Next Article in Journal
Possibilities of K-Value Determination for Active Admixtures with Respect to Durability
Previous Article in Journal
Low-Refractive-Index SiO2 Nanocolumnar Thin Films Fabricated by Oblique Angle Deposition
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Machine Learning-Assisted Discovery of Empirical Rule for Martensite Transition Temperature of Shape Memory Alloys

Key Laboratory for Anisotropy and Texture of Materials (Ministry of Education), School of Material Science and Engineering, Northeastern University, Shenyang 110819, China
*
Author to whom correspondence should be addressed.
Materials 2025, 18(10), 2226; https://doi.org/10.3390/ma18102226
Submission received: 9 April 2025 / Revised: 28 April 2025 / Accepted: 10 May 2025 / Published: 12 May 2025
(This article belongs to the Special Issue Magnetic Shape Memory Alloys: Fundamentals and Applications)

Abstract

:
Shape memory alloys (SMAs) derive their unique functional properties from martensitic transformations, with the martensitic transformation temperature (TM) serving as a key design parameter. However, existing empirical rules, such as the valence electron concentration (VEC) and lattice volume (V) criteria, are typically restricted to specific alloy families and lack general applicability. In this work, we used a data-driven methodology to find a generalizable empirical formula for TM in SMAs by combining high-throughput first-principles calculations, feature engineering, and symbol regression techniques. Key factors influencing TM were first identified and a predictive machine learning model was subsequently trained based on these features. Furthermore, an empirical formula of TM = 82 ( ρ ¯ · M P ¯ ) 700 was derived, where ρ ¯ and M P ¯ represent the weight-average value of density and melting point, respectively. The empirical formula exhibits strong generalizability across a wide range of SMAs, such as NiMn-based, NiTi-based, TiPt-based, and AuCd-based SMAs, etc., offering practical guidance for the compositional design and optimization of shape memory alloys.

1. Introduction

Shape memory alloys (SMAs) are a class of functional materials renowned for their unique phase transformation behaviors, including the shape memory effect [1], superelasticity [2], and elastocaloric effect [3]. These functionalities originate from martensitic transformation. Among various properties, the martensitic transformation temperature (TM) is a critical parameter in SMA design, as it directly determines the operational temperature window of the material. For example, Ni2MnGa, one of the most extensively studied shape memory alloys, exhibits a TM around 220 K, thereby restricting its application in ambient or high-temperature environments such as actuators or sensors. In contrast, cryogenic applications like solid-state cooling require the TM to be substantially below room temperature. Therefore, precise control and reliable prediction of the TM are essential for tailoring the functional properties of SMAs for diverse applications.
Several strategies have been developed to tune TM, including modifications of chemical composition, microstructure, atomic ordering, and vacancy concentration [4,5,6,7], with compositional tuning emerging as the most effective approach. Empirical rules, such as the valence electron concentration (VEC) rule [3,8,9,10] where the TM generally increases with higher VEC, and the lattice volume rule [11] where a reduced austenite lattice volume tends to raise the TM, have been widely employed. Based on these principles, empirical models like TM = 705(VEC) − 5067 for Ni-Mn-Ga alloys have been proposed [12]. Other composition-related empirical formulas include TM = 1000 × (Ni% − 50) for Ni-Ti with 50~60 at.% [13,14], and for Ti-Nb where each 1 at.% increase in Nb content decreases the TM by approximately 43 K [15,16]. However, these models often suffer from limited generalizability across different alloy systems due to their strong dependence on specific material families. Therefore, identifying universal key factors that govern TM across various shape memory alloys (SMAs) remains a critical challenge. Recently, machine learning (ML) has emerged as a powerful technique to model complex, nonlinear relationships and predict material properties. Although previous studies have applied ML to forecast TM in SMAs, they often rely on computationally expensive features derived from density functional theory (DFT) calculations or experimental inputs [17], or are restricted to specific alloy families [18,19]. Therefore, developing fast and generalizable models to predict TM based on readily available information holds significant theoretical and practical value.
To address this challenge, we introduced a big data-driven machine learning framework based on elemental and simple substance properties—a state-of-the-art approach for resolving multifactorial materials design problems [18,20,21,22]. In this work, high-throughput ab initio calculations were employed alongside extensive data collection from the literature to expand the training dataset. Subsequently, a five-step descriptor screening process was implemented to identify the key factors governing TM. Based on these critical features, a highly accurate random forest machine learning model was developed to predict TM. Furthermore, a simple yet robust empirical relationship between TM and the identified parameters was discovered, exhibiting exceptional generalization performance across a wide range of SMAs.

2. Methods

A random forest (RF) machine learning model, as implemented in the Scikit-Learn software package (version 1.3.2) [23] was employed in this study. The number of decision trees was set to 50. During model construction, the bootstrap sampling strategy [24] was adopted, and the maximum tree depth was left unconstrained to allow for full data-driven learning. To mitigate the risk of overfitting, 10-fold cross-validation was performed [25]. Model performance was evaluated using the mean absolute error (MAE), root mean square error (RMSE) and mean square error (MSE). To identify the dominant factors influencing TM, a comprehensive feature selection strategy was employed, involving the variance screening, the Pearson correlation screening [26], the univariate screening [27,28], the recursive elimination screening [29], and the exhaustive screening [21]. Detailed descriptions of these methods can be found in Ref. [22].
High-throughput ab initio calculations were conducted using spin-polarized density functional theory within the framework of the projector augmented-wave (PAW) method [30], as implemented in the Vienna Ab initio Simulation Package (VASP, version 6.4.2) [31]. The exchange-correlation energy was treated using the Perdew-Burke-Enzerh (PBE) functional within the generalized gradient approximation (GGA) [32,33]. The plane-wave kinetic energy cutoff, total energy convergence criterion, and force convergence criterion were set to 600 eV, 10−5 eV, and 10−4 eV/Å, respectively. A 16-atom supercell was adopted for all calculations. The first Brillouin zone was sampled using a Monkhorst-Pack [34] k-point grid of 13 × 13 × 13.

