A method for predicting thermal hysteresis ∆Thys of Ni-Mn based shape memory alloys based on machine learning

Abstract

A method for predicting the thermal hysteresis ΔThys of Ni-Mn-based shape memory alloys based on machine learning. The present invention specifically relates to a method for predicting the thermal hysteresis ΔThys of Ni-Mn-based shape memory alloys based on machine learning. The present invention is to solve the problem that the prior art lacks the thermal hysteresis ΔT of Ni-Mn-based shape memory alloys. hys Problems with prediction methods. The eigenvalues ​​and volume change rates of all Ni‑Mn-based shape memory alloys are obtained through first-principles calculations combined with machine learning, and added as key physical coupling features to the Ni‑Mn-based shape memory alloy thermal hysteresis dataset. Different machine learning models are established with elements and physical properties as features and thermal hysteresis as labels, and model tuning and model selection are performed. Finally, a random forest model is used to predict the thermal hysteresis of the new Ni‑Mn-based shape memory alloy, and a new narrow hysteresis Ni‑Mn-based shape memory alloy is obtained through experimental preparation.

Description

Technical Field

[0001] The present invention specifically relates to a method for predicting thermal hysteresis ΔThys of Ni-Mn based shape memory alloy based on machine learning. Background Art

[0002] With the rapid development of semiconductor, medical, biological and mechanical fields, the requirements of micro-electromechanical systems (MEMS) for micro-actuators and micro-sensors are increasing day by day, and advanced MEMS materials can improve the performance of MEMS actuators, sensors and systems. The materials used in MEMS can be divided into structural materials and functional materials according to the functions of the components and the manufacturing methods. The structural materials are mainly semiconductor silicon and its compounds, and the functional materials are mainly shape memory materials, magnetostrictive materials, piezoelectric materials, etc. Among them, shape memory alloys have dual functions of sensing and driving due to their unique shape memory effect (SME) and superelasticity (SE). They can be effectively used as driving elements or force-sensitive and heat-sensitive sensing components, and have attracted widespread attention. Traditional shape memory alloys (SMAs) achieve the purpose of driving through the deformation caused by martensitic phase transformation, and achieve the purpose of sensing through the temperature change or strain caused by phase transformation. Although SMAs can produce large reversible recovery strain, their response frequency is not high, so the response speed of sensors and actuators made of them needs to be further improved. Ni-Mn-based shape memory alloy is a new type of shape memory alloy. It has the characteristics of shape memory effect, superelasticity, and sensitivity to external temperature changes of traditional shape memory alloys. It can also induce shape memory effect by magnetic field. It has simple driving mode, rapid reaction and high response frequency, becoming a new type of intelligent material integrating "sensing" and "driving". Therefore, Ni-Mn-based alloy is expected to replace traditional shape memory alloys in micro-electromechanical systems and develop into a large strain, high response frequency driver and sensor material.

[0003] Ni-Mn based shape memory alloys have good temperature and magnetic field induced shape memory effect, superelasticity, magnetocaloric effect, magnetoresistance effect, elastocaloric effect and other functional properties. However, in the current study, due to the lattice mismatch between the austenite phase and the martensite phase, thermal hysteresis (ΔT hys ) phenomenon. The size of thermal hysteresis is very important for the alloy to obtain stable functional characteristics under the action of cyclic external fields. The hysteresis width determines the energy loss and conversion efficiency, and narrow hysteresis means fewer defects during the cycle, higher fatigue life, and is conducive to the reversibility of phase transformation. Therefore, minimizing ΔT hys It is very important to realize the application of Ni-Mn based SMAs. Therefore, a Ni-Mn based shape memory alloy thermal hysteresis ΔT is urgently needed. hys prediction method. Summary of the invention

[0004] The present invention aims to solve the problem that the prior art lacks the thermal hysteresis ΔT of Ni-Mn based shape memory alloy. hys In order to solve the problem of prediction method, a method for predicting thermal hysteresis ΔThys of Ni-Mn based shape memory alloy based on machine learning is provided.

[0005] The present invention proposes a method for predicting thermal hysteresis ΔThys of Ni-Mn-based shape memory alloy based on machine learning, which is specifically carried out in the following steps:

[0006] 1. Collecting ΔT from the literature hys Data information, build data sets;

[0007] Second, the eigenvalue λ2 and volume change rate ΔV / V are obtained by first-principles calculation combined with the ML model;

[0008] 3. Perform feature engineering, model building and model selection through machine learning to obtain the optimal prediction model;

[0009] 4. Use the optimal prediction model to predict the thermal hysteresis of Ni-Mn based shape memory alloy.

