You are on page 1of 12

www.acsami.

org Research Article

Discovery of Low-Modulus Ti-Nb-Zr Alloys Based on Machine


Learning and First-Principles Calculations
Camilo A. F. Salvador, Bruno F. Zornio, and Caetano R. Miranda*
Cite This: https://dx.doi.org/10.1021/acsami.0c18506 Read Online

ACCESS Metrics & More Article Recommendations *


sı Supporting Information
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

ABSTRACT: The discovery of low-modulus Ti alloys for


biomedical applications is challenging due to a vast number of
Downloaded via UNIV OF GOTHENBURG on December 14, 2020 at 01:39:30 (UTC).

compositions and available solute contents. In this work, machine


learning (ML) methods are employed for the prediction of the
bulk modulus (K) and the shear modulus (G) of optimized ternary
alloys. As a starting point, the elasticity data of more than 1800
compounds from the Materials Project fed linear models, random
forest regressors, and artificial neural networks (NN), with the
aims of training predictive models for K and G based on
compositional features. The models were then used to predict the
resultant Young modulus (E) for all possible compositions in the
Ti-Nb-Zr system, with variations in the composition of 2 at. %.
Random forest (RF) predictions of E deviate from the NN
predictions by less than 4 GPa, which is within the expected variance from the ML training phase. RF regressors seem to generate
the most reliable models, given the selected target variables and descriptors. Optimal compositions identified by the ML models were
later investigated with the aid of special quasi-random structures (SQSs) and density functional theory (DFT). According to a
combined analysis, alloys with 22 Zr (at. %) are promising structural materials to the biomedical field, given their low elastic
modulus and elevated beta-phase stability. In alloys with Nb content higher than 14.8 (at. %), the beta phase has lower energy than
omega, which may be enough to avoid the formation of omega, a high-modulus phase, during manufacturing.
KEYWORDS: machine learning, first-principles calculations, metals and alloys, elasticity, phase transitions

1. INTRODUCTION the physical properties of solids with reasonable accuracy. De


Despite the increasing demand for high-performance alloys by Jong et al. (2016) were able to predict the elastic bulk (K) and
the biomedical,1 aeronautical/aerospace,2 and oil & gas shear moduli (G) behavior of k-nary inorganic compounds (up
industries,3 the traditional materials discovery process still to quaternary) using a gradient-boost framework based on
relies on empirical methodologies, based on significant trial local multivariate regression.9 Later on, Winter et al. (2017)
and error.4−6 Not only this has been the dominant exploration identified potential superhard materials from space groups 229
methodology for centuries, but materials design also involves and 225 defining an elastic anisotropy parameter, which is a
an intricate, multidimensional optimization problem, in which function of K and G, and then confirmed their unusual elastic
the links between parameters and specific properties of interest properties via DFT.10
Low-modulus Ti alloys find applications as structural
are usually unknown.
Nevertheless, with the launch of The Materials Genome biomaterials, especially among knee and hip-joint implants.
Initiative (MGI), materials and data scientists have been In these cases, ideally, the material must have a bonelike elastic
working together to develop multiscale computational method- modulus, in order to mitigate the stress-shielding effect after
ologies to assess, evaluate, and optimize materials regarding implantation and, therefore, avoid the proximal bone
their fundamental properties.7 As an example, the initiative degeneration around the implant.11 As a reference, Ti-Nb-
named Materials Project consists of an open platform that has Ta-Zr alloys dominate this field of application, with an elastic
been creating, validating, and storing data of compounds for
which experimental data may be incomplete or absent.8 At Received: October 14, 2020
present, there are 120,612 inorganic compounds’ data available Accepted: December 1, 2020
at the Materials Project, mainly obtained via first-principles
calculations. A full elastic description of these systems (i.e.,
elastic tensor) is defined for 13,621 entries. Based on the
previous works,9,10 this dataset is more than enough to model

© XXXX American Chemical Society https://dx.doi.org/10.1021/acsami.0c18506


A ACS Appl. Mater. Interfaces XXXX, XXX, XXX−XXX
ACS Applied Materials & Interfaces www.acsami.org Research Article

Figure 1. Elastic property distributions for 1826 entries in the filtered database. (a) Bulk modulus (K). (b) Shear modulus (G). (c) Poisson ratio.
The values presented are Voigt−Reuss−Hill (VRH) averages. A summary of this data is compiled in Table S1 in the Supporting Information file.

modulus close to 55 GPa, at the solution-treated and water- structure, maxing out −1 or + 1% true deformation, within 8 steps,
quenched condition.12 The Ti-Nb-Zr system explored herein and single-point calculations were performed to obtain the energy of
each deformed state. The deformations were set as follows25
ij 1 y
was arbitrary chosen based on the recent literature, which
jj− δ
jj 2 0 zzzz
zz
reported a number of promising candidates from low-cost
jj zz
j
systems.13,14 Furthermore, studies with a ternary system might 0

Tetragonal: jjjj zz, with U (δ)


jj 0 − δ 0 zzzz
provide ground to incremental works in the near future, which
jjj zz
1

jjj zz
can explore a great range of quaternary systems derived from it. tetra

z
k 0 + δ z{
In the present paper, we aim to search for new, low-modulus 2
Ti alloys for biomedical applications with the aid of the MP 0
database and the matminer library.15 For exploration purposes, 3
= δ 2(C11 − C12)
we decided to test three different machine learning algorithms 4 (1)