3. Results and Discussion

3.1. Dataset

Based on the literature [17,18,35], a dataset comprising 82 experimental data points of the TM in various shape memory alloys was collected. These systems include NiMn-based, NiTi-based, AgCd-based, AuCd-based, PdTi-based, PtTi-based, CoNi-based, TiNb-based, TiTa-based, ZnAuCu-based, TiAu-based, and MgSc-based alloys, effectively covering nearly all known shape memory alloy families to date (details in Appendix A Table A1).
To further expand the dataset, high-throughput ab initio calculations were performed on 109 kinds of L21 structured Ni2MnGa1−xZx (x = 0.25, 0.5, 0.75 and 1) alloys, where Z represents various common elements from the periodic table, as illustrated in Figure 1. The Bain distortion model was employed to compute the energy difference (ΔE) between the martensitic and austenitic phases (see calculation details in Ref. [22]). The TM was then estimated using the relation T M = E n k B [36,37] where n is the number of atoms in the supercell, and kB is the Boltzmann constant which connects the macroscopic temperature and microscopic energy parameters. Figure 1 displays the calculated martensitic transformation temperatures (TM) for Ni2MnGa1−xZx alloys, where Z spans 35 different substituting elements and x takes values of 0.25, 0.5, 0.75, and 1.00. For each element, corresponding TM values are shown beneath the elemental symbol. A backslash symbol (“\”) indicates compositions where the austenitic phase is energetically more stable than the martensitic phase even at 0 K, corresponding to a negative ΔE. In these cases, martensitic transformation does not occur, and the alloy remains in a stable austenitic state.
To assess the applicability of known empirical rules across this heterogeneous dataset, the relationships between TM and two classic descriptors were examined: Figure 2a,b shows the correlations of the TM with the valence electron concentration (VEC) and lattice volume (V), respectively. Although prior studies have demonstrated that an increase in VEC or a decrease in lattice volume correlates tends to elevate the TM within specific alloy families [3,8,9,10,11,12] these trends do not consistently extend across a broader and more diverse dataset. In our analysis, neither VEC nor V exhibited a strong or reliable correlation with TM when considering the full dataset, thereby underscoring the limited generalizability of these traditional empirical rules. In addition, the transformation temperatures span a wide range— from approximately 100 K to 1300 K—reflecting the intrinsic diversity of thermal responses among different SMA compositions. This rich and representative dataset thus provides a robust foundation for the development of data-driven predictive models and enables cross-system generalization.

3.2. Feature Engineering

To represent the different alloys in the training set, a total of 64 constituent element-associated parameters were initially considered, given that the structural phase transition as well as the physical and chemical properties of compounds are predominantly determined by their constituent elements [38,39,40,41]. Specifically, the weight-average value and standard deviation of 31 types of element or elemental-substance-associated parameters were calculated for each alloy [39,40,42], as summarized in Table 1.
In addition to these elemental descriptors, two phase stability-related parameters, i.e., mixing entropy S m i x and atom radium difference δ [20], were included. Considering that shape memory alloys generally possess mixed metallic, covalent, and ionic chemical bonds [43,44], three different atom radius definitions, including atomic radius, covalent radius, and ionic radius, were employed to estimate δ . In total, 64 descriptors were utilized to featurize the alloys within the training dataset.

3.3. Screening of Key Features

To eliminate the redundant or insignificant descriptors and to identify the key materials parameters governing TM, a successive five-step screening procedure was employed, as illustrated in Figure 3. This process combined a series of complementary descriptor selection methods applied sequentially: variance screening (Step 1) [45], Pearson correlation screening (Step 2) [26], univariate feature selection (Step 3) [27,28], recursive feature elimination (Step 4) [29], and exhaustive screening (Step 5) (more details in Appendix B and Ref. [21]). The detailed description of these selection methods can be found in Ref. [22].
First, redundancy among descriptors was eliminated using the variance method (Step 1) [45], by removing descriptors that exhibited identical values across all data points in the training set. Second, strongly correlated descriptors were identified through Pearson correlation screening (Step 2) [26]. The Pearson correlation coefficient (r) between every pair of descriptors was computed, and for each pair with |r| > 0.75, one descriptor was removed to reduce redundancy. For instance, the atomic number (Z) and atomic weight (AW) showed a perfect correlation (r = 1); thus, one of the two, such as the atomic number in this case, was eliminated. Following this screening process, 27 descriptors were retained for further analysis.
Figure 4a–c presents the results of feature selection using univariate filtering, recursive feature elimination (RFE), and exhaustive search, respectively. Following these three successive screening steps, the number of features was reduced from the initial 66 to 16, and finally to 6. This systematic feature selection process effectively narrowed the original set of candidate descriptors down to six key features, which are summarized in Table 2.
The final six descriptors identified as critical for determining the martensitic transformation temperature (TM) are: the weight-average value of ionization energy ( I E ¯ ), weight-average value of heat capacity ( H C ¯ ), weight-average value of shear modulus ( G ¯ ), and standard deviation of atomic radius (σAR). These descriptors were subsequently used to train the machine learning models, as they represent the dominant factors influencing TM.
Using the six final selected features summarized in Table 2, a random forest (RF) regression model was retrained to predict the martensitic transformation temperature (TM) of the alloys. The model’s performance was evaluated through a ten-fold cross-validation procedure to ensure its generalization ability and predictive accuracy. The resulting coefficient of determination (R2) reached 0.82, demonstrating that the model exhibits excellent predictive performance.
Figure 5 presents the feature importance ranking obtained from the random forest model. By assessing the relative contribution of each feature to the model’s prediction capability, the following order of importance was determined: M P ¯ (39) > ρ ¯ (37) > G ¯ (10) > σAR (6) > I E ¯ (5) > H C ¯ (3). These results highlight the dominant roles of melting point and density in governing the TM of shape memory alloys.