[0010] Beneficial effects of the present invention:

[0011] The present invention combines first-principles calculations, machine learning and experiments to accelerate the material research and development process. It can greatly reduce time and money costs, quickly screen out available target alloys, and achieve the theoretical design of narrow hysteresis Ni-Mn-based shape memory alloys, thereby accelerating the exploration of Ni-Mn-based shape memory alloys.

[0012] The thermal hysteresis of the Ni-Mn-Sn based shape memory alloy material prepared by the present invention is predicted to be less than 8K, and the hysteresis is small, which is beneficial to the reversibility of the phase change. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 For you 24 Mn 18 Schematic diagram of the crystal structure of Sn6;

[0014] Figure 2 is the random forest scatter plot of the eigenvalue (λ2);

[0015] Figure 3 is the random forest scatter plot of volume change rate (ΔV / V);

[0016] Figure 4 R for different machine learning models 2 Comparison chart;

[0017] Figure 5 Predict results for the RFR model;

[0018] Figure 6 For you 45 Mn 45 DSC curves of Sn8X2 (X = Sn, Al, Si, B and Ga);

[0019] Figure 7 Comparison chart of prediction and experimental data. DETAILED DESCRIPTION

[0020] The technical solution of the present invention is not limited to the specific implementation modes listed below, but also includes any combination of the specific implementation modes.

[0021] Specific implementation method 1: In this implementation method, a method for predicting thermal hysteresis ΔThys of Ni-Mn-based shape memory alloy based on machine learning is specifically performed in the following steps:

[0022] 1. Collecting ΔT from the literature hys Data information, build data sets;

[0023] Second, the eigenvalue λ2 and volume change rate ΔV / V are obtained by first-principles calculation combined with the ML model;

[0024] 3. Perform feature engineering, model building and model selection through machine learning to obtain the optimal prediction model;

[0025] 4. Use the optimal prediction model to predict the thermal hysteresis of Ni-Mn based shape memory alloy.

[0026] Specific implementation method 2: This implementation method is different from the specific implementation method 1 in that: Step 1 collects ΔT from the literature hys The data information is 462 sets of data collected from the published literature data of Ni-Mn-based SMAs, including the basic data Ni, Mn and X (X = In, Sn, Sb) elements, and ten doping elements Ga, Ti, Fe, Co, Pd, Cu, Al, Si, V and B. The rest is the same as the first embodiment.

[0027] Specific implementation method three: This implementation method is different from the specific implementation method one in that: the ΔT in step one hys =(A s +A f -M s -M f ) / 2; where M s is the martensitic transformation starting temperature, M f is the martensitic transformation end temperature, A s is the starting temperature of reverse martensitic transformation, A f is the end temperature of the reverse martensitic transformation. Other aspects are the same as those of the first embodiment.

[0028] Specific implementation method 4: This implementation method is different from the specific implementation method 2 in that: ΔT hys The data information includes the phase transition temperature data of the alloy and the properties of the alloy itself; the properties of the alloy itself include the number of protons, the number of valence electrons, atomic radius, melting point, minimum oxidation state, maximum oxidation state, resistivity, molar volume, thermal conductivity, boiling point, linear thermal expansion coefficient, electronegativity, ionization energy, relative atomic mass, and metal radius; the properties of all elements in the alloy are multiplied by their proportion in the entire alloy through the formula, and the result is used as the physical properties of the final alloy. These properties are ΔT hys The structure-property relationship modeling provides a physical basis. The rest is the same as the second specific implementation method.

[0029] Specific implementation mode 5: This implementation mode is different from the specific implementation mode 1 in that the eigenvalue λ2 and the volume change rate ΔV / V in step 2 are obtained by calculating the lattice constant. The rest is the same as the specific implementation mode 1.

[0030] Specific embodiment 6: This embodiment differs from specific embodiment 5 in that the lattice parameters are calculated by using Vienna Ab-Initio Simulation Package based on density functional theory. The rest is the same as specific embodiment 5.