ij 0 + δ 0 yz
jj zz
in this work: multivariate linear regression models, random

Trigonal: jjjj+ δ 0 0 zzzz, with Utri(δ) = 2δ 2C44


forest regressors, and neural networks. Later on, compositions
jj zz
from a selected range were further characterized by first-
k 0 0 0{
principles calculations based on density functional theory (2)
(DFT) using special quasi-random structures (SQSs).
Elastic modulus estimations were obtained based on the Voigt−
2. COMPUTATIONAL METHODS Reuss−Hill approximation, with the elastic constants being derived
from the bulk modulus considering a body-centered cubic (BCC)
Machine learning codes were implemented in python using
system
pymatgen,16 sci-kit learn, Keras, and TensorFlow.17 For all of the
machine learning algorithms, we performed random search Bulk modulus, K = 1/9(C11 + 2(C12)) (3)
optimizations for the hyperparameters and used 5-fold cross-
validation. For the neural networks, to avoid overfitting, the root- Shear modulus, G = 1/15(C11 − C12 + 3(C44)) (4)
mean-squared error (RMSE) and mean absolute error (MAE) were
registered for each training epoch (step), and an earlier stop was Elastic modulus, E = 9KG /(3K + G) (5)
defined based on the loss function. CALPHAD calculations were
performed with the aid of the MatCalc suite, developed by
Kozeschnik.18 3. MACHINE LEARNING MODELS
DFT calculations were performed using the Vienna ab initio
simulation package (VASP).19 The projected augmented wave 3.1. Initial Dataset, Screening, and Featurization. A
(PAW)20 methodology was used to treat the core electronic density, few considerations were made regarding the original MP
and the Perdew−Burke−Ernzerhof (PBE) exchange-correlation dataset before deploying the ML algorithms. From the starting
functional was employed.21 The plane-wave energy cutoff was set to 13,619 entries, the dataset was limited to compounds
350 eV, and a 5x5x5 mesh grid sampled the Brillouin zone for all containing transition metals and post-transition metals only.
systems, to reach the convergence criteria adopted. The idea behind this screening was to exclude compounds in
To obtain an atomistic description of our body-centered cubic which bonds have a predominant ionic or covalent character.
(bcc) solid solutions, special quasi-random structures (SQSs) were Four thousand two hundred sixty-five (4265) entries remained
built with the ATAT package.22 The ATAT code aims to obtain the
best SQS for a given system size by maximizing the number of after the first filtering step. Following a few modeling tryouts,
possible radial correlations from the fully disordered state in the the search space was restricted to fit the elastic property range
supercell.23 In our case, the cells were fixed to 3x3x3 times the in which Ti alloys are found, i.e., K smaller than 150 GPa, G
primitive cell, resulting in 54 atoms. The use of SQSs in the smaller than 100 GPa, and average Poisson ratio between 0
determination of elastic properties of random alloys has been proved and 1. Only 2566 entries were kept after this step. It is worth
accurate for many systems in which long-range interactions are reminding that our approach was focused on finding elastic
relatively small. According to von Pezold et al. (2010), in 3x3x3 face- minima and subtle differences in the elastic properties between
centered cubic (fcc) supercells, the sampling error associated with this similar compositions. Therefore, we were interested in
technique is less than 1%, which is within the accuracy of DFT modeling the low-modulus region as best as possible. Using
calculations.24
From the lattice parameter optimization, we employed the all data, the relative error associated with the low-modulus
Murnaghan equation to obtain the bulk modulus of each cell. After region increases, with the model missing most of the values
volume optimization, the cells were fully relaxed, until the from our region of interest. In other words, the model would
convergence criteria of forces smaller than 0.01 eV/Å are reached. adjust to relatively high-modulus materials, which are prevalent
Then, tetragonal and trigonal strains were applied to the optimized in the dataset.26
B https://dx.doi.org/10.1021/acsami.0c18506
ACS Appl. Mater. Interfaces XXXX, XXX, XXX−XXX
ACS Applied Materials & Interfaces www.acsami.org Research Article

Figure 2. Twenty most important features to predict bulk modulus K (a) and shear modulus G (b) according to a univariate feature selection
(unsupervised). An additional analysis with the incorporation of mutual information (c,d) is also presented. As a reference for this project, a score
near 10.000 would represent a correlation factor near 0.1 between the feature and target.

For the ML model parameterization, we sought composi- nation (R2). The 20 most important features obtained with this
tion-based descriptors suitably depicting the elastic properties. methodology, considering K as a target, are shown in Figure 2.
From the matminer library, the following featurization classes For the bulk modulus (K) determination, it seems that all
were implemented: ElementProperty(magpie), BandCenter, features are well-balanced, being equally important, despite the
ValenceOrbital, IonProperty, TMetalFraction, CohesiveEnergy, following three prominent important features: mean ground-
and AtomicPackingEff iciency. A detailed description of each state volume per atom (GSVolume), the fraction of magnetic
feature generated with these classes can be found in Ward et al. transition metals (TMetalFraction) in the composition, and
(2018).15 In addition to those, we also exploited the Miedema the cohesive energy (Figure 2a). Analogous behavior was
and Yang solid solution featurizers, which are semi-empirical observed by testing a mutual info regression function (Figure
models for solid solutions. The Miedema model is based on 2c), which considers pairs of features for the fit. As can be seen
the heat of formation of binary compounds and has shown a in Figure 2c, all features have a similar score. None of them
high reproducibility in regard to experimental data.27 The stands out, which is an indication that only high-capacity
model proposed by Yang and Zhang is focused on multi- models, including a high number of descriptive features, would
component alloys and incorporates the entropy of mixing and be able to model K.
the effect of the atomic size differences in a free energy-based The same trend is observed when the shear modulus (G) is
model.28 In total, 158 descriptors were obtained. set as the target variable (Figure 2b). However, relevant
During the remaining dataset screening for the featurization features for G are not the same as for K. In this case, only the
methods, 551 entries were lost due to the absence of important Poisson ratio is highlighted. The Poisson ratio also stands out
information or because one of its elements was not included in during the mutual information regression analysis (Figure 2d).
the Miedema model. Typical compounds filtered in this step It is important to remark that this was expected. After all, the
contained Yb, Pm, and Pa in their composition. Finally, 208 Poisson ratio and G are physically related to each other.
entries with a single or exact oxidation state were excluded to However, according to a recent paper, the Poisson ratio is
avoid ionic compounds, one more time. A summary of the final challenging to be modeled based on composition, with the data
1826 instances’ K, G, and Poisson ratio values is shown in currently available.31 Therefore, incorporating the Poisson
Figure 1. ratio to the model creates a constraint that will be further
3.2. Feature Analysis. An important exploratory step in discussed in Section 3.5.
ML projects is to check which features are potentially the best Now that a preliminary data selection and an analysis of the
to describe a specific target property.29,30 To achieve that, we features were concluded, we present the models implemented
employed a univariate feature selection method implemented during this work. Based on this section, the bulk modulus (K)
in Keras as SelectKBest.17 This technique consists of performing and the shear modulus (G) are assumed to depend on all the
one linear fit for each feature-target pair, and then, it ranks the available descriptors; therefore, to avoid any bias by selecting
features based on their independent coefficient of determi- preferred features by hand, the initial models incorporated all
C https://dx.doi.org/10.1021/acsami.0c18506
ACS Appl. Mater. Interfaces XXXX, XXX, XXX−XXX
ACS Applied Materials & Interfaces www.acsami.org Research Article

