of smelting data through the synergistic effect of physical information feature engineering and instance weight transfer learning, thereby achieving the goal of accurately predicting the phase transformation temperature of small-sample additive manufacturing alloys.

[0006] 1. The technical solution of the present invention is: a method for predicting the phase transformation temperature of NiTi-based shape memory alloys in additive manufacturing based on physical feature transfer learning, characterized by comprising the following steps:

[0007] A. Construct a multi-source dataset for NiTi-based alloys, including a large amount of literature data on NiTi-based alloys under traditional smelting processes, and a small amount of experimental data on NiTi-based alloys manufactured by additive manufacturing.

[0008] A1. Mining and cleaning literature data on NiTi-based alloys prepared by traditional smelting processes as source domain datasets;

[0009] A2. Prepare and measure experimental data of NiTi-based alloys under additive manufacturing processes as the target domain dataset;

[0010] A3. The original input of the multi-source dataset is the atomic percentage of all elements, and the output is the phase transformation temperature, including the end temperature of austenite phase transformation Af and the start temperature of martensite phase transformation Ms.

[0011] B. Based on the multi-source dataset in step A, perform feature engineering based on physical feature information. Calculate a set of physical descriptors that can characterize the thermodynamic laws of phase transition using the principles of materials science. Map the original component features to the physical enhancement feature space and use the physical information processed by feature engineering as feature input.

[0012] C. Based on the feature engineering multi-source dataset in step B, perform transfer learning optimization. The specific steps are as follows:

[0013] C1. Extract a small number of samples from the target domain dataset as a calibration set, and merge them with the source domain dataset to construct a hybrid dataset;

[0014] C2. The loss function of the mixed dataset is reconstructed by introducing instance weight assignment. The formula for the generalized loss function J(θ) is:

[0015]

[0016] C3, Instance Weight Allocation w i Prioritize smaller calibration sets, defined as follows: when a data point belongs to the source domain dataset, the instance weight coefficient is 1; when a data point belongs to the calibration set, the instance weight coefficient is λ.

[0017] C4. Employ multiple machine learning models to apply a transfer learning training strategy based on instance weight allocation to a mixed dataset;

[0018] C5. Perform rigorous grid search combined with cross-validation on each model to simultaneously optimize the model-specific hyperparameters and instance weight coefficients;

[0019] D. Based on the transfer learning training and optimization of all machine learning models in step C, predict the phase transition temperatures Af and Ms of the validation set in the target domain dataset.

[0020] E. Based on the prediction results of all optimized models in step D, compare the model fit and prediction performance using multiple evaluation metrics.

[0021] 2. The method as described in claim 1, wherein in step A, the third element in the NiTi-based ternary alloy is selected from Cu, Nb, and Hf;

[0022] 3. The method as described in claim 1, characterized in that, in step A2, the target domain dataset, due to its limited data volume, strictly adopts a calibration set-validation set partitioning scheme, wherein the calibration set participates in the transfer learning optimization process, and the validation set is used to rigorously test the model's generalization ability on unseen samples;

[0023] 4. The method as described in claim 1, characterized in that, in step B, the physical feature information processed by feature engineering includes valence electron concentration VEC, atomic size difference δ, electronegativity difference ∆χ, and enthalpy of mixing ∆H. mix Mixed entropy ∆S mix ;

[0024] 5. The method as described in claim 1, wherein in step C, the machine learning model is first pre-trained using the source domain dataset;

[0025] 6. The method as described in claim 1, wherein in step C3, the instance weight coefficient λ of the calibration set ranges from 1 to 50;

[0026] 7. The method as described in claim 1, wherein in step C4, the plurality of machine learning models include Lasso Regression, Ridge Regression, ElasticNet, Decision Tree, Random Forest Regression, Extra Trees, Gradient Boosting Regressor, AdaBoost, XGBoost, LightGBM, K-Nearest Neighbors, Support Vector Regression, etc.;

[0027] 8. The method as described in claim 1, wherein in step E, the evaluation metrics for model fit and predictive performance include the coefficient of determination R. 2Mean absolute error (MAE) and root mean square error (RMSE);

[0028] 9. The method as described in claim 1, wherein the domain drift between the additive manufacturing process and the conventional smelting process is corrected by physical information feature engineering in step B and instance weight transfer learning in step C, thereby enabling the prediction of the phase transformation temperature of the NiTi-based alloy in the additive manufacturing process.

[0029] The beneficial effects of this invention are as follows: This invention proposes a method for predicting the phase transformation temperature of NiTi-based shape memory alloys in additive manufacturing based on physical feature transfer learning. Through the synergistic effect of physical information feature engineering and instance weight transfer learning, it solves the problem of cross-process prediction bias in the context of small sample sizes in additive manufacturing, achieving effective knowledge transfer from large-scale traditional smelting data to small-scale additive manufacturing data. This significantly improves the design and development efficiency of novel NiTi-based additive manufacturing alloys and provides a low-cost technical solution for the customized production of complex functional components. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of a method for predicting the phase transformation temperature of NiTi-based shape memory alloys based on physical feature transfer learning in additive manufacturing, as described in this invention.