[0031] Specific implementation method 7: This implementation method is different from the specific implementation method 1 in that: after step 2 is completed, data cleaning is performed and 205 sets of data are retained. The rest is the same as the specific implementation method 1.

[0032] Specific implementation eight: This implementation is different from specific implementation one in that: in step three, 77 component alloys of different components are selected, and their lattice constants a0, a, b, c and angle β are calculated using the first principle; λ2 and ΔV / V of the above 77 component alloys are calculated by formula; machine learning models of λ2 and ΔV / V are established respectively, with elements and physical properties as features, λ2 and ΔV / V as labels, and the relationship between features and labels is established. The rest is the same as specific implementation one.

[0033] Specific implementation method 9: This implementation method is different from the specific implementation method 1 in that the optimal prediction model in step 3 is a Random Forest Regressor. The rest is the same as the specific implementation method 1.

[0034] Specific embodiment 10: This embodiment differs from the specific embodiment 1 in that: the ternary alloy Ni with narrow thermal hysteresis is obtained through the prediction in step 4. 45 Mn 45 Sn 10 and the ternary alloy Ni 45 Mn45 Sn 10 The specific preparation process is as follows: Ni, Mn, Sn, Al, Si, Ga, and B with a purity of 99.99wt.% are used as raw materials, and a non-consumable vacuum arc melting furnace is used to melt and prepare under the protective atmosphere of argon gas. First, a mechanical pump and a molecular pump are used to evacuate to 4×10 -3 Pa, and then backfilled with high-purity argon to 0.07MPa for gas washing, and smelting was carried out after four cycles; all sample ingots were magnetically stirred during the smelting process, and repeatedly turned over and smelted 5 times to ensure the accuracy and uniformity of the chemical composition of the alloy, and then taken out after cooling, and cut along the diameter by wire cutting; all alloy samples were first mechanically polished to remove surface impurities, and then cleaned with acetone and alcohol and packaged in a vacuum of 10 -3 Pa quartz tube, kept at 1173K for 1 hour for homogenization annealing, and quenched in ice water. Others are the same as the first embodiment.

[0035] The following examples are used to verify the beneficial effects of the present invention:

[0036] Embodiment 1: In this embodiment, a method for predicting thermal hysteresis ΔThys of Ni-Mn-based shape memory alloy based on machine learning is specifically performed in the following steps:

[0037] 1. Collecting ΔT from the literature hys Data information, data set establishment; 462 sets of data were collected from the published literature on Ni-Mn-based SMAs, and thermal hysteresis was defined as ΔT hys =(A s +A f -M s -M f ) / 2. Among them, M s is the martensitic transformation starting temperature, M f is the martensitic transformation end temperature, A s is the starting temperature of reverse martensitic transformation, A f is the end temperature of the reverse martensitic phase transformation. It includes the basic data Ni, Mn and X (X = In, Sn, Sb) elements, as well as ten doping elements Ga, Ti, Fe, Co, Pd, Cu, Al, Si, V and B;

[0038] In order to predict the ΔT of Ni-Mn based alloys hysIn addition to the phase transition temperature data of the alloy, the properties of the alloy itself are also required, including the number of protons, the number of valence electrons, atomic radius, melting point, minimum oxidation state, maximum oxidation state, resistivity, molar volume, thermal conductivity, boiling point, linear thermal expansion coefficient, electronegativity, ionization energy, relative atomic mass, metal radius, etc. The properties of all elements in the alloy are multiplied by their proportion in the entire alloy through the formula, and the result is used as the properties of the final alloy. These properties are ΔT hys It provides a relatively simple physical basis for modeling the structure-property relationship.

[0039] 2. Using first-principles calculations combined with the ML model to obtain the eigenvalue λ2 and volume change rate ΔV / V; ΔT of Ni-Mn-based shape memory alloys hys It is closely related to the eigenvalue (λ2) and the volume change rate (ΔV / V). λ2 and ΔV / V can be calculated by lattice constants. The present invention calculates the lattice parameters by Vienna Ab-Initio Simulation Package (VASP) based on density functional theory (DFT). Since the thermal hysteresis data set of Ni-Mn-based shape memory alloy is large, it is very time-consuming to obtain the lattice constants of all alloys in the composition space through traditional DFT calculations. Therefore, this paper proposes a method for obtaining the lattice constants of all data by combining first-principles calculations with machine learning (ML). 77 component alloys of different components are selected, and their lattice constants a0, a, b, c and angle β are obtained by first-principles calculations, such as Figure 1 As shown. The λ2 and ΔV / V of these 77 alloy components were calculated by formula. Machine learning models of λ2 and ΔV / V were established respectively, with elements and physical properties as features, λ2 and ΔV / V as labels, and the relationship between features and labels was established. The Random Forest Regressor (RFR) model performed well, as shown in Figure 2 Figure 3 As shown, it can be used to predict λ2 and ΔV / V of the entire Ni-Mn-based thermal hysteresis data set, and added to the Ni-Mn-based shape memory alloy thermal hysteresis data set as a key physical coupling feature;