of the 158 available features. The properties K and G were not With the random forest model optimized, we could use the
available as descriptors, in any case. built-in Extra Tree Classifier to obtain the relative importance
3.3. Linear and Random Forest Regressors. Linear of each feature to the final forest. This method roughly
regression models are least-squares linear fits in a multidimen- estimates how important one variable (descriptor) is in terms
sional space. Usually, linear models establish a baseline for of predicting values in a multidimensional space. The top 20
more elaborate models. However, they can hardly ever solve features identified for each target for the best random forest
complex regression problems. In this case, the least absolute regressor are displayed in Figure 4. For the shear modulus (G)
shrinkage and selection operator (LASSO) regularization is prediction, the Poisson ratio seems to be a dominant
often used to create sparse solutions to the model via variable descriptor, as expected. Properties related to the ground-state
selection and ultimately improve the prediction performance.31 volume per atom (GSVolume), cohesive energy, and melting
On the other hand, nonlinear machine learning algorithms, temperature were selected by both models, thus being relevant
such as decision trees, have become popular across the physical to both target variables. Other relevant properties seem to be
sciences.32 Decision trees are hierarchical models that aim to associated with the covalent radius, electronegativity, the
find a target value by asking the fewer if-else questions. Each number of p-valence electrons, and the position in the periodic
node on a tree splits the dataset once according to whether a table (period and group).
specific parameter, or a combination of them, is above or Having obtained two baseline models, we proceeded with
below a threshold value.33 Random forest (RF) regressors are the use of neural networks (NN). However, unlike the others,
simply a combination of multiple regression trees. The the implementation of NN is not simple, as complex networks
predictive power of RF regressors comes from the diversity are prone to overfit.35,36
of the trees/branches on them. However, an excessive number 3.4. Neural Networks. Neural networks are nonlinear
of trees, or estimators, can lead to overfitting. Other essential models that mimic the operation of the human brain. A
parameters of RF models are the criterion to define a split network of neurons, also called perceptrons, takes an input
(usually, a minimization of the RMSE), the number of features value (e.g., a variable descriptor or a combination of them),
considered for splitting a node, the maximum number of nodes performs basic arithmetic operations, and then subjects it to an
(splits), and the maximum allowed depth. activation function. In our case, we implemented sequential
At first, the predictive accuracies of linear regression and neural networks with three hidden layers. The number of
random forest regressors were evaluated. The random forest neurons varied between 30 and 300, which was optimized for
models were optimized based on a random search algorithm, each case via random search. The activation function was the
searching for 50 different combinations of hyperparameters. All rectified linear activation function, also known as ReLU, and
models used 80% of the data as the training set to predict the the optimization occurred via the stochastic gradient
unseen 20%. For the full set of features, the following RMSEs descendant (adam), implemented with the default hyper-
were obtained after optimization (Table 1). More details about parameters from Keras/TensorFlow.17 An illustration of the
each model will be discussed further on. neural networks used throughout this work is shown in the
Supporting Information file (Figure S3).
Table 1. Linear Models, Random Forest Regressors, and The performance of the NN is indicated in Figure 5. Models
Neural NetworksRoot-Mean-Square Errors (RMSE) for (a) and (c) (left-hand side), which were implemented with the
each Target, K (Bulk) and G (Shear)a full set of descriptors (147) and a higher number of neurons,
tend to overfit easily, while models (b) and (d) displayed a
model RMSE for K RMSE for G reduced test-set RMSE, despite the higher dispersion for the
LM training (testing) 11.38 (15.87) 6.49 (7.45) training data (in blue). Although high-capacity models might
LM+LASSO, training (testing) 12.81 (14.67) 7.48 (7.58) be able to achieve remarkably low RMSE during training, they
LM+ElasticNet, training (testing) 15.05 (16.47) 11.28 (12.39) diverge from a generalized model, as confirmed by the
RF, training (testing) 4.99 (12.39) 2.38 (5.91) scattering of test-set points (gray) in Figure 5a,c. In this way,
full NN, training (testing) 3.90 (13.62) 2.05 (5.97) only low-capacity neural networks could surpass the test
reduced NN, training (testing) 8.84 (12.04) 3.36 (5.18) prediction capabilities of the random forest regressors (Table
a
The RMSE of the models adopted for Figures 3 and 4 are in bold; 1). The results with NN suggest that some knowledge on
full and reduced configurations are presented in Figure 5. feature importance can help to obtain generalized models
avoiding overfit. On the other hand, the implementation of the
random forest method is much more straightforward since
For the linear regression models, introducing LASSO feature selection is built into the model.
regularization improved the test-set predictive performance In the pursuit of ultrahard materials, Mansouri Tehrani et
from 15.87 to 14.67 (RMSE) for K (Figure 3). However, it was al.26 employed support vector machine (SVM) regressors to
not efficient regarding the shear modulus G. The inclusion of model K and G in a workflow similar to the one presented
ElasticNet, another regularization method, did not improve here. The authors have demonstrated a good agreement
any of the models. Based on our observations, the inclusion of between SVR predictions and DFT data, especially for
regularization methods seems to be ineffective when the G is compounds with bonds of metallic or covalent nature.
the target variable. Regarding feature importance, our models share most of the
For the bulk modulus (K), the random forest regressor relevant parameters highlighted in Mansouri Tehrani et al.’s
achieved a notably low RMSE on the training step (4.99) and a work, corroborating with their analyses. The main difference
quite good RMSE during testing (12.39) (Table 1). Recent lies in the absence of structural parameters in our models.
works also displayed the power of random forest regressors to Thus, in this paper, descriptors associated with the electro-
predict the bulk moduli of crystalline materials based on negativity, valence electrons, and cohesive energy are highly
experimental data.34 influential for G predictions (Figure 2). It is worth noting that
D https://dx.doi.org/10.1021/acsami.0c18506
ACS Appl. Mater. Interfaces XXXX, XXX, XXX−XXX
ACS Applied Materials & Interfaces www.acsami.org Research Article

Figure 3. (a,c) Modeling the bulk (K) and shear (G) modulus with linear and (b,d) random forest regressors (in red). The performance of the RF
is notably better, especially when comparing the test-set absolute error distributions (insets). LASSO regularization was included in the linear
model displayed in (a).

Figure 4. Feature importance according to an Extra Tree Classifier for (a) bulk (K) and (b) shear (G) modulus. The final forests had an average of
1928 and 874 nodes and an average maximum depth of 22 and 18 for K and G, respectively.

despite including 14 structural descriptors to their models, predictions, we assumed the Poisson ratio for all alloys in
Mansouri Tehrani et al. also reported difficulty in determining the Ti-Nb-Zr system being proportional to the configuration
the shear modulus (G), and a cross-validated RMSE of 16.5 entropy.37 The idea behind it is to identify candidates far from
GPa was obtained for G. In our case, we could obtain a lower the diagram’s edges (pure elements). To obtain elastic moduli
RMSE due to the use of the Poisson ratio as a descriptor and a (E) comparable with experimental results, it is important to set
much narrower dataset (low-modulus centered, excluding ionic the Poisson ratio to reasonable values. In our case, we have
compounds). interpolated the Poisson ratio between 0.25 and 0.45, typical
3.5. Elastic Property Predictions. Using ML models to values among Ti alloys. While this might be considered a
predict the elastic properties of alloys from a system of interest methodological artifact, such assumptions are often seen in the
is the main goal of this study. Looking forward to G specialized literature, for example, when estimating the elastic
E https://dx.doi.org/10.1021/acsami.0c18506
ACS Appl. Mater. Interfaces XXXX, XXX, XXX−XXX
ACS Applied Materials & Interfaces www.acsami.org Research Article