[0031] Figure 2 This is a comparison of the goodness of fit and prediction performance of the original training model, the physical information feature engineering processing model, and the instance weight transfer learning optimization model in this invention on the target domain validation set: (ac) R 2 ;(df) MAE;(gi) RMSE;

[0032] Figure 3 These are the prediction results of the optimized Extra Trees (ET) model in this invention on the training set, test set, and validation set: (a) Af result; (b) Ms result.

[0033] Figure 4 The following are the predicted and actual phase transition temperatures of three representative dual-wire arc additive manufacturing NiTiCu, NiTiNb, and NiTiHf alloys in this invention: (a) Af result; (b) Ms result. Detailed Implementation

[0034] To make the objectives and technical solutions of this invention clearer, detailed explanations are provided below with reference to specific examples. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this invention.

[0035] like Figure 1The diagram shown illustrates the application process of a method for predicting the phase transformation temperature of NiTi-based shape memory alloys in additive manufacturing based on physical feature transfer learning, according to the present invention. The application of this method includes the following steps:

[0036] A. Construct a multi-source dataset for NiTi-based alloys, including a large amount of literature data on NiTi-based alloys under traditional smelting processes, and a small amount of experimental data on NiTi-based alloys manufactured by additive manufacturing.

[0037] A1. Mining and cleaning literature data on NiTi-based alloys prepared by traditional smelting processes as source domain datasets;

[0038] A2. Prepare and measure experimental data of NiTi-based alloys under additive manufacturing processes as the target domain dataset;

[0039] A3. The original input of the multi-source dataset is the atomic percentage of all elements, and the output is the phase transformation temperature, including the end temperature of austenite phase transformation Af and the start temperature of martensite phase transformation Ms.

[0040] B. Based on the multi-source dataset in step A, perform feature engineering based on physical feature information. Calculate a set of physical descriptors that can characterize the thermodynamic laws of phase transition using the principles of materials science. Map the original component features to the physical enhancement feature space and use the physical information processed by feature engineering as feature input.

[0041] C. Based on the feature engineering multi-source dataset in step B, perform transfer learning optimization. The specific steps are as follows:

[0042] C1. Extract a small number of samples from the target domain dataset as a calibration set, and merge them with the source domain dataset to construct a hybrid dataset;

[0043] C2. The loss function of the mixed dataset is reconstructed by introducing instance weight assignment. The formula for the generalized loss function J(θ) is:

[0044]

[0045] C3, Instance Weight Allocation w i Prioritize smaller calibration sets, defined as follows: when a data point belongs to the source domain dataset, the instance weight coefficient is 1; when a data point belongs to the calibration set, the instance weight coefficient is λ.

[0046] C4. Employ multiple machine learning models to apply a transfer learning training strategy based on instance weight allocation to a mixed dataset;

[0047] C5. Perform rigorous grid search combined with cross-validation on each model to simultaneously optimize the model-specific hyperparameters and instance weight coefficients;

[0048] D. Based on the transfer learning training and optimization of all machine learning models in step C, predict the phase transition temperatures Af and Ms of the validation set in the target domain dataset.

[0049] E. Based on the prediction results of all optimized models in step D, compare the model fit and prediction performance using multiple evaluation metrics.

[0050] This invention takes the dual-wire arc additive manufacturing of NiTi-based shape memory alloys as an example. The two wires are, respectively, a NiTi (50.5 at.% Ni) alloy wire and one of a high-purity copper wire, niobium wire, or hafnium wire. The arc additive manufacturing employs a robotic tungsten inert gas deposition system equipped with an ultra-high frequency pulsed arc power supply.

[0051] A. Construct a multi-source dataset for NiTi-based alloys, including a large amount of literature data on NiTi-based alloys under traditional smelting processes, and a small amount of experimental data on NiTi-based alloys manufactured by dual-wire arc additive manufacturing.

[0052] In step A, the third element in the NiTi-based ternary alloy is selected from one of Cu, Nb, and Hf;

[0053] A1. Mining and cleaning literature data on NiTi-based alloys prepared by traditional smelting processes as source domain datasets;

[0054] In step A1, data on NiTi-based alloys prepared by traditional smelting processes, including arc melting and induction melting, were mined from academic papers. After screening and cleaning, 152 data points were finally selected as the source domain dataset.