3.4. Linear Relationship Between ρ ¯ · M P ¯ and TM

To further investigate the influence of the six key features identified by machine learning on the martensitic transformation temperature (TM) of shape memory alloys (SMAs), detailed visual analyses were conducted.
Figure 6a illustrates the relationship between the weight-average value of density ( ρ ¯ ), the weight-average value of melting point ( M P ¯ ), and TM. As shown, the feature space defined by ρ ¯ and M P ¯ can be effectively divided into two distinct regions corresponding to different TM ranges. The green region in the lower left corner is characterized by low values of both M P ¯ and ρ ¯ ; this region corresponds to relatively low martensitic transformation temperatures (TM < 300 K). The red region in the upper right corner is marked by higher values of both M P ¯ and ρ ¯ ; this region corresponds to higher transformation temperatures (TM > 300 K). This separation suggests that M P ¯ and ρ ¯ act as discriminative features that jointly determine TM in SMAs. Figure 6b provides an enlarged view of the black-boxed region from Figure 6a, further confirming the observed trend: higher values of the weight-average of melting point ( M P ¯ ) and weight-average value of density ( ρ ¯ ) correspond to higher martensitic transformation temperatures (TM). This observation reinforces the conclusion that increases in M P ¯ and ρ ¯ are closely associated with elevated TM values. Figure 6c illustrates the relationships between the weight-average value of shear modulus ( G ¯ ), the weight-average value of ionization energy ( I E ¯ ), and TM. The results reveal weak correlations between these features and TM, suggesting that G ¯ and I E ¯ have a minimal influence on the martensitic transformation temperature. Similarly, Figure 6d depicts the relationships between the standard deviation of atomic radius (σAR), the weight-average value of heat capacity ( H C ¯ ), and TM. Again, no clear trends are observed, indicating that these descriptors also exert relatively minor influence on TM. In sum, the weight-average value of melting point ( M P ¯ ) and weight-average value of density ( ρ ¯ ) are identified as the dominant factors affecting the martensitic transformation temperature. Consequently, these two features were selected for the construction and analysis of an empirical predictive formula.
Using M P ¯ and ρ ¯ as the two key features, a symbolic regression approach was employed to develop an empirical formula for predicting the martensitic transformation temperature (TM) of SMAs. Symbolic regression [46] is a data-driven technique for discovering functional relationships, aiming to autonomously derive optimal analytical expressions within a predefined mathematical operator space. In this study, symbolic regression was implemented using a genetic algorithm [46]. The mathematical operator space consisted of basic arithmetic operations, including addition, multiplication, subtraction, and division. The input features were M P ¯ and ρ ¯ , and the target variable was TM. The fitness of each candidate formula was evaluated based on the correlation coefficient (r) between the predicted and actual TM values.
After 500 generations of iterative optimization, the final empirical formula was identified:
T M = 82 ( ρ ¯ · M P ¯ ) 700
where ρ ¯ and M P ¯ represent weight-average values of density and melting point, respectively, and TM is the martensite transition temperature of SMAs. This result highlights that TM scales approximately linearly with the combined descriptor ρ ¯ · M P ¯ .
Figure 7 illustrates the relationship between ρ ¯ · M P ¯ and TM across the entire dataset. A strong linear correlation is observed (r = 0.81 and MAE = 120 K), confirming that the empirical relationship generated via symbolic regression accurately captures the dependence of TM on the key features. Additionally, the distribution of the absolute error δ(TM) is shown in the inset of Figure 7. Notably, for the majority of datapoints (58%), δ(TM) is less than 100 K. To further evaluate the model’s performance, both the 95% confidence interval and the 95% prediction interval were constructed in Figure 7. Most data points fall within the prediction intervals, are closely aligned with the confidence intervals, and are distributed symmetrically around the regression line. These results indicate that the model provides an accurate and reliable estimation of the overall trend in TM across different compositions.
This simple formula offers an effective method to predict TM across a diverse range of alloy families, including NiTi-based, TiPd-based, NiMn-based, TiAu-based, MgSc-based, AgCd-based, AuCd-based, PdTi-based, PtTi-based, CoNi-based, TiNb-based, TiTa-based, and ZnAuCu-based SMAs. This stands in contrast to traditional empirical rules such as the valence electron concentration (VEC) and lattice volume (V) rules, or other composition-related rules, which are typically applicable only to specific alloy families (see Figure 2 and Refs. [3,8,9,10,11,12]). While previous studies have used ML to forecast TM in SMAs, many faced similar limitations to traditional empirical approaches: (1) some models are restricted to specific alloy systems [18,19], and (2) models capable of predicting across multiple systems often rely on computationally expensive features derived from density functional theory (DFT) calculations and require complex frameworks such as artificial neural networks [17]. These DFT-based models can take hours or even days to process a single data point, and their computational cost increases rapidly with prediction set size. In contrast, the empirical formula proposed in this work enables near real-time predictions of TM for hundreds or even thousands of alloy compositions, as it relies solely on simple substance properties that are readily available. The key advantage of our formula lies in its ability to quickly and efficiently predict the martensitic transformation temperature across multiple SMA systems.