[0040] 3. Perform feature engineering, model building and model selection through machine learning to obtain the optimal prediction model; Due to differences in measurement methods, some data do not have both M s 、M f , A s , A f, some data are wrong, and some thermal hysteresis is negative, so data cleaning is performed, and finally 205 sets of data are retained. After data cleaning, feature engineering and model parameter optimization, the present invention evaluates the performance of three machine learning models: Random Forest Regressor (RFR), XGBRegressor (XGBR) and Gradient BoostedRegressor Trees (GBRT), such as Figure 4 As shown. Among them, the score R of the RFR model on the test set 2 Highest (R 2 =0.902), the fitting effect is the best, and it can be used to predict the thermal hysteresis of the new Ni-Mn based shape memory alloy;

[0041] Fourth, the optimal prediction model is used to predict the thermal hysteresis of Ni-Mn based shape memory alloy;

[0042] The RFR model is used to predict the thermal hysteresis of new Ni-Mn based shape memory alloys, and to find a ternary alloy Ni with narrow thermal hysteresis from a wide range of Ni-Mn based shape memory alloy search space. 45 Mn 45 Sn 10 , through Al, Si, B and Ga to the ternary alloy Ni 45 Mn 45 Sn 10 Doping further reduces the ΔT of the alloy hys ,like Figure 5 shown.

[0043] Experimental preparation and performance characterization of a novel narrow hysteresis Ni-Mn based shape memory alloy

[0044] Based on the prediction results of machine learning, a Ni 45 Mn 45 Sn8X2 (X = Sn, Al, Si, B and Ga) and Ni 45 Mn 45 Sn 10-x Al x All samples were prepared by melting Ni, Mn, Sn, Al, Si, Ga, and B with a purity of 99.99wt.% in a non-consumable vacuum arc melting furnace under an argon atmosphere. The vacuum was first evacuated to 4×10 -3Pa, and then backfilled with high-purity argon to 0.07MPa for gas washing, and smelting was carried out after four cycles. All sample ingots were magnetically stirred during the smelting process, and repeatedly turned over and smelted 5 times to ensure the accuracy and uniformity of the chemical composition of the alloy. After cooling, they were taken out and cut along the diameter by wire cutting. All alloy samples were first mechanically polished to remove surface impurities, then cleaned with acetone and alcohol and packaged in a vacuum of 10 -3 Pa quartz tube, homogenized at 1173 K for 1 hour, and quenched in ice water to obtain a high degree of chemical order.

[0045] The martensitic transformation behavior of the alloy samples was tested using a Diamond differential scanning calorimeter produced by PerkinElmer. All test samples were of uniform size of approximately The wafer was placed in an aluminum pot after removing the oxide layer on the surface for sample preparation. The temperature change rate during the test was 10K / min. The martensitic transformation start temperature (M) of the corresponding sample was obtained from the obtained DSC curve using the tangent method. s ), the final temperature of the martensitic transformation temperature (M f ), the starting temperature of the reverse martensitic transformation (A s ), and the final temperature of the reverse martensitic transformation (A f ), through ΔT hys =(A s +A f -M s -M f ) / 2 to calculate the ΔT of the sample hys .

[0046] from Figure 6 It can be seen that Al, Si, B and Ga have a significant influence on Ni 45 Mn 45 Sn 10 The doping of the ternary alloy further reduces the thermal hysteresis, and the error between the two is small. The machine learning prediction results are consistent with the experimentally measured ΔT hys The results match, such as Figure 7 As shown. Among them, the quaternary alloy Ni after Al doping 45 Mn 45 Sn8Al2 has the smallest thermal hysteresis (ΔT hys =7.99K).