[0055] A2. Prepare and measure experimental data of NiTi-based alloys under the dual-wire arc additive manufacturing process as the target domain dataset;

[0056] In step A2, the process for preparing NiTi-based alloys using dual-wire arc additive manufacturing is as follows: base current 80-140 A, rectangular ultra-high frequency pulse current 60-70 A, deposition rate 180-360 mm / min, NiTi wire feed speed 2200 mm / min, and high-purity argon protective gas flow rate 15 L / min. The X-wire feed speed is adjusted according to the NiTi-based alloy composition and the NiTi wire feed speed. Differential scanning calorimetry is used to measure the phase transition temperature at a heating and cooling rate of 10 °C / min. A total of 14 data points are prepared as the target domain dataset. Due to the limited amount of data in the target domain dataset, a strict calibration set-validation set partitioning scheme is adopted. The calibration set participates in the transfer learning optimization process, and the validation set is used to rigorously test the model's generalization ability on unseen samples.

[0057] A3. The original input of the multi-source dataset is the atomic percentage of all elements, and the output is the phase transformation temperature, including the end temperature of austenite phase transformation Af and the start temperature of martensite phase transformation Ms.

[0058] In step A3, the original input in the multi-source dataset is specifically the atomic percentage of five elements: Ni, Ti, Cu, Nb, and Hf.

[0059] B. Based on the multi-source dataset in step A, perform feature engineering based on physical feature information. Calculate a set of physical descriptors that can characterize the thermodynamic laws of phase transition using the principles of materials science. Map the original component features to the physical enhancement feature space and use the physical information processed by feature engineering as feature input.

[0060] In step B, the physical characteristic information processed by feature engineering includes valence electron concentration (VEC), atomic size difference (δ), electronegativity difference (∆χ), and enthalpy of mixing (∆H). mix Mixed entropy ∆S mix In addition to these descriptions based on physical information, elemental composition ratios are also introduced as supplementary feature inputs, including Ni / Ti, Cu / (Ni+Ti), Nb / (Ni+Ti) and Hf / (Ni+Ti).

[0061] C. Based on the feature engineering multi-source dataset in step B, perform transfer learning optimization. The specific steps are as follows:

[0062] In step C, the machine learning model is first pre-trained using the source domain dataset; during the pre-training process, the source domain dataset is divided into a training set and a test set in a 7:3 ratio.

[0063] C1. Extract a small number of samples from the target domain dataset as a calibration set, and merge them with the source domain dataset to construct a hybrid dataset;

[0064] In step C1, five data points are selected from the target domain dataset as a calibration set;

[0065] C2. The loss function of the mixed dataset is reconstructed by introducing instance weight assignment. The formula for the generalized loss function J(θ) is:

[0066]

[0067] C3, Instance Weight Allocation w i Prioritize smaller calibration sets, defined as follows: when a data point belongs to the source domain dataset, the instance weight coefficient is 1; when a data point belongs to the calibration set, the instance weight coefficient is λ.

[0068] In step C3, the instance weight coefficient λ of the calibration set ranges from 1 to 50;

[0069] C4. Employ multiple machine learning models to apply a transfer learning training strategy based on instance weight allocation to a mixed dataset;

[0070] In step C4, multiple machine learning models include Lasso Regression (LR), Ridge Regression (RR), ElasticNet (EN), Decision Tree (DT), Random Forest Regression (RFR), Extra Trees (ET), Gradient Boosting Regressor (GBR), AdaBoost (AdaB), XGBoost (XGB), LightGBM (LGBM), K-Nearest Neighbors (KNN), Support Vector Regression (SVR), etc.

[0071] C5. Perform rigorous grid search combined with cross-validation on each model to simultaneously optimize the model-specific hyperparameters and instance weight coefficients;

[0072] In step C5, the grid search of 5-fold cross-validation is used to optimize the model hyperparameters and instance weight coefficients;

[0073] D. Based on the transfer learning training and optimization of all machine learning models in step C, predict the phase transition temperatures Af and Ms of the validation set in the target domain dataset.

[0074] In step D, nine data points in the target domain dataset are used as the validation set, and the corresponding actual experimental phase transition temperatures are shown in Table 1.

[0075] Table 1. Actual experimental phase transition temperatures of the target domain validation set.

[0076] Serial Number Real Af (°C) Real Ms (°C) #1 26.4 11.6 #2 31.9 12.7 #3 47.6 27.2 #4 125.0 36.0 #5 75.7 34.5 #6 72.5 34.1 #7 71.2 33.8 #8 77.0 35.0 #9 330.4 274.0

[0077] E. Based on the prediction results of all optimized models in step D, compare the model fit and prediction performance using multiple evaluation metrics.