4. Conclusions

In this study, a systematic data-driven approach was developed to derive an interpretable and generalizable empirical formula for predicting the martensitic transformation temperature (TM) of shape memory alloys (SMAs). By integrating high-throughput first-principles calculations, extensive feature selection techniques, and symbolic regression methods, we identified the key physical factors that govern TM behavior. The final empirical model, expressed as TM = 82 ( ρ ¯ · M P ¯ ) 700 , highlights the combined influence of material density and melting point on martensitic transformation characteristics. The proposed formula demonstrates strong predictive capability across diverse classes of SMAs, such as NiMn-based, NiTi-based, TiPt-based, and AuCd-based SMAs, etc. The reliability of the model was quantitatively assessed by evaluating performance metrics and the associated confidence/prediction intervals were reported to ensure accuracy of the model. Compared to previous empirical rules that were often limited to narrow alloy families, the present work offers a unified and practically applicable tool for SMA composition design, offering significant potential for high-throughput material design and optimization.

Author Contributions

Conceptualization, H.-X.L. and H.-L.Y.; methodology, H.-X.L. and N.J.; software, H.-X.L. and B.Y.; validation, B.Y., Z.L. and X.Z.; formal analysis, L.Z.; investigation, H.-X.L.; writing—original draft preparation, H.-X.L.; writing—review and editing, H.-L.Y. and X.Z.; visualization, H.-X.L.; supervision, L.Z.; funding acquisition, H.-L.Y. All authors have read and agreed to the published version of the manuscript.

Funding

This work is supported by the National Natural Science Foundation of China (Grant No. 52471003).

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The original contributions presented in this study are included in the article. Further inquiries can be directed to the corresponding author.

Conflicts of Interest

The authors declare no conflicts of interest.

Appendix A

Table A1. Chemical compositions and TM values of alloys from experimental (Exp.) in the literature.
Table A1. Chemical compositions and TM values of alloys from experimental (Exp.) in the literature.
CompositionsTM (K)Ref.CompositionsTM (K)Ref.
Ti500Ni452Cu10Fe38248[14]Ti500Ni468Cu9Fe20Pd3290[14]
Ti500Ni444Cu20Fe36239[14]Ti500Ni442Cu19Fe38Pd1243[14]
Ti500Ni428Cu40Fe32231[14]Ti500Ni467Cu8Fe23Pd2282[14]
Ti500Ni436Cu30Fe34242[14]Ti500Ni442Cu19Fe39242[14]
Ti500Ni460Cu0Fe40234[14]Ti500Ni481Cu2Fe15Pd2302[14]
Ti500Ni445Cu15Fe30Pd10223[14]Ti500Ni445Cu16Fe37Pd2244[14]
Ti500Ni340Cu130Pd30299[14]Ti500Ni482Cu6Fe9Pd3320[14]
Ti500Ni340Cu100Pd60303[14]Ti500Ni465Cu11Fe22Pd2284[14]
Ti500Ni340Cu120Pd40311[14]Ti500Ni483Fe16Pd1302[14]
Ti500Ni420Cu50Fe30226[14]Ti500Ni490Fe2Pd8358[14]
Ti500Ni350Cu120Pd30321[14]Ti500Ni486Fe9Pd5332[14]
Ti500Ni400Pd100279[14]Ti500Ni435Cu20Fe45226[14]
Ti500Ni340Cu140Pd20319[14]Ti500Ni250Pd250456[14]
Ti500Ni340Cu160333[14]Ag51Cd4948[13]
Ti500Ni440Cu10Fe10Pd40239[14]Au505Cd495308[13]
Ti500Ni340Pd160331[14]Au50Ti50843[13]
Ti500Ni364Cu120Fe16286[14]Pd50Ti50761[13]
Ti500Ni412Cu60Fe28220[14]Pt50Ti501291[13]
Ti500Ni380Cu100Fe20274[14]Ni50Ti50252[13]
Ti500Ni348Cu140Fe12271[14]Co50Ni24Ga26336[13]
Ti500Ni396Cu80Fe24258[14]Ti505Ni245Pd250454[13]
Ti500Ni500364[14]Ti505Ni245Pt250721[13]
Ti500Ni440Cu20Fe40220[14]Ti50Pt25Ir251395[13]
Ti500Ni445Cu21Fe34248[14]Ni50Mn25Ga25278[13]
Ti500Ni440Cu23Fe36Pd1241[14]Ti75Nb22476[13]
Ti500Ni404Cu46Fe10Pd40270[14]Ti75Nb25381[13]
Ti500Ni457Fe43231[14]Ti75Ta25558[13]
Ti500Ni458Fe42234[14]Mg81Sc1983[13]
Ti500Ni428Cu36Fe28Pd8234[14]Ag51Cd49223[30]
Ti500Ni439Cu21Fe40233[14]Au50Cd50308[30]
Ti500Ni445Cu17Fe37Pd1245[14]Ti50Au50843[30]
Ti500Ni457Cu12Fe30Pd1264[14]Zn45Au30Cu25235[30]
Ti500Ni440Cu23Fe37234[14]Co50Ni24Ga26336[30]
Ti500Ni451Cu14Fe35248[14]Ti76Nb24338[30]
Ti500Ni446Cu19Fe34Pd1250[14]Ni50Mn25Ga25278[30]
Ti500Ni438Cu26Fe36239[14]Ti50Ni50252[30]
Ti500Ni439Cu15Fe46228[14]Ti50Pd50761[30]
Ti500Ni445Cu19Fe34Pd2243[14]Ti50Pt501291[30]
Ti500Ni460Cu11Fe28Pd1261[14]Ti50.5Ni24.5Pd25454[30]
Ti500Ni438Cu20Fe41Pd1231[14]Ti50.5Ni24.5Pt25721[30]
Ti500Ni439Cu20Fe40Pd1232[14]Ti50Pt25Ir251395[30]