Claims

1. A method for predicting thermal hysteresis ∆Thys of Ni-Mn-based shape memory alloy based on machine learning, characterized in that The method for predicting thermal hysteresis ∆Thys of Ni-Mn-based shape memory alloy based on machine learning is specifically carried out in the following steps:

1. Collecting ∆T from the literature hys Data information, build data set; ∆T hys The data information includes the phase transition temperature data of the alloy and the properties of the alloy itself; the properties of the alloy itself include the number of protons, the number of valence electrons, atomic radius, melting point, minimum oxidation state, maximum oxidation state, resistivity, molar volume, thermal conductivity, boiling point, linear thermal expansion coefficient, electronegativity, ionization energy, relative atomic mass, and metal radius; the properties of all elements in the alloy are multiplied by their proportion in the entire alloy through the formula, and the result is used as the physical properties of the final alloy. These properties are ∆T hys Provide a physical basis for modeling the structure-property relationship; Second, the eigenvalue λ2 and volume change rate ∆V / V are obtained by first-principles calculation combined with the ML model; 3. 77 alloys of different compositions were selected, and their lattice constants a0, a, b, c and angle β were calculated by first principles. λ2 and ∆V / V of the above 77 alloys were calculated by formula. Machine learning models of λ2 and ∆V / V were established respectively, with elements and physical properties as features, λ2 and ∆V / V as labels, the relationship between features and labels was established, and Random Forest Regressor was selected as the optimal prediction model.

4. Use the optimal prediction model to predict the thermal hysteresis of Ni-Mn based shape memory alloy.

2. The method for predicting thermal hysteresis ∆Thys of Ni-Mn-based shape memory alloy based on machine learning according to claim 1, characterized in that Step 1: Collect ∆T from the literature hys The data information is 462 sets of data collected from the published literature data of Ni-Mn-based SMAs, including the basic data Ni, Mn and X elements, and ten doping elements Ga, Ti, Fe, Co, Pd, Cu, Al, Si, V and B; the X element is In, Sn or Sb.

3. The method for predicting thermal hysteresis ∆Thys of Ni-Mn-based shape memory alloy based on machine learning according to claim 1, characterized in that ∆T described in step 1 hys =( A s + A f − M s − M f ) / 2; where M s is the martensitic transformation starting temperature, M f is the martensitic transformation end temperature, A s is the starting temperature of reverse martensitic transformation, A f is the end temperature of the reverse martensitic transformation.

4. The method for predicting thermal hysteresis ∆Thys of Ni-Mn-based shape memory alloy based on machine learning according to claim 1, characterized in that The eigenvalue λ2 and volume change rate ∆V / V described in step 2 are obtained by calculating the lattice constant.

5. The method for predicting thermal hysteresis ∆Thys of Ni-Mn-based shape memory alloy based on machine learning according to claim 4, characterized in that The lattice parameters were calculated by the Vienna Ab-Initio Simulation Package based on density functional theory.

6. The method for predicting thermal hysteresis ∆Thys of Ni-Mn-based shape memory alloy based on machine learning according to claim 1, characterized in that After step 2 is completed, data cleaning is performed and 205 sets of data are retained.

7. The method for predicting thermal hysteresis ∆Thys of Ni-Mn-based shape memory alloy based on machine learning according to claim 1, characterized in that The narrow thermal hysteresis of the ternary alloy Ni is obtained by the prediction of step 4. 45 Mn 45 Sn 10 and the ternary alloy Ni 45 Mn 45 Sn 10 The specific preparation process is as follows: Ni, Mn, Sn, Al, Si, Ga, and B with a purity of 99.99wt.% are used as raw materials, and a non-consumable vacuum arc melting furnace is used to melt and prepare under the protective atmosphere of argon. First, a mechanical pump and a molecular pump are used to evacuate to 4×10 -3 Pa, and then backfilled with high-purity argon to 0.07MPa for gas washing, and smelting was carried out after four cycles; all sample ingots were magnetically stirred during the smelting process, and repeatedly turned over and smelted 5 times to ensure the accuracy and uniformity of the chemical composition of the alloy, and then taken out after cooling, and cut along the diameter by wire cutting; all alloy samples were first mechanically polished to remove surface impurities, and then cleaned with acetone and alcohol and packaged in a vacuum of 10 -3 Pa quartz tube, kept at 1173K for 1 hour for homogenization annealing, and quenched in ice water.

Images