[0078] In step E, the evaluation metrics for model fit and predictive performance include the coefficient of determination R. 2 Mean absolute error (MAE) and root mean square error (RMSE); comparison of the fit and prediction performance of the original training model, the physical information feature engineering processing model, and the instance weight transfer learning optimization model on the target domain validation set, such as... Figure 2 As shown, the prediction performance of each model is significantly improved after feature engineering and transfer learning, with the optimized Extra Trees (ET) model being the best performing prediction model. The ET model exhibits high-precision prediction convergence on the training, test, and validation sets, as shown in the figure. Figure 3 As shown; further, three representative NiTiCu, NiTiNb, and NiTiHf alloys were selected in the validation set for dual-wire arc additive manufacturing, and the predicted and actual values ​​of the phase transformation temperature were compared, as shown. Figure 4 As shown, the proposed hybrid optimization framework can accurately predict the phase transformation temperature of alloys manufactured in small-sample dual-wire arc additive manufacturing.

[0079] The domain drift between the dual-wire arc additive manufacturing process and the conventional smelting process is corrected through physical information feature engineering in step B and instance weight transfer learning in step C, thereby enabling the prediction of the NiTiX phase transition temperature in the dual-wire arc additive manufacturing process.

[0080] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for predicting the phase transformation temperature of additive manufacturing NiTi-based shape memory alloys based on physical feature transfer learning, characterized in that, Includes the following steps: A. Construct a multi-source dataset for NiTi-based alloys, including a large amount of literature data on NiTi-based alloys under traditional smelting processes, and a small amount of experimental data on NiTi-based alloys manufactured by additive manufacturing. A1. Mining and cleaning literature data on NiTi-based alloys prepared by traditional smelting processes as source domain datasets; A2. Prepare and measure experimental data of NiTi-based alloys under additive manufacturing processes as the target domain dataset; A3. The original input of the multi-source dataset is the atomic percentage of all elements, and the output is the phase transformation temperature, including the end temperature of austenite phase transformation Af and the start temperature of martensite phase transformation Ms. B. Based on the multi-source dataset in step A, perform feature engineering based on physical feature information. Calculate a set of physical descriptors that can characterize the thermodynamic laws of phase transition using the principles of materials science. Map the original component features to the physical enhancement feature space and use the physical information processed by feature engineering as feature input. C. Based on the feature engineering multi-source dataset in step B, perform transfer learning optimization. The specific steps are as follows: C1. Extract a small number of samples from the target domain dataset as a calibration set, and merge them with the source domain dataset to construct a hybrid dataset; C2. The loss function of the mixed dataset is reconstructed by introducing instance weight assignment. The formula for the generalized loss function J(θ) is: C3, Instance Weight Allocation w i Prioritize smaller calibration sets, defined as follows: when a data point belongs to the source domain dataset, the instance weight coefficient is 1; when a data point belongs to the calibration set, the instance weight coefficient is λ. C4. Employ multiple machine learning models to apply a transfer learning training strategy based on instance weight allocation to a mixed dataset; C5. Perform rigorous grid search combined with cross-validation on each model to simultaneously optimize the model-specific hyperparameters and instance weight coefficients; D. Based on the transfer learning training and optimization of all machine learning models in step C, predict the phase transition temperatures Af and Ms of the validation set in the target domain dataset. E. Based on the prediction results of all optimized models in step D, compare the model fit and prediction performance using multiple evaluation metrics.

2. The method as described in claim 1, characterized in that, In step A, the third element in the NiTi-based ternary alloy is selected from one of Cu, Nb, and Hf.

3. The method as described in claim 1, characterized in that, In step A2, due to the limited amount of data in the target domain dataset, a strict calibration set-validation set partitioning scheme is adopted. The calibration set participates in the transfer learning optimization process, while the validation set is used to rigorously test the model's generalization ability on unseen samples.

4. The method as described in claim 1, characterized in that, In step B, the physical characteristic information processed by feature engineering includes valence electron concentration (VEC), atomic size difference (δ), electronegativity difference (∆χ), and enthalpy of mixing (∆H). mix Mixed entropy ∆S mix。 5. The method as described in claim 1, characterized in that, In step C, the machine learning model is first pre-trained using the source domain dataset.

6. The method as described in claim 1, characterized in that, In step C3, the instance weight coefficient λ of the calibration set ranges from 1 to 50.

7. The method as described in claim 1, characterized in that, In step C4, multiple machine learning models include Lasso Regression, Ridge Regression, ElasticNet, Decision Tree, Random Forest Regression, Extra Trees, Gradient Boosting Regressor, AdaBoost, XGBoost, LightGBM, K-Nearest Neighbors, Support Vector Regression, etc.

8. The method as described in claim 1, characterized in that, In step E, the evaluation metrics for model fit and predictive performance include the coefficient of determination R. 2 Mean absolute error (MAE) and root mean square error (RMSE).

9. The method as described in claim 1, characterized in that, The domain drift between the additive manufacturing process and the conventional smelting process is corrected through physical information feature engineering in step B and instance weight transfer learning in step C, thereby enabling the prediction of the phase transformation temperature of NiTi-based alloys in the additive manufacturing process.

Images