Figure 5. (a,b) Modeling K (bulk modulus) and (c,d) G (shear modulus) with neural networks. The models (a) and (c) were implemented with
the full set of descriptors. For models (b) and (d), only the selected descriptors (Figures 2 and 4) were made available, and neurons were limited to
50.

modulus via nanoindentation measurements.38 As stated impressive yield strength, higher than 1 GPa.40 The second
before, the idea of this work was to make elastic predictions region of interest is located at near 80% Ti. On this island,
based on the composition only. In this sense, the average highlighted by a red circle (Figure 6e), the elastic modulus
Poisson’s ratio is an important feature to improve the shear seems to be homogeneous and lower than that in its vicinity.
modulus predictions since it carries pieces of structural The region is extended while Ti is being substituted with Nb,
information linked with the elastic anisotropy. Given that i.e., to the northwest of the diagram. Kim et al. recently
our models are phase agnostic, an average Poisson ratio can be explored Ti-Nb-Zr alloys within this compositional range and
assumed (or measured) for multiphase alloys, thus enabling found out a surprisingly low elastic modulus, of 39 GPa, among
one to make use of the models to explore multiphase systems Ti-14Nb-18Zr (at. %) samples.41 Also, a few commercial
without previous structural knowledge. More information alloys, such as Ti-13Nb-13Zr, and gum metals could be fit into
about the Poisson ratio assumptions can be found in the this region. In other words, the ML models’ results are well-
Supporting Information file. supported by previous experimental observations.
The elastic property predictions by the RF (left side) and In general, variations on elastic properties predicted via NN
NN models (right side) are depicted in the ternary diagrams are more subtle. The reduced model performed well while
presented in Figure 6. The RF regressors do not capture well predicting K. On the other hand, G predictions were flat, with
variations of the bulk modulus (Figure 6a). Nevertheless, they no significant modulus variation along with the ternary
exhibit two standout regions (Figure 6c,e) with a low shear and diagram. Therefore, since we were looking specifically for
low resultant Young modulus. Both regions are spotted close divergent points in which the G and E could be both low, we
to the bottom line of the diagram. The first region is the Ti-Zr returned to the NN with all descriptors to make the
near-equiatomic proportions. In a recent assessment, Shiraishi predictions for G, even though the test-RMSE was slightly
et al. identified a minimum Young modulus for the Ti-Zr worse than the reduced one (test-RMSE of 5.97 > 5.18). The
binary system close to this region. According to the authors, minimum elastic modulus obtained for both of the models is
Ti-60Zr alloys present an elastic modulus of 89 GPa, with a around 67 GPa (Table 2). Despite the lower test-set RMSE for
relatively low shear modulus, 33 GPa.39 Recently explored Ti- NN, the final predictions of each model converge in the critical
Zr-Ta ternary alloys close to this composition displayed an regions of the compositional space. However, we do not expect
F https://dx.doi.org/10.1021/acsami.0c18506
ACS Appl. Mater. Interfaces XXXX, XXX, XXX−XXX
ACS Applied Materials & Interfaces www.acsami.org Research Article

Figure 6. (a, c, and e) Elastic property predictions by the RF and (b, d, and f) NNs. The region of interest is highlighted in (e) and (f), in red.

Table 2. Predictions of the Ti-22Zr-Nb (at. %) Alloys’ tional theory (DFT) to evaluate the stability of the beta phase
Elastic Modulus (E) from the Selected Region in several Ti-Nb-Zr alloys selected from the optimal region
highlighted in Figure 6.
Nb content (at. %) NN E (GPa) RF E (GPa)
At RT (room temperature), pure Ti exhibits a hexagonal
0 67.0 78.3 phase known as alpha (α-phase). At 882 °C, Ti displays an
4 71.0 78.7 allotropic transformation known as β-transus, which marks the
7 72.8 78.9 transition from α-phase to β (bcc). Alloying elements
11 74.7 79.5 considered that β-stabilizers such as Nb, Ta, Mo, and V may
15 77.3 80.6 reduce the β-transus temperature, allowing the β-phase to form
18 79.6 81.4 at lower temperatures.42 However, when β-structures are fast
22 82.2 81.9 cooled, an athermal omega phase can be formed on quenching
78 137.1 130.3 due to a first-order structural transition. Athermal ω is
metastable and has a size limited to a few nanometers, and
the absolute values predicted via ML models (Table 2) to be its formation occurs by the collapse of alternate pairs of (222)
exact compared to experimental verification. β planes, through a diffusionless mechanism.43 It often has a
deleterious effect on biomedical alloys, increasing the elastic
4. FIRST-PRINCIPLES AND CALPHAD CALCULATIONS modulus.44
4.1. Phase Stability. While our ML models do not require In the compositional range explored in this work, alloys
any structural information, they also cannot predict which should present a predominant beta-phase matrix when fast
phase or crystal structure might be related to the forecast cooled to room temperature, given that omega formation can
properties shown in Figure 6. To address this limitation and be avoided.45 Therefore, an assessment of the energy
also compare the ML results with a traditional approach, we formations of metastable phases, such as the omega
employed first-principles calculations based on density func- phase,46,47 is presented. To build the omega supercells, we
G https://dx.doi.org/10.1021/acsami.0c18506
ACS Appl. Mater. Interfaces XXXX, XXX, XXX−XXX
ACS Applied Materials & Interfaces www.acsami.org Research Article

Figure 7. Phase stability of beta (bcc) relative to the omega phase. (a) Energy difference between the two phases. (b) Degree of collapse of the
omega structures after relaxation. Illustrations of omega (c) and beta (d) structures after relaxation. Note that (c) and (d) are shown in perspective
and aligned to the (110) beta zone axis, a direction in which the collapse of (222) planes is evident. The supercells use CPK standard coloring for
the atoms.