Appendix B

The exhaustive method was used to identify the best combination of descriptors. Every possible combination of the remaining descriptors after the recursive elimination was used to train the ML model. The combination of descriptors in which the ML model possesses the lowest error was selected as the critical factors that determine the TM of SMA. The detailed flow is as follows (Figure A1):
(i)
Generate All Combinations of Remaining Descriptors: After eliminating irrelevant features, every possible combination of the remaining descriptors is generated for testing.
(ii)
Train ML Model: Each combination of descriptors is used to train a ML model, ensuring a diverse exploration of feature subsets.
(iii)
Evaluate Model Performance: The performance of each model is evaluated using standard error metrics like R2, RMSE, and MSE to assess its accuracy.
(iv)
Select Optimal Combination: The combination of descriptors that results in the lowest error is chosen as the most significant contributors to the TM prediction.
Figure A1. Workflow of exhaustive screening.
Figure A1. Workflow of exhaustive screening.
Materials 18 02226 g0a1

References

  1. Adachi, Y.; Kato, M.; Kanomata, T.; Kikuchi, D.; Xu, X.; Umetsu, R.Y.; Kainuma, R.; Ziebeck, K.R.A. The effect of pressure on the magnetic and structural transition temperatures of the shape memory alloys Ni50+x(Mn0.5Fe0.5)25Ga25−x. Phys. B Conden. Matt. 2015, 464, 83–87. [Google Scholar] [CrossRef]
  2. Chen, H.Y.; Wang, Y.D.; Nie, Z.H.; Li, R.G.; Cong, D.Y.; Liu, W.J.; Ye, F.; Liu, Y.Z.; Cao, P.Y.; Tian, F.Y.; et al. Unprecedented non-hysteretic superelasticity of 001-oriented NiCoFeGa single crystals. Nat. Mater. 2020, 19, 712–718. [Google Scholar] [CrossRef]
  3. Qu, Y.H.; Cong, D.Y.; Li, S.H.; Gui, W.Y.; Nie, Z.H.; Zhang, M.H.; Ren, Y.; Wang, Y.D. Simultaneously achieved large reversible elastocaloric and magnetocaloric effects and their coupling in a magnetic shape memory alloy. Acta Mater. 2018, 151, 41–55. [Google Scholar] [CrossRef]
  4. Kumar Pathak, A.; Dubenko, I.; Stadler, S.; Ali, N. The effect of partial substitution of In by Si on the phase transitions and respective magnetic entropy changes of Ni50Mn35In15 Heusler alloy. J. Phys. D Appl. Phys. 2008, 41, 202004. [Google Scholar] [CrossRef]
  5. Ghosh, S.; Ghosh, S. Role of composition, site ordering, and magnetic structure for the structural stability of off-stoichiometric Ni2MnSb alloys with excess Ni and Mn. Phys. Rev. B 2019, 99, 064112. [Google Scholar] [CrossRef]
  6. Huang, X.-M.; Wang, L.-D.; Liu, H.-X.; Yan, H.-L.; Jia, N.; Yang, B.; Li, Z.-B.; Zhang, Y.-D.; Esling, C.; Zhao, X.; et al. Correlation between microstructure and martensitic transformation, mechanical properties and elastocaloric effect in Ni–Mn-based alloys. Intermetallics 2019, 113, 106579. [Google Scholar] [CrossRef]
  7. Unzueta, I.; de R-Lorente, D.A.; Cesari, E.; Sánchez-Alarcos, V.; Recarte, V.; Pérez-Landazábal, J.I.; García, J.A.; Plazaola, F. Experimental observation of vacancy-assisted martensitic transformation shift in Ni-Fe-Ga Alloys. Phys. Rev. Lett. 2019, 122, 165701. [Google Scholar] [CrossRef] [PubMed]
  8. Sutou, Y.; Imano, Y.; Koeda, N.; Omori, T.; Kainuma, R.; Ishida, K.; Oikawa, K. Magnetic and martensitic transformations of NiMnX(X = In,Sn,Sb) ferromagnetic shape memory alloys. Appl. Phys. Lett. 2004, 85, 4358–4360. [Google Scholar] [CrossRef]
  9. Zayak, A.T.; Adeagbo, W.A.; Entel, P.; Rabe, K.M. e/a dependence of the lattice instability of cubic Heusler alloys from first principles. Appl. Phys. Lett. 2006, 88, 111903. [Google Scholar] [CrossRef]
  10. Tan, C.; Tai, Z.; Zhang, K.; Tian, X.; Cai, W. Simultaneous enhancement of magnetic and mechanical properties in Ni-Mn-Sn alloy by Fe doping. Sci. Rep. 2017, 7, 43387. [Google Scholar] [CrossRef]
  11. Jiang, C.; Feng, G.; Gong, S.; Xu, H. Effect of Ni excess on phase transformation temperatures of NiMnGa alloys. Mater. Sci. Eng. A 2003, 342, 231–235. [Google Scholar] [CrossRef]
  12. Chernenko, V.A.; Cesari, E.; Kokorin, V.V.; Vitenko, I.N. The development of new ferromagnetic shape memory alloys in Ni-Mn-Ga system. Scr. Metall. Mater. 1995, 33, 1239–1244. [Google Scholar] [CrossRef]
  13. Li, W.; Zhao, C. Microstructure and Phase Transformation Analysis of Ni50−xTi50Lax Shape Memory Alloys. Crystals 2018, 8, 345. [Google Scholar] [CrossRef]