explicitly modified the position of 36/54 atoms on the <111> temperature for alloys with Zr content near 22 at. % is close
beta direction to mimic the collapse of the {111} beta planes to 700 °C (Figure 8). Below this temperature, the alpha phase
observed in the first-order beta to omega phase transition
(Figure 7). From the fully collapsed structure, we estimated
total energies before and after relaxation with the aid of DFT.
The differences in energy between beta and omega after
relaxation are presented in Figure 7a. Regarding the structures’
energies, it could be inferred that an Nb content higher than
11.1 at. % is needed to stabilize the beta phase relative to
omega.
Still, to further evaluate the omega-phase stability, we take
the omega structures after relaxation, check each atomic
position, and compare it to the prior beta structure. A degree
of collapse (DOC) of 1 means that the atoms are at the same
positions expected for a fully collapsed, standard omega crystal
structure. On the other hand, a DOC = 0 means that the
omega structure fully relaxed into beta, i.e., the atoms returned
to their original positions (as in the prior beta phase). The
DOC for Ti-22Zr alloys with Nb contents varying from 0 to
22.2 at. % can be seen in Figure 7b. As expected, the DOC
tends to 1 in compositions whose omega energy is lower than Figure 8. CALPHAD assessments (beta-transus of three selected
beta energy, such as in the Ti-22Zr binary alloy. On the other alloys with Zr = 22 at. % (from left to right) are approx. 725, 680, and
hand, if the Nb content is higher than 11.1 at. %, then a 665 °C, respectively).
partially collapsed omega is obtained after relaxation, in a
similar fashion as reported by Devaraj et al.48 Accordingly,
among these compositions, the energy of the resultant beta should be formed by rejecting Nb to the beta matrix. The beta-
phase is lower than of the omega phase, after relaxation. stabilizing effect of Nb on Ti alloys is well-documented in the
Following our results, if formed, then the omega phase should literature.42 Our results also display this effect since the
be partially collapsed in alloys Ti-22Zr-14.8Nb and Ti-22Zr- increase in Nb causes a reduction of the beta-phase energy.
18.5Nb and completely unstable if the Nb content is higher However, while beta must be the dominant phase, it is still
than 22.2 Nb (at. %). possible to find the alpha phase (hexagonal, HCP) when the
Elements like Zr and Sn may help destabilizing omega, alloys are slowly cooled from temperatures above the beta-
which is often associated with a high modulus, and benefit the transus. Therefore, to keep the elastic modulus low, alloys must
alpha” phase formation. In this case, alpha” is a metastable be fast cooled to retain as much beta phase as possible.51 Water
phase, which has composition and structure much similar to quenching may be enough to avoid the formation of alpha,
beta, and it is assumed that their elastic properties should be eventually causing a minor formation of alpha” instead.52
similar.49 According to experimental evidence, the presence of 4.3. Elastic Properties. The analyses presented in Section
alpha” is pronounced in this system because Zr might help to 4.1 suggest that it is reasonable to assume that the beta phase
stabilize the beta phase energetically but not elastically, as will be dominant among these alloys since omega is unstable in
suggested by Hu et al. (2008).50 compositions with an Nb content higher than 18.5 at. %. Also,
4.2. Beta-transus. Now let us turn our attention to the it is established that alpha-phase formation is diffusion-
formation of the equilibrium alpha phase among these alloys. controlled and can be avoided by water quenching of alloys
Thermodynamic databases indicate that the beta-transus from the beta-phase field.53 All of the elastic properties
H https://dx.doi.org/10.1021/acsami.0c18506
ACS Appl. Mater. Interfaces XXXX, XXX, XXX−XXX
ACS Applied Materials & Interfaces www.acsami.org Research Article

Table 3. Elastic Coefficients and the Average Elastic Modulus (E), Shear Modulus (G), Poisson Ratio, and Voigt Elastic
Anisotropy (Eanis.), Obtained via DFT
Nb content (at. %) K (GPa) C11 (GPa) C12 (GPa) C44 (GPa) E (GPa) G (GPa) Poisson ratio Eanis. (Young)
0.0 102.3 109.2 98.9 14.3 30.8 10.6 0.45 2.70
3.7 90.6 96.8 87.5 14.3 30.2 10.4 0.44 2.97
7.4 90.6 96.5 87.6 17.6 35.4 12.3 0.43 3.78
11.1 122.9 127.9 120.4 14.6 29.9 10.3 0.46 3.78
14.8 104.9 110.0 102.3 15.1 30.8 10.6 0.45 3.79
18.5 104.9 109.5 102.6 15.2 30.5 10.5 0.45 4.25
22.2 135.7 133.6 139.9 14.9 27.4 9.4 0.47 13.72
77.8 144.6 151.9 141.0 19.8 40.9 14.1 0.45 3.52

Figure 9. (a−c) 2D projections of the Young modulus, (d−f) shear modulus, and (g−i) Poisson ratio, along with the crystal orientation. For the
shear modulus and the Poisson ratio, maximum possible values are plotted in blue and minimum possible values in green. Negative values are
plotted in red (f). Since the selected alloys have similar stiffness coefficients, they are represented by only one composition, at the central column.

presented herein were modeled within 3x3x3 BCC supercells, constants from compositions in the optimal region identified
yielding 54 atoms, as an initial structure. via ML are similar to each other. Selected alloys, with Nb
The stiffness tensor coefficients (Cij, GPa) for a given contents varying from 0 to 22 at. %, present an elastic modulus
structure are defined as the derivative of the potential energy smaller than 40 GPa, with the elastic anisotropy growing with
with respect to small strain deformations, at corresponding the Nb content. These alloys also satisfy the Born−Huang
local minima. For cubic structures, only three coefficients are stability criteria, i.e., C11 − C12 > 0 and C12 > C44, which points
needed.25 In the present work, C11 and C12 were obtained via to the high elastic stability of the beta phase in this
tetragonal strains, while C44 was obtained via trigonal strains, compositional range.
subjecting the structure to shear (more details are provided in Still, the origins of such a low elastic modulus (40 GPa) are
the Methods section). The elastic coefficients presented herein unknown. Previous authors suggested that finding alloys with
were obtained based on strains from −1 to 1% of the the difference between elastic constants C11 and C22 tending to
optimized structure and were based on the Voigt−Reuss−Hill zero is fundamental. As a consequence of that, the single-
(VRH) approximation. As one can see in Table 3, the elastic crystal Young modulus in the <001> direction and the shear
I https://dx.doi.org/10.1021/acsami.0c18506
ACS Appl. Mater. Interfaces XXXX, XXX, XXX−XXX
ACS Applied Materials & Interfaces www.acsami.org Research Article

modulus along <111>, {011}, {112}, and {123} slip systems be associated with elastic instabilities of the beta phase in these
will tend to zero.54 Also, a low elastic modulus is often configurations. According to our methodology, alloys with
associated with the reduction of C44, which leads to a low shear 22Zr and Nb additions higher than 14.8 (at. %) are the most
modulus. promising candidates among ternary Ti-Nb-Zr alloys for
The difference between the elastic anisotropy for the biomedical applications. This work suggests that there are
selected alloys and the binaries Ti-Zr22 and Nb-Zr22 is sweet spots in the compositional map, which could be further
depicted in Figure 9. Two-dimensional projections were explored experimentally.
obtained using the Marmier et al.55 representation via the
elastic tensor analysis (ELATE) suite, developed by Gaillac et
al.56 Furthermore, our results indicate that an excess of Nb

*
ASSOCIATED CONTENT
sı Supporting Information
atoms in the solid solution might enhance the stiffness of the
The Supporting Information is available free of charge at
bonds, increasing the resultant elastic constants of the beta
https://pubs.acs.org/doi/10.1021/acsami.0c18506.
phase (Nb-22Zr). This is evidenced in Figure 9 while
comparing the elastic modulus (E) and the shear modulus Summary of the elastic properties in the starting
(G) from the right-hand column (Figure 9c,f) with the others. database, after screening; properties of nine different
In summary, both of the ML algorithms predict a relatively supercells after relaxation obtained via DFT; Poisson’s
extended compositional field in which low-elastic modulus ratio distributions assumed for the machine learning
compositions can be found. There are attractive candidates models; predictions without assuming a Poisson ratio;
more centered in the diagram (where the TNZT alloys would architecture of two neural networks after random
be) and also low-cost compositions, close to the binary Ti- optimization (PDF)


22Zr (at. %), which have a relatively low electron-per-atom
ratio. In fact, stability assessments here presented show that the
beta phase in Ti-Zr binary alloys should exhibit an elastic AUTHOR INFORMATION
modulus smaller than in the ternary system (Table 3); Corresponding Author
however, omega could still be formed and compromise the Caetano R. Miranda − Instituto de Física, DFMT,
elastic modulus obtained in final products, after cooling. Universidade de São Paulo, 05315-970 São Paulo, SP,
The divergence between the ML predictions and DFT Brazil; orcid.org/0000-0002-8008-4907;
calculations regarding the Young modulus (E) was within the Email: cmiranda@if.usp.br
expected range. For example, for Ti-22Zr-14.8Nb (at. %), the
ML models predicted a minimum E between 67 and 82 GPa, Authors
while the DFT results indicated smaller values, as low as 30 Camilo A. F. Salvador − Instituto de Física, DFMT,
GPa. It is worth reminding that via DFT, the elastic modulus is Universidade de São Paulo, 05315-970 São Paulo, SP,
underestimated since it was assessed for the beta phase only. Brazil
Bruno F. Zornio − Instituto de Física, DFMT, Universidade
5. CONCLUSION de São Paulo, 05315-970 São Paulo, SP, Brazil;
orcid.org/0000-0002-9464-2738
Despite the previous record of experimental data addressing
bonelike modulus alloys, this work aimed to obtain generalized Complete contact information is available at:
models of the elastic properties of Ti alloys, to then find novel, https://pubs.acs.org/10.1021/acsami.0c18506
prospective candidates for biomedical applications in the Ti-
Nb-Zr system. Author Contributions
First, we could state that the current elasticity data available C.A.F.S. and C.R.M. conceived this research. C.A.F.S. designed
at the Materials Project allows obtaining good models to the machine learning models and SQS structures. DFT
describe the elastic properties of Ti-Nb-Zr alloys. The calculations and data analyses were performed by C.A.F.S.
implementation of packages such as pymatgen and matminer and B.F.Z. C.R.M. supervised the study. The manuscript was
was successful, given their speed and good reproducibility written through the contributions of all authors. All authors
generating the features to feed the models. Both random forest have approved the final version of the manuscript.
and neural network architectures were able to model the elastic Notes
properties. Their predictions regarding the elastic modulus The authors declare no competing financial interest.


converged to some areas of the compositional space.
Although there are quantitative differences between DFT ACKNOWLEDGMENTS
and experimental values regarding elastic constants, these
differences are systematic; thus, the models provide an effective The authors gratefully acknowledge the financial support from
method for showing trends in elastic behavior as a function of the Brazilian National Council for Scientific and Technological
composition. However, the absolute values of elastic constants Development (CNPq) through grants 155650/2018-1,
reported here must be used with care. 308333/2015-2, and 307064/2019-0 and the São Paulo
Within the first-principles methodology, DFT calculations Research Foundation (FAPESP, 17/02317-2). We would like
provided relevant information on the elastic properties of the to thank Logan Ward (Argonne National Laboratory) and Qi
selected alloys. Based on DFT results, the Poisson ratio range Wang (Lawrence Berkeley National Laboratory) for the
assumed for the ML predictions (between 0.25−0.45) was support regarding the matminer package.
accurate for this system.
In cases such as Ti-22Zr-3.7Nb and Ti-22Zr-7.4Nb, despite
the low elastic modulus estimations via ML, DFT results point
■ REFERENCES
(1) Kolli, R.; Devaraj, A. A Review of Metastable Beta Titanium
to an omega-phase formation at low temperatures. This might Alloys. Metals (Basel). 2018, 8, 506.

J https://dx.doi.org/10.1021/acsami.0c18506
ACS Appl. Mater. Interfaces XXXX, XXX, XXX−XXX
ACS Applied Materials & Interfaces www.acsami.org Research Article

(2) Huda, Z.; Edi, P. Materials Selection in Design of Structures and (22) Van De Walle, A.; Tiwary, P.; De Jong, M.; Olmsted, D. L.;
Engines of Supersonic Aircrafts: A Review. Mater. Des. 2013, 46, Asta, M.; Dick, A.; Shin, D.; Wang, Y.; Chen, L. Q.; Liu, Z. K.
552−560. Efficient Stochastic Generation of Special Quasirandom Structures.
(3) ANSI/NACE MR0175/ISO 15156. Petroleum and Natural Gas Calphad 2013, 42, 13−18.
Industries, Materials for Use in H2S Containing Environments in Oil and (23) Zunger, A.; Wei, S.-H.; Ferreira, L. G.; Bernard, J. E. Special
Gas Production; NACE International: Houston 2002. Quasirandom Structures. Phys. Rev. Lett. 1990, 65, 353−356.
(4) Abdel-Hady, M.; Hinoshita, K.; Morinaga, M. General Approach (24) von Pezold, J.; Dick, A.; Friák, M.; Neugebauer, J. Generation
to Phase Stability and Elastic Properties of β-Type Ti-Alloys Using and Performance of Special Quasirandom Structures for Studying the
Electronic Parameters. Scr. Mater. 2006, 55, 477−480. Elastic Properties of Random Alloys: Application to Al-Ti. Phys. Rev.
(5) Senkov, O. N.; Miracle, D. B. A New Thermodynamic Parameter B 2010, 81, 094203.
to Predict Formation of Solid Solution or Intermetallic Phases in High (25) Chen, K.; Zhao, L. R.; Tse, J. S. Ab Initio Study of Elastic
Entropy Alloys. J. Alloys Compd. 2016, 658, 603−607. Properties of Ir and Ir3X Compounds. J. Appl. Phys. 2003, 93, 2414−
(6) Senkov, O. N.; Jensen, J. K.; Pilchak, A. L.; Miracle, D. B.; Fraser, 2417.
H. L. Compositional Variation Effects on the Microstructure and (26) Mansouri Tehrani, A.; Oliynyk, A. O.; Parry, M.; Rizvi, Z.;
Properties of a Refractory High-Entropy Superalloy AlMo 0.5 NbTa Couper, S.; Lin, F.; Miyagi, L.; Sparks, T. D.; Brgoch, J. Machine
0.5 TiZr. Mater. Des. 2018, 139, 498−511. Learning Directed Search for Ultraincompressible, Superhard
(7) Materials Genome Initiative for Global Competitiveness. 2011. Materials. J. Am. Chem. Soc. 2018, 140, 9844−9853.
https://doi.org/https:/www.mgi.gov/. (27) Miedema, A. R.; de Châtel, P. F.; de Boer, F. R. COHESION
(8) Jain, A.; Ong, S. P.; Hautier, G.; Chen, W.; Richards, W. D.; IN ALLOYS - FUNDAMENTALS OF A SEMI-EMPIRICAL
Dacek, S.; Cholia, S.; Gunter, D.; Skinner, D.; Ceder, G.; Persson, K. MODEL. Physica B+C 1980, 1−28.
A. Commentary: The Materials Project: A Materials Genome (28) Yang, X.; Zhang, Y. Prediction of High-Entropy Stabilized
Approach to Accelerating Materials Innovation. APL Mater. 2013, Solid-Solution in Multi-Component Alloys. Mater. Chem. Phys. 2012,
1, No. 011002. 132, 233−238.
(9) de Jong, M.; Chen, W.; Notestine, R.; Persson, K.; Ceder, G.; (29) Liu, Y.; Zhao, T.; Ju, W.; Shi, S. Materials Discovery and Design
Jain, A.; Asta, M.; Gamst, A. A Statistical Learning Framework for Using Machine Learning. J. Mater. 2017, 3, 159−177.
Materials Science: Application to Elastic Moduli of k-Nary Inorganic (30) Xue, D.; Balachandran, P. V.; Hogden, J.; Theiler, J.; Xue, D.;
Polycrystalline Compounds. Sci. Rep. 2016, 6, 34256. Lookman, T. Accelerated Search for Materials with Targeted
(10) Winter, I. S.; de Jong, M.; Asta, M.; Chrzan, D. C. Computing Properties by Adaptive Design. Nat. Commun. 2016, 7, 11241.
Elastic Anisotropy to Discover Gum-Metal-like Structural Alloys. (31) Deng, B. Machine Learning on Density and Elastic Property of
Phys. Rev. Mater. 2017, 1, No. 030601. Oxide Glasses Driven by Large Dataset. J. Non-Cryst. Solids 2020, 529,
(11) Van Rietbergen, B.; Huiskes, R.; Weinans, H.; Sumner, D. R.;
119768.
Turner, T. M.; Galante, J. O. The Mechanism of Bone Remodeling
(32) Kim, G.; Diao, H.; Lee, C.; Samaei, A. T.; Phan, T.; De Jong,
and Resorption around Press-Fitted THA Stems. J. Biomech. 1993, 26,
M.; An, K.; Ma, D.; Liaw, P. K.; Chen, W. First-Principles and
369−382.
Machine Learning Predictions of Elasticity in Severely Lattice-
(12) Gepreel, M. A.-H.; Niinomi, M. Biocompatibility of Ti-Alloys
Distorted High-Entropy Alloys with Experimental Validation. Acta
for Long-Term Implantation. J. Mech. Behav. Biomed. Mater. 2013, 20,
Mater. 2019, 181, 124−138.
407−415.
(33) Ahneman, D. T.; Estrada, J. G.; Lin, S.; Dreher, S. D.; Doyle, A.
(13) Lima, D. D.; Mantri, S. A.; Mikler, C. V.; Contieri, R.; Yannetta,
C. J.; Campo, K. N.; Lopes, E. S.; Styles, M. J.; Borkar, T.; Caram, R.; G. Predicting Reaction Performance in C−N Cross-Coupling Using
Banerjee, R. Laser Additive Processing of a Functionally Graded Machine Learning. Science 2018, 360, 186−190.
Internal Fracture Fixation Plate. Mater. Des. 2017, 130, 8−15. (34) Furmanchuk, A.; Agrawal, A.; Choudhary, A. Predictive
(14) Dal Bó, M. R.; Salvador, C. A. F.; Mello, M. G.; Lima, D. D.; Analytics for Crystalline Materials: Bulk Modulus. RSC Adv. 2016,
Faria, G. A.; Ramirez, A. J.; Caram, R. The Effect of Zr and Sn 6, 95246−95251.
Additions on the Microstructure of Ti-Nb-Fe Gum Metals with High (35) Brun, F.; Yoshida, T.; Robson, J. D.; Narayan, V.; Bhadeshia, H.
Elastic Admissible Strain. Mater. Des. 2018, 160, 1186−1195. K. D. H.; MacKay, D. J. C. Theoretical Design of Ferritic Creep
(15) Ward, L.; Dunn, A.; Faghaninia, A.; Zimmermann, N. E. R.; Resistant Steels Using Neural Network, Kinetic, and Thermodynamic
Bajaj, S.; Wang, Q.; Montoya, J.; Chen, J.; Bystrom, K.; Dylla, M.; Models. Mater. Sci. Technol. 2013, 15, 547−554.
Chard, K.; Asta, M.; Persson, K. A.; Snyder, G. J.; Foster, I.; Jain, A. (36) Wu, C.-Y.; Chang, H.-T.; Wu, C.-Y.; Chen, S.-W.; Huang, S.-Y.;
Matminer: An Open Source Toolkit for Materials Data Mining. Huang, M.; Pan, Y.-T.; Bradbury, P.; Chou, J.; Yen, H.-W. Machine
Comput. Mater. Sci. 2018, 152, 60−69. Learning Recommends Affordable New Ti Alloy with Bone-like
(16) Ong, S. P.; Richards, W. D.; Jain, A.; Hautier, G.; Kocher, M.; Modulus. Mater. Today 2019, 34, 1.
Cholia, S.; Gunter, D.; Chevrier, V. L.; Persson, K. A.; Ceder, G. (37) Ye, Y. F.; Wang, Q.; Lu, J.; Liu, C. T.; Yang, Y. High-Entropy
Python Materials Genomics (Pymatgen): A Robust, Open-Source Alloy: Challenges and Prospects. Mater. Today 2016, 19, 349−362.
Python Library for Materials Analysis. Comput. Mater. Sci. 2013, 68, (38) Heinrich, C.; Waas, A. M.; Wineman, A. S. Determination of
314−319. Material Properties Using Nanoindentation and Multiple Indenter
(17) Chollet, F.; others. Keras. GitHub 2015. https://github.com/ Tips. Int. J. Solids Struct. 2009, 46, 364−376.
fchollet/keras (39) Shiraishi, T.; Yubuta, K.; Shishido, T.; Shinozaki, N. Elastic
(18) Kozeschnik E, B. B. MatCalc-a Simulation Tool for Multi- Properties of As-Solidified Ti-Zr Binary Alloys for Biomedical
component Thermodynamics, Diffusion and Phase Transformation Applications. Mater. Trans. 2016, 57, 1986−1992.
Kinetics. Math. Model. Weld Phenom. 5 2001, Book 738, 349−361. (40) Biesiekierski, A.; Ping, D.; Li, Y.; Lin, J.; Munir, K. S.; Yamabe-
(19) Kresse, G.; Hafner, J. Ab Initio Molecular Dynamics for Liquid Mitarai, Y.; Wen, C. Extraordinary High Strength Ti-Zr-Ta Alloys
Metals. Phys. Rev. B 1993, 47, 558−561. through Nanoscaled, Dual-Cubic Spinodal Reinforcement. Acta
(20) Blöchl, P. E. Projector Augmented-Wave Method. Phys. Rev. B Biomater. 2017, 53, 549−558.
1994, 50, 17953−17979. (41) Kim, K. M.; Kim, H. Y.; Miyazaki, S. Effect of Zr Content on
(21) Perdew, J. P.; Ruzsinszky, A.; Csonka, G. I.; Vydrov, O. A.; Phase Stability, Deformation Behavior, and Young’s Modulus in Ti−
Scuseria, G. E.; Constantin, L. A.; Zhou, X.; Burke, K. Restoring the Nb−Zr Alloys. Materials 2020, 476.
Density-Gradient Expansion for Exchange in Solids and Surfaces. (42) Banerjee, D.; Williams, J. C. Perspectives on Titanium Science
Phys. Rev. Lett. 2008, 100, 136406. and Technology. Acta Mater. 2013, 61, 844−879.

K https://dx.doi.org/10.1021/acsami.0c18506
ACS Appl. Mater. Interfaces XXXX, XXX, XXX−XXX
ACS Applied Materials & Interfaces www.acsami.org Research Article

(43) Fontaine, D. D. Mechanical Instabilities in the B.c.c. Lattice and


the Beta to Omega Phase Transformation. Acta Metall. 1970, 18,
275−279.
(44) Geetha, M.; Singh, A. K.; Asokamani, R.; Gogia, A. K. Ti Based
Biomaterials, the Ultimate Choice for Orthopaedic Implants − A
Review. Prog. Mater. Sci. 2009, 54, 397−425.
(45) Salvador, C. A. F.; Opini, V. C.; Lopes, E. S. N.; Caram, R.
Microstructure Evolution of Ti−30Nb−(4Sn) Alloys during Classical
and Step-Quench Aging Heat Treatments. Mater. Sci. Technol. 2017,
33, 400−407.
(46) de Fontaine, D.; Kikuchi, R. Bragg-Williams and Other Models
of the Omega Phase Transformation. Acta Metall. 1974, 22, 1139.
(47) Cao, P.; Tian, F.; Wang, Y. Theoretical Investigation of β-ω
Phase Transformation Mechanism in Ti-Mo Alloys. Comput. Mater.
Sci. 2020, 171, 109277.
(48) Devaraj, A.; Nag, S.; Srinivasan, R.; Williams, R. E. A.; Banerjee,
S.; Banerjee, R.; Fraser, H. L. Experimental Evidence of Concurrent
Compositional and Structural Instabilities Leading to ω Precipitation
in Titanium−Molybdenum Alloys. Acta Mater. 2012, 60, 596−609.
(49) Salvador, C. A. F.; Lopes, E. S. N.; Ospina, C. A.; Caram, R.
Orthorhombic Martensite Formation upon Aging in a Ti-30Nb-4Sn
Alloy. Mater. Chem. Phys. 2016, 183, 238−246.
(50) Hu, Q. M.; Li, S. J.; Hao, Y. L.; Yang, R.; Johansson, B.; Vitos,
L. Phase Stability and Elastic Modulus of Ti Alloys Containing Nb,
Zr, and/or Sn from First-Principles Calculations. Appl. Phys. Lett.
2008, 93, 121902.
(51) Lopes, E. S. N.; Cremasco, A.; Afonso, C. R. M.; Caram, R.
Effects of Double Aging Heat Treatment on the Microstructure,
Vickers Hardness and Elastic Modulus of Ti−Nb Alloys. Mater.
Charact. 2011, 62, 673−680.
(52) Hao, Y. L.; Li, S. J.; Sun, S. Y.; Yang, R. Effect of Zr and Sn on
Young’s Modulus and Superelasticity of Ti−Nb-Based Alloys. Mater.
Sci. Eng., A 2006, 441, 112−118.
(53) Meng, M.; Yang, H.; Fan, X. G.; Yan, S. L.; Zhao, A. M.; Zhu, S.
On the Modeling of Diffusion-Controlled Growth of Primary Alpha in
Heat Treatment of Two-Phase Ti-Alloys. J. Alloys Compd. 2017, 691,
67−80.
(54) Saito, T.; Furuta, T.; Hwang, J.-H.; Kuramoto, S.; Nishino, K.;
Suzuki, N.; Chen, R.; Yamada, A.; Ito, K.; Seno, Y.; Nonaka, T.;
Ikehata, H.; Nagasako, N.; Iwamoto, C.; Ikuhara, Y.; Sakuma, T.
Multifunctional Alloys Obtained via a Dislocation-Free Plastic
Deformation Mechanism. Science 2003, 300, 464−467.
(55) Marmier, A.; Lethbridge, Z. A. D.; Walton, R. I.; Smith, C. W.;
Parker, S. C.; Evans, K. E. ElAM: A Computer Program for the
Analysis and Representation of Anisotropic Elastic Properties.
Comput. Phys. Commun. 2010, 181, 2102−2115.
(56) Gaillac, R.; Pullumbi, P.; Coudert, F. X. ELATE: An Open-
Source Online Application for Analysis and Visualization of Elastic
Tensors. J. Phys. Condens. Matter 2016, 28, 275201.

L https://dx.doi.org/10.1021/acsami.0c18506
ACS Appl. Mater. Interfaces XXXX, XXX, XXX−XXX

You might also like