  14. Yang, X.; Ma, L.; Shang, J. Martensitic transformation of Ti50(Ni50−xCux) and Ni50(Ti50−xZrx) shape-memory alloys. Sci. Rep. 2019, 9, 3221. [Google Scholar] [CrossRef]
  15. Kim, H.Y.; Hashimoto, S.; Kim, J.I.; Hosoda, H.; Miyazaki, S. Mechanical Properties and Shape Memory Behavior of Ti-Nb Alloys. Mater. Trans. 2004, 45, 2443–2448. [Google Scholar] [CrossRef]
  16. Fang, H.; Xu, X.; Zhang, H.; Sun, Q.; Sun, J. Alloying Effect on Transformation Strain and Martensitic Transformation Temperature of Ti-Based Alloys from Ab Initio Calculations. Materials 2023, 16, 6069. [Google Scholar] [CrossRef]
  17. Minami, D.; Uesugi, T.; Takigawa, Y.; Higashi, K. Artificial neural network assisted by first-principles calculations for predicting transformation temperatures in shape memory alloys. Int. J. Mod. Phys. B 2019, 33, 1950055. [Google Scholar] [CrossRef]
  18. Xue, D.Z.; Xue, D.Q.; Yuan, R.H.; Zhou, Y.M.; Balachandran, P.V.; Ding, X.D.; Sun, J.; Lookman, T. An informatics approach to transformation temperatures of NiTi-based shape memory alloys. Acta Mater. 2017, 125, 532–541. [Google Scholar] [CrossRef]
  19. Trehern, W.; Ortiz-Ayala, R.; Atli, K.C.; Arroyave, R.; Karaman, I. Data-driven shape memory alloy discovery using Artificial Intelligence Materials Selection (AIMS) framework. Acta Mater. 2022, 228, 117751. [Google Scholar] [CrossRef]
  20. Pei, Z.; Yin, J.; Hawk, J.A.; Alman, D.E.; Gao, M.C. Machine-learning informed prediction of high-entropy solid solution formation: Beyond the Hume-Rothery rules. NPJ Comput. Mater. 2020, 6, 50. [Google Scholar] [CrossRef]
  21. Zhang, H.; Fu, H.; He, X.; Wang, C.; Jiang, L.; Chen, L.-Q.; Xie, J. Dramatically enhanced combination of ultimate tensile strength and electric conductivity of alloys via machine learning screening. Acta Mater. 2020, 200, 803–810. [Google Scholar] [CrossRef]
  22. Liu, H.-X.; Yan, H.-L.; Jia, N.; Tang, S.; Cong, D.; Yang, B.; Li, Z.; Zhang, Y.; Esling, C.; Zhao, X.; et al. Machine learning informed tetragonal ratio c/a of martensite. Comp. Mater. Sci. 2023, 233, 112735. [Google Scholar] [CrossRef]
  23. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
  24. Singh, K.; Xie, M. Bootstrap method. In International Encyclopedia of Education; Peterson, P., Baker, E., McGaw, B., Eds.; Elsevier: Oxford, UK, 2010; pp. 46–51. [Google Scholar]
  25. Ward, L.; Agrawal, A.; Choudhary, A.; Wolverton, C. A general-purpose machine learning framework for predicting properties of inorganic materials. NPJ Comput. Mater. 2016, 2, 16028. [Google Scholar] [CrossRef]
  26. Pearson, K.; Galton, F. VII. Note on regression and inheritance in the case of two parents. Proc. R. Soc. Lond. 1895, 58, 240–242. [Google Scholar]
  27. Hoque, N.; Bhattacharyya, D.K.; Kalita, J.K. MIFS-ND: A mutual information-based feature selection method. Expert Syst. Appl. 2014, 41, 6371–6385. [Google Scholar] [CrossRef]
  28. Kraskov, A.; Stögbauer, H.; Grassberger, P. Estimating mutual information. Phys. Rev. E 2004, 69, 066138. [Google Scholar] [CrossRef] [PubMed]
  29. Guyon, I.; Weston, J.; Barnhill, S.; Vapnik, V. Gene selection for cancer classification using support vector machines. Mach. Learn. 2002, 46, 389–422. [Google Scholar] [CrossRef]
  30. Blöchl, P.E. Projector augmented-wave method. Phys. Rev. B 1994, 50, 17953–17979. [Google Scholar] [CrossRef]
  31. Hafner, J. Ab-initio simulations of materials using VASP: Density-functional theory and beyond. J. Comput. Chem. 2008, 29, 2044–2078. [Google Scholar] [CrossRef]
  32. Perdew, J.P.; Burke, K.; Ernzerhof, M. Generalized gradient approximation made simple. Phys. Rev. Lett. 1996, 77, 3865–3868. [Google Scholar] [CrossRef] [PubMed]
  33. Kresse, G.; Furthmuller, J. Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set. Phys. Rev. B 1996, 54, 11169–11186. [Google Scholar] [CrossRef]
  34. Monkhorst, H.J.; Pack, J.D. Special points for Brillouin-zone integrations. Phys. Rev. B 1976, 13, 5188–5192. [Google Scholar] [CrossRef]
  35. Lee, J.; Ikeda, Y.; Tanaka, I. First-principles screening of structural properties of intermetallic compounds on martensitic transformation. NPJ Comput. Mater. 2017, 3, 52. [Google Scholar] [CrossRef]
  36. Entel, P.; Gruner, M.E.; Comtesse, D.; Wuttig, M. Interaction of phase transformation and magnetic properties of Heusler alloys: A density functional theory study. Jom 2013, 65, 1540–1549. [Google Scholar] [CrossRef]
  37. Schröter, M.; Herper, H.C.; Grünebohm, A. Tuning the magnetic phase diagram of Ni-Mn-Ga by Cr and Co substitution. J. Phys. D Appl. Phys. 2021, 55, 025002. [Google Scholar] [CrossRef]
  38. Sun, L.F.; He, Z.F.; Jia, N.; Guo, Y.X.; Jiang, S.; Yang, Y.L.; Liu, Y.X.; Guan, X.J.; Shen, Y.F.; Yan, H.L.; et al. Local chemical order enables an ultrastrong and ductile high-entropy alloy in a cryogenic environment. Sci. Adv. 2024, 10, eadq6398. [Google Scholar] [CrossRef]
  39. He, Z.F.; Guo, Y.X.; Sun, L.F.; Yan, H.L.; Guan, X.J.; Jiang, S.; Shen, Y.F.; Yin, W.; Zhao, X.L.; Li, Z.M.; et al. Interstitial-driven local chemical order enables ultrastrong face-centered cubic multicomponent alloys. Acta Mater. 2023, 243, 118495. [Google Scholar] [CrossRef]
  40. Zhao, Y.; Yan, H.L.; Xiang, H.Y.; Jia, N.; Yang, B.; Li, Z.; Zhang, Y.; Esling, C.; Zhao, X.; Zuo, L. Valence electron concentration and ferromagnetism govern precipitation in NiFeGa magnetic shape memory alloys. Acta Mater. 2024, 264, 119592. [Google Scholar] [CrossRef]
  41. Perez-Checa, A.; Feuchtwanger, J.; Barandiaran, J.M.; Sozinov, A.; Ullakko, K.; Chernenko, V.A. Ni-Mn-Ga-(Co, Fe, Cu) high temperature ferromagnetic shape memory alloys: Effect of Mn and Ga replacement by Cu. Scrip. Mater. 2018, 154, 131–133. [Google Scholar] [CrossRef]
  42. Lu, Y.; Yu, H.X.; Sisson, R.D. The effect of carbon content on the c/a ratio of as-quenched martensite in Fe-C alloys. Mater. Sci. Eng. A 2017, 700, 592–597. [Google Scholar] [CrossRef]
  43. Yan, H.-L.; Liu, H.-X.; Zhao, Y.; Jia, N.; Bai, J.; Yang, B.; Li, Z.; Zhang, Y.; Esling, C.; Zhao, X.; et al. Impact of B alloying on ductility and phase transition in the Ni–Mn-based magnetic shape memory alloys: Insights from first-principles calculation. J. Mater. Sci. Technol. 2021, 74, 27–34. [Google Scholar] [CrossRef]
  44. Yan, H.-L.; Wang, L.-D.; Liu, H.-X.; Huang, X.-M.; Jia, N.; Li, Z.-B.; Yang, B.; Zhang, Y.-D.; Esling, C.; Zhao, X.; et al. Giant elastocaloric effect and exceptional mechanical properties in an all-d-metal Ni–Mn–Ti alloy: Experimental and ab-initio studies. Mater. Des. 2019, 184, 108180. [Google Scholar] [CrossRef]
  45. Ong, S.P.; Cholia, S.; Jain, A.; Brafman, M.; Gunter, D.; Ceder, G.; Persson, K.A. The materials application programming interface (API): A simple, flexible and efficient API for materials data based on representational state transfer (REST) principles. Comput. Mater. Sci. 2015, 97, 209–215. [Google Scholar] [CrossRef]
  46. gplearn: Genetic Programming in Python. Available online: https://github.com/trevorstephens/gplearn (accessed on 9 May 2025).
Figure 1. TM of Ni2MnGa1−xZx alloys (x = 0.25, 0.5, 0.75 and 1) from high-throughput ab initio calculations.
Figure 1. TM of Ni2MnGa1−xZx alloys (x = 0.25, 0.5, 0.75 and 1) from high-throughput ab initio calculations.
Materials 18 02226 g001
Figure 2. The relationship between the martensitic transformation temperature TM of alloys in the dataset and known empirical rules (a) VEC rule and (b) Volume rule. Neither VEC nor V exhibited a strong or reliable correlation with TM when considering the full dataset.
Figure 2. The relationship between the martensitic transformation temperature TM of alloys in the dataset and known empirical rules (a) VEC rule and (b) Volume rule. Neither VEC nor V exhibited a strong or reliable correlation with TM when considering the full dataset.
Materials 18 02226 g002
Figure 3. Workflow of adopted five-step descriptor screening.
Figure 3. Workflow of adopted five-step descriptor screening.
Materials 18 02226 g003
Figure 4. Machine learning screening of key parameters affecting martensitic transformation temperature. (a) Univariate screening; (b) Recursive elimination; and (c) Exhaustive screening.
Figure 4. Machine learning screening of key parameters affecting martensitic transformation temperature. (a) Univariate screening; (b) Recursive elimination; and (c) Exhaustive screening.
Materials 18 02226 g004
Figure 5. The result of 10-cross validation and the importance of descriptors in machine learning models.
Figure 5. The result of 10-cross validation and the importance of descriptors in machine learning models.
Materials 18 02226 g005
Figure 6. The visualization of the relationship between the target TM and the screened key descriptors. (a) M P ¯ and ρ ¯ ; (b) Enlarged view of the black rectangular region in panel (a); (c) G ¯ and I E ¯ ; (d) σAR and H C ¯ .
Figure 6. The visualization of the relationship between the target TM and the screened key descriptors. (a) M P ¯ and ρ ¯ ; (b) Enlarged view of the black rectangular region in panel (a); (c) G ¯ and I E ¯ ; (d) σAR and H C ¯ .
Materials 18 02226 g006
Figure 7. The visualization of the relationship between the TM and ρ ¯ · M P ¯ of the dataset. The black line represents the fitted empirical formula TM = 82 ( ρ ¯ · M P ¯ ) 700 . The red and pink shaded region correspond to the 95% confidence interval and 95% prediction interval, respectively. The inset plot shows the distribution of prediction errors.
Figure 7. The visualization of the relationship between the TM and ρ ¯ · M P ¯ of the dataset. The black line represents the fitted empirical formula TM = 82 ( ρ ¯ · M P ¯ ) 700 . The red and pink shaded region correspond to the 95% confidence interval and 95% prediction interval, respectively. The inset plot shows the distribution of prediction errors.
Materials 18 02226 g007
Table 1. Adopted elemental, simple substance, and phase stability-associated features.
Table 1. Adopted elemental, simple substance, and phase stability-associated features.
Feature CategoryFeature Description Abbreviation
Elemental propertiesAtomic number Z
Periodic table columnC
Atomic weightAW
Mendeleev numberMN
Periodic table rowPR
Atomic radiusAR
Number of s valence electronsNs
Number of p valence electronsNp
Number of d valence electronsNd
Number of f valence electronsNf
Number of total valence electronsNt
Number of unfilled s statesUs
Number of unfilled p statesUp
Number of unfilled d statesUd
Number of unfilled f statesUf
Number of total unfilled statesUt
Simple substance propertiesMelting pointMP
Boiling pointBP
Heat capacityHC
Heat fusionHF
Pauling ElectronegativityEN
Covalent radiusCR
Ionic radiusIR
Density ρ ¯
Magnetic momentM
VolumeV
Band gapGap
First ionization energyE
Space group numberSG
Bulk modulusB
Shear modulusG
Phase stability propertiesMixing entropy S m i x
Atomic size difference δ
Table 2. The selected descriptors for the univariate screening, recursive elimination, and exhaustive screening methods for the dataset.
Table 2. The selected descriptors for the univariate screening, recursive elimination, and exhaustive screening methods for the dataset.
Descriptors (Abbr.)Descriptors (Abbr.)
Univariate screening Standard deviation of melting point (σMP)Weight-average value of p valence electron ( V E ¯ p)
Weight-average value of valence electron ( V E ¯ )Weight-average value of s valence electron ( V E ¯ s)
Weight-average value magnetic moment ( M a g ¯ )Weight-average value of bulk modulus ( G ¯ )
Standard deviation of Mendeleev number (σMN)Standard deviation of heat fusion (σHF)
Standard deviation of valence electron (VEsd)Standard deviation of s valence electron (σVEs)
Recursive elimination screeningExhaustive screeningWeight-average value of melting point ( M P ¯ )Standard deviation of atomic radius (σAR)
Weight-average value of ionization energy ( I E ¯ )Weight-average value of shear modulus ( G ¯ )
Weight-average value of heat capacity ( H C ¯ )Weight-average value of density ( ρ ¯ )
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Liu, H.-X.; Yan, H.-L.; Jia, N.; Yang, B.; Li, Z.; Zhao, X.; Zuo, L. Machine Learning-Assisted Discovery of Empirical Rule for Martensite Transition Temperature of Shape Memory Alloys. Materials 2025, 18, 2226. https://doi.org/10.3390/ma18102226

AMA Style

Liu H-X, Yan H-L, Jia N, Yang B, Li Z, Zhao X, Zuo L. Machine Learning-Assisted Discovery of Empirical Rule for Martensite Transition Temperature of Shape Memory Alloys. Materials. 2025; 18(10):2226. https://doi.org/10.3390/ma18102226

Chicago/Turabian Style

Liu, Hao-Xuan, Hai-Le Yan, Nan Jia, Bo Yang, Zongbin Li, Xiang Zhao, and Liang Zuo. 2025. "Machine Learning-Assisted Discovery of Empirical Rule for Martensite Transition Temperature of Shape Memory Alloys" Materials 18, no. 10: 2226. https://doi.org/10.3390/ma18102226

APA Style

Liu, H.-X., Yan, H.-L., Jia, N., Yang, B., Li, Z., Zhao, X., & Zuo, L. (2025). Machine Learning-Assisted Discovery of Empirical Rule for Martensite Transition Temperature of Shape Memory Alloys. Materials, 18(10), 2226. https://doi.org/10.3390/ma18102226

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop