Materials Today Communications 38 (2024) 108477

Contents lists available at ScienceDirect

Materials Today Communications

journal homepage: www.elsevier.com/locate/mtcomm

A combined EBSD and machine learning study of predicting deformation

twinning in BCC Fe81Ga19 alloy
Zongwen Ma a, Yuanxiang Zhang a, *, Yukun Xia a, b, Yuchen Wang a, R.D.K. Misra c, Feng Fang a,
Yang Wang a, Xiaoming Zhang a, *
a
The State Key Laboratory of Rolling and Automation (RAL), Northeastern university, Shenyang 110819, PR China
b
Center of Electron Microscopy, State Key Laboratory of Silicon and Advanced Semiconductor Materials, School of Materials Science and Engineering, Zhejiang
University, Hangzhou 310027, PR China
c
Department of Metallurgical, Materials and Biomedical Engineering, University of Texas at EL Paso, El Paso, TX 79968, USA

A R T I C L E I N F O A B S T R A C T

Keywords: Compared to face-centered cubic (FCC) metals, body-centered cubic (BCC) metals demonstrate more intricate
Fe-Ga alloy deformation behavior and microstructural characteristics because of the influence of intrinsic unstable stacking
Twin nucleation faults and defect structure. To investigate the unique phenomenon of deformation twinning in Fe81Ga19 alloy
EBSD
with BCC structure at moderate temperature and low strain rate, four twin nucleation models were constructed
Machine learning
using a combination of electron backscatter diffraction (EBSD) technique and machine learning. These models
are aimed to predict twin nucleation of grains and explore the relationship between twin nucleation and
microstructural attributes. The study revealed that amongst the 235 grains analyzed during in-situ compressive
deformation, 32 grains experienced twinning. Nine attributes influencing twin nucleation were selected and
ranked according to their importance. Grain area, average image quality, neighboring grain count, and the
Schmid factor of twinning systems exhibited significant influence on twin nucleation. Decision tree and tree
ensemble (XGBoost) achieved 94% and 96% accuracy, respectively, on the test set, showcasing robust gener­
alization capability. Due to the optimization of hyperparameters, support vector machine (SVM) and artificial
neural network (ANN) exhibited outstanding performance. The ANN model achieved an accuracy of 97% and an
F1 score of 0.91 on the test set, while the SVM model achieved an accuracy of 98% and an F1 score of 0.94,
indicating superior performance. Furthermore, the study revealed that the twinning stress in Fe81Ga19 alloy
exhibited weak responsiveness to temperature during deformation, and the intrinsic factors of grains were
identified as crucial factors influencing twin nucleation.

1. Introduction deformation twinning [3] was observed during the rolling deformation
process (at 573 K and at a strain rate of 101 s− 1), which is an unusual
The Fe-Ga alloy, a new generation of magnetostrictive material, has phenomenon for body-centered cubic (BCC) metals.
wide-range of applications in various fields such as transducers, actua­ Plastic deformation in crystals is predominantly accomplished
tors, and sensors, owing to its vast potential [1]. However, due to the through dislocation glide and twinning [4]. The stacking fault energy of
intrinsic brittleness of the Fe-Ga alloy, the occurrence of cracking during a crystal determines the ease of dislocation cross-slip, thus influencing
the thin sheets manufacturing process is highly probable [2]. The rolling its plastic deformation. Materials with a high stacking fault energy are
technique employed for the production of Fe-Ga alloy sheets faces lim­ prone to dislocation glide, while those with a low stacking fault energy
itations, significantly impeding the progress and utilization of Fe-Ga tend to form incomplete dislocations, making dislocation glide difficult.
alloy. In our previous investigation, we successfully fabricated However, dislocation glide is considered the primary deformation
Fe81Ga19 alloy using a twin-roll strip casting method, and it was mechanism in BCC metals because of its higher stacking fault energy and
observed that the deformation mechanism followed a sequential pattern abundant slip systems. The formation of deformation twinning typically
of slip-twin-shear during deformation. Intriguingly, the occurrence of occurs under conditions of low temperature (≤ 173 K) and high strain

* Corresponding authors.
E-mail addresses: [email protected] (Y. Zhang), [email protected] (X. Zhang).

https://doi.org/10.1016/j.mtcomm.2024.108477
Received 20 December 2023; Received in revised form 11 February 2024; Accepted 23 February 2024
Available online 25 February 2024
2352-4928/© 2024 Elsevier Ltd. All rights reserved.
Z. Ma et al. Materials Today Communications 38 (2024) 108477

rates (≥ 102 s− 1) [5–7]. circumvent intricate and uncertain deformation mechanisms, thereby
Compared to face-centered cubic (FCC) metals, BCC metals exhibit offering a new perspective for exploring the relationship between
more intricate deformation behavior and microstructural features microstructural attributes and twinning nucleation. Essentially, the
because of the influence of intrinsic crystal symmetry and defect struc­ initiation of deformation twinning in Fe81Ga19 alloy through
ture. Specifically, this includes: 1) BCC metals possess three slip systems, medium-temperature rolling can be seen as a binary classification issue
namely {110} <111>, {112} <111>, and {123} <111>, with slip rooted in grain properties. Using ample grain data and the imple­
activation influenced by factors such as temperature, stress state, and mentation of diverse classification algorithms, a robust correlation is
loading rate. 2) The deformation behavior in BCC metals deviates from established between the microstructural attributes extracted from the
Schmid’s law, primarily due to the core structure of screw dislocations. material’s EBSD data and the occurrence of twin nucleation instances.
3) Propensity of screw dislocations to expand around their core. 4) BCC This augments the comprehension of the interconnection between the
metals display low-temperature brittleness. 5) Unique softening microstructural attributes and twin nucleation in Fe81Ga19 alloy.
behavior of solid solution The aforementioned distinct deformation In this study, we present a strategy that integrates in-situ EBSD with
behavior in BCC metals can be largely attributed to the nonplanar core four machine learning algorithms to enhance understanding of the
structure of screw dislocations. However, the early deformation twin­ correlation between microstructural attributions and twin nucleation in
ning theories, which primarily relied on dislocation decomposition Fe81Ga19 alloy with BCC structure. The workflow of this investigation is
models in FCC metals, yielded inconsistencies when applied to under­ depicted in Fig. 1. Information derived from the in-situ EBSD data serves
standing deformation twinning in BCC metals. as input for machine learning, with the occurrence or nonoccurrence of
The studies on deformation twinning in BCC metals at room tem­ twin nucleation designated as the target label. We employ artificial
perature and low strain rates (< 102 s− 1) have mainly focused on neural networks (ANN) and support vector machines (SVM), two models
nanocrystals or nanowires [8–11] due to their larger surface area to considered pure black boxes, to formulate two classifiers for forecasting
volume ratio, resulting in limited deformation carriers. In contrast to twin nucleation events, independent of the decision framework under­
bulk materials, nanostructured crystals require elevated stress levels for lying the predictions. Furthermore, we establish two interpretable de­
plastic deformation. Even at room temperature and low strain rates (< cision tree (DT) and eXtreme Gradient Boosting ensemble tree
102 s− 1), such a high-stress environment is more conducive to promoting (XGBoost) models. The performance of the four different algorithms is
the occurrence of twinning phenomena. Nevertheless, the observation of compared, and the factors that influence twin nucleation in Fe-Ga alloys
twinning in deformed BCC bulk crystals at room temperature and low are ranked. The study seeks to enhance and expedite the investigation of
strain rates (< 102 s− 1) is relatively scarce. Intrinsic factors such as grain deformation twinning in the Fe81Ga19 alloy based on EBSD, with the
size, crystal orientation, alloying, and extrinsic factors including tem­ objective of unveiling the microstructural attributes behind the twin
perature, strain rate, loading direction, and pre-deformation collectively nucleation event. This research approach not only focuses on exploring
exert influence upon the nucleation of deformation twinning. In-situ twin nucleation but also offers the potential for studying other intricate
EBSD [12], numerical analysis [13], and finite element methods [14] phenomena observed via EBSD, thus opening up new pathways and
are commonly used to study deformation twinning, but these methods perspectives to expedite research on rare and complex occurrences. In
require extensive knowledge of materials science and involve lengthy addition, other factors that could impact twin nucleation are also
simulations. Various models, such as the pole mechanism [15], the considered to deepen the understanding and cognition of twinning
dislocation core decomposition model [16], the dislocation interaction deformation in the Fe-Ga alloy, extending beyond the outcomes derived
model [17], and the double cross-slip model [18], have been proposed to from machine learning.
explain twin nucleation and growth in BCC metals. However, these
models primarily focus on explaining deformation mechanisms and a 2. Materials and methods
comprehensive understanding of the microstructural properties that
affect the twinning nucleation of BCC metal is still insufficient. The experiment employed a Fe81Ga19 (wt%) alloy strip with a
Currently, there is no effective criterion to predict which grains will thickness of 2 mm, which was directly obtained by hot rolling using a
undergo twin nucleation during plastic deformation of BCC metal. And twin-roll caster. After removing the oxide layer, the strip was subjected
the special phenomenon of deformation twinning in Fe81Ga19 alloy with to a warm rolling process at 300 ◦ C with a reduction per pass of 0.1 mm.
BCC structure during medium-temperature (573 K) rolling deformation When a thickness of 1.8 mm was reached, samples of dimensions 12 mm
has not been well explained. Therefore, employing machine learning in × 8 mm were cut from the warm-rolled strip. The samples were then
the investigation of material microstructure provides us with a new polished and etched using a 5% HNO3 ethanol solution, followed by
approach to seeking patterns from the data itself. observation using an OLYMPUS-BX53M optical microscope. Subse­
The emergence of the big data era has led to the widespread utili­ quently, the samples were electrolytically polished using a 12.5% HClO4
zation of machine learning in various domains [19–21]. The amalgam­ ethanol solution at 25 V for 20 s, and scanning electron microscopy
ation of machine learning and EBSD presents a pragmatic approach for (SEM) examination was performed using a Zeiss Ultra55 field-emission
extracting information on the microstructure of alloys. Compared to scanning electron microscope equipped with an EBSD probe (accelera­
numerical analysis and finite element methods, the application of ma­ tion voltage: 20 kV, working distance: 14.5 mm). Transmission electron
chine learning techniques can significantly reduce the time required for microscopy (TEM) observations were performed on the FEI Tecnai G2
material analysis. Furthermore, machine learning has garnered consid­ F20 transmission electron microscope, operating at a voltage of 200 kV.
erable attention within the realm of material microstructure analysis. To investigate the twinning nucleation phenomenon in the Fe81Ga19
For HCP metals with fewer slip systems at room temperature, re­ alloy, we conducted in-situ compression experiments on an original cast
searchers such as Tong [22] and Gui [23] have utilized a variety of strip specimen using the Gatan MICROTEST 2000 dynamic in-situ ten­
machine learning algorithms to predict twin nucleation events by sile testing machine. The specimen underwent two compression tests
extracting information from EBSD data. Additionally, Orme et al. [24] along the rolling direction (RD), with a strain rate of 1×10− 3 s− 1 and
employed decision tree methods to capture the correlation between deformation temperature at room temperature. Initially, the specimen
microstructure and twin nucleation in AZ31magnesium. However, to was compressed to 1.3% strain, followed by further compression to 5%
date, there has been no investigation utilizing a combined approach of strain. After the experiment concluded, we selected the image with the
machine learning and EBSD to elucidate the relationship between highest identification rate for EBSD image analysis. The scanning area of
microstructural information and twinning nucleation in BCC metals. the sample was 330×320 μm2, with a step size of 1μm.
Considering the complexity of the mechanism of deformation twinning The EBSD data were post-processed using Aztec software to obtain
in BCC metals, the advantage of machine learning lies in its ability to initial grain information before and after deformation. This information

2
Z. Ma et al. Materials Today Communications 38 (2024) 108477

Fig. 1. Research flowchart for the twin nucleation event in the Fe-Ga alloy.

includes: grain area, fitted ellipse angle, fitted ellipse sspect ratio, mean ( )
cov(Xi , Y)
grain orientation spread, mean band contrast, MAD, neighbor grain ρ Xi , Y = (2)
count, TwSF, and SlSF. The extracted attributes, along with their brief σ(Xi )σ (Y)
descriptions, are listed in Table 1. Among these attributes, the SlSF Among them, cov represents covariance, and σ represents standard
(maximum Schmid factor of slip systems) is calculated using the Schmid deviation.
law (Eq. (1)), while twinning is considered a result of partial dislocation Due to the lack of various local attribute information near the edges
slip [25–27]. The TwSF (maximum Schmid factor of twinning systems) of the scanning area, these grains are excluded from the scope of sta­
is determined within the twinning system {112} <111>. tistical data. The total number of datasets used in this study is 235,
SF = cos λ cos ϕ (1) extracted from images of a single specimen. During training, a hybrid
hyperparameter optimization approach incorporating manual tuning
Among these, λ denotes the inclination between the loading axis and and grid parameter search was utilized. To evaluate the performance of
the slip direction, while ϕ signifies the inclination between the loading the model, a ten-fold cross-validation strategy was adopted. Specifically,
axis and the normal vector of the slip plane. Furthermore, through we divided the training set into 10 equally sized subsets, with 9 subsets
observation of grain structures before and after in-situ compression used as training data (totaling 157 for training dataset) and the
deformation, we identified all grains exhibiting deformation twins remaining 1 subset used as validation data (totaling 78 for the validation
within the scanning area, categorizing them into two groups: "twin" and dataset). Cross-validation was applied to predict twin nucleation for
"no twin", labeled as "1" and "0" respectively. both the training and test set, enabling comparison of algorithm per­
To enhance the precision of attribute analysis regarding twin crystal formance in terms of accuracy and F1 score. To address the issue of data
nucleation, the Pearson correlation coefficient ρ(Xi, Y) [28] was calcu­ bias caused by class imbalance, a strategy was employed. Specifically,
lated to evaluate the influence of each attribute Xi on the label of the we divided the majority class (i.e., the class with more samples) into
class Y, as shown below: multiple subsets, then randomly selected one subset to combine with the
minority class (i.e., the class with fewer samples) to construct a balanced
training set. This approach effectively alleviates the problem of class
Table 1 imbalance and enhances the performance and generalization ability of
Possible influential attributes on the twin nucleation of Fe-Ga alloy and their
the model.
description.
Attribute Description 3. Machine learning algorithms and performance metrics
A Area of grain
Fitted Ellipse Angle/EA The angle of the major axis of the fitted object’s Utilizing the scikit-learn library in Python [29], four models were
ellipse
constructed to predict twin nucleation in the Fe-Ga alloy. The models
Fitted Ellipse Aspect Ratio/AR The aspect ratio of the fitting ellipse of grain
Mean Grain Orientation Spread/ The average value of orientation dispersion of
employed include decision tree, XGBoost ensemble tree, artificial neural
GOS each grain network, and support vector machine.
Mean Band Contrast/BC The average band contrast value of the grain
MAD Average angle deviation value of grains
Neighbor Grain Count/NG The number of adjacent grains around the grain
TwSF The maximum Schmid factor of the twin system
SlSF The maximum Schmid factor of the slip system

3
Z. Ma et al. Materials Today Communications 38 (2024) 108477

3.1. Machine learning algorithms used for classification tasks score [35] normalization as described in Eq. (5).
f (x) = x, x > 0; f (x) = 0, x ≤ 0 (3)
3.1.1. Decision tree
The decision tree is an extensively utilized and robust machine
f (x) = 1/(1 + e− x ) (4)
learning algorithm, particularly suitable to identify trends. They make
prediction and decision by constructing a tree-like structure. During the Z = (x − μ)/σ (5)
prediction process, decision trees employ a "divide-and-conquer" strat­
egy, traversing the tree based on feature values. The dataset is parti­ Where Z symbolizes the normalized data, x represents the raw data
tioned into distinct branches, with each node representing a feature and within the dataset, while μ and σ respectively indicate the mean and
each leaf node representing a class. The decision tree model is not fully standard deviation of the data in attribute x.
opaque, by visualizing the tree, the classification criteria and paths can
be elucidated [22,24]. However, decision tree models are susceptible to 3.1.4. Support vector machine
overfitting, necessitating pruning to enhance their generalization The Support Vector Machine [36] is a prevalent and robust machine
prowess. In the present study, the CART (Classification and Regression learning algorithm employed for identifying the optimal hyperplane or
Tree) algorithm from the scikit-learn library was utilized. The Gini co­ decision boundary to achieve maximum-margin classification. SVM
efficient was employed for attribute splitting [30] to model twin exhibits its efficacy not only on linearly separable data, but also by
nucleation event in Fe81Ga19 alloy. The following parameter combina­ utilizing kernel functions to map samples onto high-dimensional feature
tions are used for pruning processing: criterion = "gini", splitter = "best", spaces, thereby facilitating linear separability in those domains. SVM
max_depth = 4, min_samples_leaf =1, min_samples_split =2. relies solely on the positions of support vectors for establishing the de­
cision boundary, rather than the entire dataset, rendering it highly
3.1.2. XGBoost ensemble trees adaptable to handling large sample sizes and high-dimensional data.
Ensemble learning, through the aggregation of multiple learners, can SVM demonstrates proficiency in the management of high-
attain superior generalization performance compared to a solitary dimensional data and in addressing non-linear problems. Nonetheless,
learner, thus mitigating overfitting. The XGBoost algorithm represents the efficacy of SVM’s generalization hinges upon the adeptness in
an enhanced and refined iteration of the gradient boosting tree algo­ selecting appropriate hyperparameters, particularly the Kernel function.
rithm, devised by Chen et al. [31]. This algorithm embraces the concept Furthermore, it is imperative to acknowledge that SVM exhibits sensi­
of gradient boosting, in which numerous weak learners are constructed tivity toward feature scaling. Therefore, undertaking normalization
in each iteration, with each learner trained based on the outcomes of the preprocessing during SVM model training becomes indispensable. This
preceding iteration. Moreover, XGBoost incorporates regularization refinement increases the model’s performance and stability, thereby
terms to govern the complexity of the model, guarding against over­ fortifying its resilience and advancing convergence speed and classifi­
fitting, while upholding exceptional robustness. An XGBoost model was cation accuracy.
constructed using the given hyperparameters: gamma = 0.001, lear­ In this study, a predictive model for twin nucleation events in
ning_rate = 0.2, max_depth = 4, n_estimators = 123. And the relative Fe81Ga19 alloy was developed using the SVM package from the scikit-
importance of each attribute is calculated using the built-in function learn library. Before training, the dataset was standardized with Z
’xgboost.plot_importance’ of the XGBoost model [32]. scores to ensure a distribution with zero mean and unit variance. Mul­
tiple models of kernel function, including linear, polynomial, sigmoid,
3.1.3. Artificial neural network and Gaussian Radial Basis Function (RBF), were constructed to obtain
The artificial neural network (ANN) [33,34] is a computational the best model. Through comprehensive performance evaluation, the
model that emulates the functioning of biological neural systems. It finds RBF was selected as the kernel function due to its exceptional perfor­
extensive application in the domains of machine learning and pattern mance in handling nonlinear problems. The adjustment of two key
recognition. The fundamental unit of an ANN is a neuron that facilitates hyperparameters (C and γ) in the SVM model of RBF kernel function is
the transmission of information between neurons through weighted detailed in Section 4.3.
connections. Each neuron receives a set of input signals and these signals
are subjected to weighting and summation using connection weights. 3.2. Evaluation metrics
The outcome is then transformed non-linearly using an activation
function. The transformed output can be propagated to subsequent The evaluation metrics employed in machine learning tasks serve to
neurons or used as the final output. ANN commonly employs a assess and compare the performance and efficacy of diverse models. For
feed-forward architecture, wherein neurons are organized in layers, binary classification tasks, the confusion matrix stands as a widely uti­
enabling the flow of information from the input layer, through hidden lized evaluation metric. Accuracy (Acc, Eq. (6), derived from the
layers, to the output layer. The artificial neural network possesses confusion matrix, quantifies the correct classification rate based on the
remarkable expressive capabilities and adaptability, enabling it to instance count for each class. Recall (R, Eq. (7)), on the other hand,
comprehend intricate non-linear relationships and achieve notable ad­ gauges the ratio of accurately classified positive samples among all
vancements in diverse domains. Furthermore, the determination of actual positive samples, while precision (P, Eq. (8)) quantifies the ratio
hyperparameters such as the number of nodes in hidden layers and the of correctly classified positive samples among all samples classified as
selection of activation functions plays a pivotal role in optimizing neural positive. Lastly, the F1 score (F1, Eq. (9)), the harmonic mean of pre­
network models. cision and recall, presents a more robust metric for contrasting different
The ANN model was constructed using the MLPClassifier package classifiers.
from the scikit-learn library. To determine the optimal number of nodes
in the hidden layers, a grid parameter optimization method was Acc = (fTP + fTN )/(fTP + fTP + fTP + fTP ) (6)
employed. A two-hidden-layer network was selected as the best archi­
R = fTP /(fTP + fFN ) (7)
tecture, with 6 and 14 nodes in the hidden layers, respectively, to avoid
overfitting the model. The "Adam" solver was used for weight optimi­
P = fTP /(fTP + fFP ) (8)
zation, and the Rectified Linear Unit (ReLU) function (Eq. (3)) was
chosen as the activation function for the hidden layers, while the Sig­ /(
1 1
)
moid function (Eq. (4)) was used as the activation function for the F1 = 2 + (9)
P R
output layer. Before model training, the dataset was pre-processed for Z-

4
Z. Ma et al. Materials Today Communications 38 (2024) 108477

In this study, "true positives" (TP) correspond to instances where the 4.2. Correlation between attributes and target label (indicating twin
model accurately predicts the occurrence of twins in grains following nucleation)
deformation. "False negatives" (FN) relate to cases where the model
predicts the absence of twins in grains, despite their actual occurrence. To enhance our comprehension of the correlation between the
On the other hand, "false positives" (FP) refer to situations where the possible attributes that affect twinning and twin nucleation, we gener­
model incorrectly predicts twinning in grains when it does not occur in ated stacked histograms for each attribute. Kindly refer to Fig. 5 for the
reality. Lastly, "true negatives" (TN) indicate cases where the model specific results.
correctly predicts the absence of twins in grains. The grain size holds substantial significance in influencing twin
deformation. Compared to smaller grains, larger grains are more sus­
4. Results and discussion ceptible to stress concentration and are prone to twinning during
deformation. As depicted in Fig. 4, deformation twins are typically
4.1. The microstructure of cast strip before and after deformation observed in grains of larger sizes, while their presence in smaller grains
is difficult to discern. This occurrence can be attributed to the coordi­
Fig. 2(a) shows the micrograph of the Fe81Ga19 alloy cast strip, which nated deformation among smaller grains during plastic deformation,
is predominantly composed of equiaxed grains. After being subjected to which facilitates adjustments in grain orientation and promotes dislo­
warm rolling at 300 ◦ C with a 10% thickness reduction, Fig. 2(b) shows cation slip. However, discerning the size favoring twin nucleation solely
the resulting microstructure. The formation of deformation bands within based on the assessment of attribute A (Area) in Fig. 5 poses a challenge.
multiple grains is evident. It is worth mentioning that these bands do not Based on the observations in Fig. 5, it can be inferred that certain
traverse the grain boundaries, as indicated by the black dashed box. This attributes exhibit an effective discriminative power in determining the
observation is consistent with our previous research [3]. The orientation formation of deformation twinning. For instance, when the AR attribute
relationship between the deformation band and the matrix of the sample of the grains exceeds 2.44, no twin nucleation occurs in any of the
subjected to a 10% reduction was investigated using TEM. Fig. 3(a) grains. Furthermore, by calculating the Schmid factor, we can roughly
displays the bright-field TEM micrograph of the sample, and Fig. 3(b) identify which grains are more favorable for twinning formation, as
exhibits the selected-area electron diffraction (SAED) pattern obtained higher values of the TwSF indicate a higher probability of twinning
along the [110] zone axis in the twinning region (highlighted by the occurrence [38,39]. However, the applicability of the Schmid law is not
white box in Fig. 3(a)). Through SAED analysis, a clear mirror symmetric always valid for non-close-packed BCC metals, mainly due to the unique
relationship with the twinning plane (-112) between the deformation nonplanar core structure of their screw dislocations [40].
twin and the matrix was observed, indicating that the twinning belongs When SlSF or TwSF of the grains in Fig. 5 is less than 0.39, nearly all
to the (-112) [1–11] twin system. grains are categorized as "no twin." However, when the SlSF or TwSF of
To approximate the rolling deformation process, an "H"-shaped the grains exceeds 0.39, not all grains exhibit twin nucleation. Although
specimen was cut from the Fe81Ga19 alloy strip and an in-situ the analysis in Fig. 5 offers an initial evaluation of twin nucleation in the
compression experiment was carried out. The image of the inverse deformed Fe81Ga19 alloy, it is not scientifically conclusive to rely solely
pole figure (IPF) of the specimen before deformation is shown in Fig. 4 on a single attribute to explain the occurrence of twin nucleation.
(a), which exhibits an equiaxed grain structure without any evidence of Currently, there is no definitive attribute established as a criterion for
twinning. Under a compression rate of 0.033 mm/min, the IPF image of twin nucleation. Hence, we employ an integrated approach that com­
the sample is obtained at 5.0% compression strain (Fig. 4(b)). Clearly, bines in-situ EBSD and four machine learning algorithms to compre­
numerous deformation twins are observed in the specimen after hensively consider various attributes that may influence twin
compression, which has previously been confirmed as {112}<111> nucleation. The objective is to achieve a more comprehensive under­
deformation twins formed by sequential shear along 1/6[111] on standing of the relationship between twin nucleation and microstruc­
adjacent {112} planes. Interestingly, the deformation twins can effec­ tural properties in Fe-Ga alloy during the mid-temperature rolling
tively connect neighboring grains, possibly due to the diffusion-driven process and to gain deeper insights.
twinning behavior between grains [37]. The Pearson correlation coefficient ρ indicates the linear correlation
After performing statistical analysis on the grains in the IPF map, a between the attribute Xi and the class label Y. We plotted the Pearson
dataset comprising 235 grains was chosen to investigate twin nucleation correlation heatmap between attributes, as depicted in Fig. 6. The range
phenomena in the Fe81Ga19 alloy with BCC structure. After applying a of ρ is − 1–1, where a larger absolute value signifies a stronger linear
compressive strain of 5%, twinning was observed in 32 grains. relationship between attributes. Based on the findings in Fig. 6, we
observed a substantial correlation between attributes A, GOS, and NG

Fig. 2. The microstructure of the Fe81Ga19 alloy. (a) Cast strip; (b) 10% Compression.

5
Z. Ma et al. Materials Today Communications 38 (2024) 108477

Fig. 3. (a) Bright-field TEM micrograph of the sample after a 10% reduction; (b) Corresponding selected-area electron diffraction pattern with a zone axis of [110].
The subscripts M and T indicate the matrix and deformation twin, respectively.

Fig. 4. IPF diagram of the (a) original cast strip; (b) compression ratio of the 5% specimen.

with twin nucleation. Notably, attribute A exhibits a correlation coef­ 0.84. However, for the test set, although the accuracy is high (89%), the
ficient of 0.42 with twin nucleation, aligning with our understanding of F1 score is relatively low (0.63), indicating the presence of overfitting.
twin nucleation events. Furthermore, the GOS attribute reflects the To address the challenge posed by class imbalance and mitigate the
mean orientation spread of the grains and aids in identifying deformed impact of overfitting, pruning the model becomes imperative.
grains in the principal strain analysis. We identified a strong correlation During the training phase of the decision tree model, adjusting the
between GOS and NG attributes with twin nucleation. These results ’max_depth’ hyperparameter provides control over the depth and in­
imply that while GOS and NG attributes may often be overlooked in our tricacy of the tree structure. Smaller values may result in insufficient
typical assessment of twin nucleation, they possess significance in fitting, whereas larger values may lead to excessive fitting. To determine
uncovering such events. Although these correlation coefficients slightly the appropriate value of "max_depth", an evaluation was conducted to
exceed 0.35, they indicate a certain degree of the nonlinear relationship examine its impact on cross-validation accuracy and F1 score. Fig. 8
between twin nucleation events and attributes. A single attribute is shows the evaluation results, where it is evident that the accuracy and F1
insufficient to predict the nucleation of deformation twins in the alloy. score reach a high level when the "max_depth" value is set to 4. As
Thus, it becomes imperative to predict twin nucleation events in "max_depth" continues to increase, the F1 score shows a marginal
Fe81Ga19 alloy with BCC structure employing machine learning algo­ improvement, while the accuracy experiences a notable decline.
rithms. Subsequently, we will present the training and evaluation out­ Therefore, in this study, we opted to set ’max_depth’ to 4 as the most
comes of various models. suitable value.
The pruned decision tree model is illustrated in Fig. 9. In the con­
structed decision tree model, the root node represents attribute A (Area).
4.3. Machine learning algorithms for predicting twin nucleation events
Two distinct paths are formed from the root node. The first path en­
compasses the relationship between attributes A, MAD, TwSF, and BC.
4.3.1. The CART decision tree model
MAD denotes the average angle deviation fitting value used to assess the
Fig. 7 illustrates the decision tree model constructed on the training
match between simulated EBSP and actual positions, with a smaller
set, with a maximum depth of 7 and without pruning. The model ach­
MAD indicating better alignment. TwSF signifies the maximum value of
ieved a training set prediction accuracy of 99% and a high F1 score of

6
Z. Ma et al. Materials Today Communications 38 (2024) 108477

Fig. 5. Stacked histogram of the quantity distribution for "twin" and "no twin" occurrences in each attribute.

7
Z. Ma et al. Materials Today Communications 38 (2024) 108477

Fig. 6. The Pearson correlation heat map between different attributes.

the Schmid factor in the twin system {112} <111>, with a higher TwSF among attributes, thus playing a more substantial role in the modeling
implying a lower stress requirement for twin nucleation. Specifically, and prediction of nonlinear problems.
when A ≤ 7640, the model considers the MAD attribute. Along this path, The confusion matrix for the decision tree model is provided in
deformation twinning occurs only when MAD > 0.955, whereas twin­ Table 2. The model attained a training set precision of 99% and an F1
ning does not occur when MAD ≤ 0.955 and TwSF ≤ 0.485. However, score of 0.97. In the case of the test set, the accuracy and F1 score stood
when TwSF ≥ 0.495 and BC > 146, the model identifies twin nucleation at 96% and 0.84, respectively. It is evident that the pruned model, in
events within the grains. The other path involves attributes A, NG, and comparison to the unpruned decision tree model, not only mitigates the
BC. When 7640 ≤ A ≤ 8724 and NG > 6, or when 8724 ≤ A ≤ 28098 and risk of overfitting but also enhances accuracy and F1 score on both the
BC ≤ 139, the model predicts the formation of deformation twins within training and test sets.
the grains. These paths and decision rules comprise the decision tree
model, offering valuable insights into twin nucleation in grains of the 4.3.2. XGBoost
BCC structure Fe81Ga19 alloy. In our previous study, it was observed that A total of 123 CART trees were employed to predict the twinning
grains exhibiting deformation twinning were larger and had TwSF pri­ event, and the schematic diagram of the XGBoost model is shown in
marily concentrated around 0.46 [3]. However, the decision tree model Fig. 11. Fig. 12 shows one of the trees in the ensemble, with the leaf
integrates attributes A, TwSF, NG, BC, and MAD, providing more spe­ nodes displaying absolute values representing the confidence of the
cific criteria. This decision tree-based model presents an exploratory prediction. Larger values indicate a higher likelihood of "twin". The final
approach to gain a deeper understanding of the relationship between prediction is determined by aggregating the predictions from all the
grain attributes and twin nucleation. decision trees. The confusion matrix for the XGBoost model is provided
Additionally, we conducted an attribute of importance to analysis in in Table 2. On the training set, accuracy was achieved 100% with an F1
the model, and the specific outcomes are illustrated in Fig. 10. Among score of 0.99. On the test set, the accuracy stood at 96% with a relatively
them, attribute A exhibits a slightly higher importance score of over 0.4, high F1 score of 0.88. The Fe81Ga19 alloy twinning model developed
while attributes BC and NG demonstrate scores around 0.2. The score for using the XGBoost algorithm demonstrated superior performance
TwSF is 0.12, whereas the significance of MAD is comparatively low, compared to the decision tree model in terms of evaluation metrics.
merely 0.05. It is noteworthy that the exploration of BC, NG, and MAD Fig. 13 displays a graph representing the importance scores of the
attributes in twin nucleation studies has been limited. However, by attributes that influence twin nucleation. Among them, grain area (A)
employing the decision tree algorithm, we can broaden our compre­ stands out as the most prominent attribute, followed by BC, EA, MAD,
hension of the correlation between these attributes and twin nucleation, GOS, AR, TwSF, SlSF, and NG attributes. These findings offer valuable
enabling us to delve into novel insights. Furthermore, notable disparities insights into the contributions of attributes to twin nucleation.
were observed between the linear Pearson correlation coefficients In comparing the plots of attribute importance scores between de­
among attributes depicted in Fig. 6 and the importance levels assigned cision tree and XGBoost model, noticeable discrepancies emerge.
by the decision tree algorithm. For instance, attribute NG, which pos­ Nonetheless, irrespective of the model employed, the attribute A (Area)
sesses the highest correlation coefficient, only obtained an importance of the grains demonstrates significant importance in the scores. The
score of 0.17 in the decision tree, while attribute GOS, with a relatively results from the decision tree model in Fig. 9 reveal a substantial sup­
high correlation coefficient, appears to have no significant impact on the pression of twin nucleation within the small grains of the alloy. On the
decision tree algorithm. These results indicate that linear correlation other hand, the BC attribute assumes a secondary position in both
coefficients and the decision tree algorithm, which is more suitable for models, exhibiting a linear correlation coefficient of 0. Its magnitude
non-linear problems, represent two distinct evaluation criteria. The reflects the dislocation density of grains in their undeformed state [22]
decision tree algorithm showcases greater flexibility and adaptability, and it is often ignored in the assessment of twin nucleation. Further­
empowering it to capture nonlinear relationships and interactions more, Fig. 5 indicates that grains with lower BC values exhibit minimal

8
Z. Ma et al. Materials Today Communications 38 (2024) 108477

Fig. 7. The unpruned CART decision tree model.

twin nucleation events. This can be attributed to its inherently elevated XGBoost model (Fig. 13) was used to construct nine different ANN
initial dislocation density, enabling slip deformation to adequately models, with each model’s input dataset comprising the first n (n = 1 ~
accommodate compression and diminishing the propensity for twin 9) attributes. Performance scores of the nine models with varying
nucleation. Moreover, Fig. 13 demonstrates that attributes like EA, numbers of input attributes were compared on the training and testing
MAD, and GOS exhibit varying degrees of association with twin nucle­ sets, as illustrated in Fig. 15. The results demonstrate a significant
ation. Relying solely on individual attributes for judgment has limita­ improvement in model performance with an increasing number of input
tions in explaining the nucleation of twinning phenomena. Recognizing attributes. However, the model’s performance noticeably declines when
the interplay and intricate relationships among attributes is pivotal for there are only 3 input attributes. Subsequently, as the number of input
accurately determining the occurrence of twin nucleation event. attributes continues to increase, the model’s performance stabilizes and
reaches its peak when constructed on the dataset with n = 9 attributes.
4.3.3. Artificial neural network Table 2 showcases the confusion matrix of the model on the dataset,
The schematic diagram of the double hidden layer artificial neural achieving an accuracy of 97% and an F1 score of 0.91 on the test set.
network used in the study can be referred to as Fig. 14. Excessive noise Overall, in terms of predicting twin nucleation, the ANN model out­
attributes have the potential to diminish the accuracy of the ANN model. performs the decision tree and the XGBoost model, exhibiting superior
To enhance the precision and predictive performance of the model, a performance.
selection of the top-ranked attributes based on their importance in the

9
Z. Ma et al. Materials Today Communications 38 (2024) 108477

Fig. 8. Impact of the hyperparameter "max_depth" on cross-validation accuracy

and F1 score during training process.
Fig. 10. The importance scores of features in the decision tree.
4.3.4. Support vector machine
To obtain an SVM model with a highly generalizable RBF kernel
Table 2
function, it is imperative to fine-tune two pivotal hyperparameters: C
The confusion matrix values, Acc and F1 values for predicting twin nucleation
and γ [41]. C represents the penalty parameter, where excessively events on both the training and test sets using various machine learning
elevated values can engender overfitting, while lower values may lead to algorithms.
underfitting. The γ parameter determines the influence of individual
algorithms The values of the The values of the Acc and F1 Acc and
training samples on the decision boundary. A smaller γ value signifies a confusion matrix confusion matrix value of the F1 value
broader impact range of the decision boundary and greater inter-sample (TP, FP, FN, TN) in (TP, FP, FN, TN) training set of the
influence, thereby yielding a complex or overfitted decision boundary. the training set in the test set test set
Conversely, larger γ values yield the opposite effect. Fig. 16 elucidates Decision 142, 0, 3, 19 59, 2, 0, 10 99%/0.97 94%/
the validation curve of the SVM model with the RBF kernel function Tree 0.84
across varying C and γ values. The subplots (a) (b) and (c) (d) manifest XGBoost 138, 0, 0, 19 64, 1, 2, 11 100%/1 96%/
0.88
the impact of different γ and C on accuracy and F1 score of the model
ANN 134, 0, 0, 23 68, 1, 0, 9 100%/1 97%/
through 10-fold cross-validation. The graph unequivocally demonstrates 0.91
that the model achieves optimal accuracy and F1 score when γ is set to 1 SVM 135, 0, 0, 22 66, 2,0,10 100%/1 98%/
and C is set to 10. 0.94

Fig. 9. The pruned decision tree model.

10
Z. Ma et al. Materials Today Communications 38 (2024) 108477

Fig. 11. Schematic diagrams of the XGBoost model.

Fig. 12. One CART tree within the ensemble of 123 trees in the XGBoost model.

Fig. 13. Importance scores of the nine attributes for evaluating the
XGBoost model.
Fig. 14. Diagram of the artificial neural network.

The hyperparameters C and γ underwent optimization using a grid

of the SVM model on both the training and test set, the confusion matrix
parameter search procedure, spanning from 10− 4 to approximately 103.
is displayed in Table 2. The test set yielded a remarkable accuracy of
According to the heatmap depicted in Fig. 17, the model demonstrated
98%, accompanied by an impressive F1 score of 0.94. This SVM model
its optimal performance at γ = 1 and C = 10. Assessing the performance
outperformed the remaining three learning algorithms, likely benefiting

11
Z. Ma et al. Materials Today Communications 38 (2024) 108477

Furthermore, the F1 scores of all four models exhibited gradual

improvement, demonstrating commendable generalization capability.
Within this quartet of models, the SVM and ANN models benefit from
fine-tuning their hyperparameters, demonstrating outstanding perfor­
mance not only on the training set but also on the test set, exhibiting
strong generalization abilities. Both models achieve an F1 score
exceeding 0.9. The ANN model, with two hidden layers, constructed
based on the dataset containing all attributes and optimized using grid
search, achieves an accuracy rate of 97% and an F1 score of 0.91 on the
test set, showcasing its robust generalization capability. The SVM model
with the RBF kernel, optimized through parameter tuning, achieves an
accuracy rate of 98% and an F1 score of 0.94 on the test set, demon­
strating the highest predictive ability and strongest generalization per­
formance, making it the optimal model among the four algorithms. The
exceptional abilities of ANN and SVM in handling high-dimensional
nonlinear decision boundaries further accentuate the fact that the twin
Fig. 15. The influence of input attributes (n = 1,2,.,9) on the evaluation nucleation of grains is not simply a linear relationship but rather a
metrics of twin nucleation occurrences. nonlinear problem.

from its exceptional performance on datasets with limited size. 4.5. Discussion of factors influencing twin nucleation

Through the extraction of in-situ EBSD data from the Fe81Ga19 alloy,
4.4. Comparison of different models the interpretable decision tree and XGBoost model have been success­
fully established. The application of these models unveils essential in­
Fig. 18 shows a comparison of the accuracy and F1 scores of four sights into the microstructural attributes and twin nucleation. In the
machine learning algorithms utilized for the prediction of twin nucle­ decision tree model, the influential attributes are ranked in descending
ation event in Fe81Ga19 alloy. Notably, only the decision tree model fell order of importance as A, BC, NG, TwSF, and MAD. In the XGBoost
short of attaining flawless accuracy and F1 score on the training set, model, the attribute importance ranking is A, BC, EA, MAD, GOS, AR,
whereas the remaining three models exhibited impeccable outcomes. TwSF, SlSF, and NG. These findings provide valuable understanding
The decision tree, XGBoost, ANN, and SVM model showcased a rising regarding the relationship between microstructural attributes and twin
trend in accuracy on the test set, surpassing 95% in all cases. nucleation. Particularly noteworthy is the consistently high importance

Fig. 16. Validation curves for different parameters of the SVM model with RBF kernel. (a) and (b) represent the impact of γ on accuracy and F1 score, respectively. (c)
and (d) depict the influence of C on accuracy and F1 score, respectively.

12
Z. Ma et al. Materials Today Communications 38 (2024) 108477

Fig. 17. Heatmap results of grid parameter search for optimal values of C and γ. The combination of C = 0 and γ = 1 yielded the highest levels of (a) accuracy and (b)
F1 score.

model, the machine learning-based model not only improves the pre­
diction of twin nucleation but also offers an exploratory avenue for
comprehending the intricate relationship between microstructural
properties of grains and twin nucleation. Furthermore, these insights can
potentially be incorporated into crystal plasticity models to enhance
their predictive capabilities and interpretability.
Moreover, some inherent factors of the grains also play a pivotal role
in twin nucleation. Among them, the grain orientation assumes a deci­
sive function in the activation of deformation twinning in alloys [44–46]
Not all grain orientations induce twinning. Our previous investigation
has identified that in Fe81Ga19 alloy, the matrix orientations prone to
twinning predominantly cluster near {111} <112> [3]. The substantial
presence of Ga elements further reduces the stacking fault energy [38],
facilitating the decomposition of full dislocations into partial dislo­
cations, thus decreasing the likelihood of cross-slip and heightening the
propensity for deformation twinning. Alloying also exerts an influence
Fig. 18. Radar chart depicting the performance of four models on both the on deformation twinning. When alloying elements exist as substitution
training and test set. atoms in the BCC lattice, they typically foster the occurrence of defor­
mation twinning. Stiegler et al. [47] revealed that the introduction of
of the A attribute in both models, which is expected. Grain size plays a solid solution atoms elevates the lattice friction stress and decreases
pivotal role in facilitating deformation twinning [42]. Typically, crystals cross-slip capability, thereby promoting the initiation of deformation
with larger grain sizes exhibit greater susceptibility to deformation twinning. Bolling et al. [48] demonstrated that the inclusion of solution
twinning. The effect of grain size on slip and twinning stresses follows atoms reduces the rate of dislocation migration and decreases the
the Hall-Petch equation, with twinning stress showing a higher de­ number of mobile dislocations, resulting in a transition from slip to
pendency on grain size [43]. Consequently, with an increase in grain twinning. In the Fe81Ga19 alloy, Ga elements exist in the form of sub­
size, there is a simultaneous decrease in both slip and twinning stresses. stitution atoms in the BCC lattice, which increases the possibility of
However, the rate of decrease in twinning stress is higher, facilitating deformation twinning.
the occurrence of deformation twinning. According to certain perspectives presented in the literature[4] [49,
The decision tree model enables the determination of twin nucle­ 50], a decrease in temperature generally enhances the formation of
ation occurrence within the grains of deformed Fe81Ga19 alloy, deformation twins. This phenomenon arises from the reduction in
employing the following criteria: 1) In the case of smaller sized grains, if dislocation mobility associated with lower temperatures. Compared to
the model satisfies the conditions: A ≤ 7640, MAD ≤ 0.955, TwSF ≥ the critical stress for twin nucleation, the critical stress for dislocation
0.49, and BC > 146, it infers the presence of twin nucleation within the slip is more temperature sensitive. As the temperature is lowered beyond
grains. Here, the MAD, TwSF, and BC attributes impose constraints on a certain threshold, deformation twins reach the stress level necessary
grains. 2) For moderate sized grains, if the model satisfies the condition: for their initiation before dislocation slip, thereby assuming a dominant
7640 ≤ A ≤ 8724 and NG > 6, it identifies twin nucleation within the role in the plastic deformation process. Nevertheless, research con­
grains. 3) Larger sized grains are determined by the conditions: 8724 ≤ ducted by Meyers et al. [51] has indicated that the twin stress of various
A ≤ 28098 and BC ≤ 139. These rules provide valuable insights into twin BCC metals remains constant within the temperature range of 4–400 K.
nucleation within the grains of Fe81Ga19 alloy. Instead, the formation of deformation twins is influenced by increased
In previous investigations, we solely considered the characteristics of resistance to dislocation slip. In the case of in-situ compressive defor­
the Schmid factor and grain size based on the traditional model. The mation of Fe81Ga19 alloy, the occurrence of deformation twins appears
results revealed that deformation twins possess larger grain sizes, with a to be unaffected by temperature, possibly due to the alloy’s low
major concentration of TwSF around 0.46 [3]. However, by integrating responsiveness to temperature variations.
attributes such as A, TwSF, NG, BC, and MAD, the decision tree model Furthermore, studies have indicated that an increase in the rate at
establishes more definitive criteria. Contrasted with the traditional which strain is applied can facilitate the development of deformation

13
Z. Ma et al. Materials Today Communications 38 (2024) 108477

twinning [5,52]. When the motion of dislocations becomes insufficient while the SVM model achieves an accuracy of 98% and an F1
to accommodate the applied strain, twinning can activate additional slip score of 0.94, performing the best. The combined framework of
systems by adjusting the orientation of grains, thus contributing to EBSD and machine learning provides the possibility to explore
plastic deformation. Consequently, twinning acts as a complement to the efficient observation of stochastic events between micro­
dislocation slip and actively participates in the plastic deformation structure and material response during in-situ microscopy
process. According to the Ekelund formula, the Fe81Ga19 alloy exhibits a activities.
strain rate of 5.6 s− 1 during the rolling process. However, whether in the (4) It is observed that the Fe81Ga19 alloy exhibits poor responsiveness
rolling deformation process or in-situ compression deformation process, to temperature during deformation, and the intrinsic factors of
the strain rates in both plastic deformation procedures do not reside the grains, including grain size, grain orientation, stacking fault
within the domain of elevated strain rates. Consequently, the manifes­ energy, and alloying, are the primary influences on the nucle­
tation of twinning at lower strain rates in the Fe81Ga19 alloy implies its ation of deformation twinning.
susceptibility to lower strain rates.
The introduction of high-density deformation twins has been CRediT authorship contribution statement
extensively employed in the advancement of FCC metals to improve the
strength and ductility of the material. However, due to the instability of Yang Wang: Funding acquisition. Xiaoming Zhang: Writing – re­
intrinsic stacking faults in BCC metals and the rigorous conditions view & editing, Investigation, Funding acquisition, Data curation. R.D.
required for deformation twinning, the introduction of high-density K. Misra: Writing – review & editing. Feng Fang: Validation, Software.
deformation twins by manipulating the stacking faults poses a signifi­ Yukun Xia: Data curation. Yuchen Wang: Writing – review & editing,
cant challenge [53]. In this investigation, the constraints of low tem­ Validation. Zongwen Ma: Writing – review & editing, Writing – original
perature and high strain rate were surmounted during the rolling draft, Investigation, Funding acquisition, Formal analysis. Yuanxiang
deformation process of the Fe18Ga19 alloy with BCC-structured, enabling Zhang: Writing – review & editing, Funding acquisition, Formal
the formation of deformation twins in BCC metals under conditions of analysis.
moderate temperature and low strain rate, which is not an exception. Xie
et al. [25] conducted a study on the tensile deformation of the Fe-6.5 wt Declaration of Competing Interest
%Si alloy at moderate temperature (0.5 mm/min) and observed that the
presence of a considerable number of deformation twins reduced the The authors declare that they have no known competing financial
incompatibility of strain and the concentration of stress at grain interests or personal relationships that could have appeared to influence
boundaries, consequently enhancing the tensile ductility of the alloy. the work reported in this paper.
The dynamic slip-twinning mechanism offers novel prospects for BCC
metals to overcome the intrinsic brittleness of materials. Data availability

5. Conclusions Data will be made available on request.

This study investigates the twin nucleation behavior of Fe81Ga19 Acknowledgments

alloy with BCC structure during room temperature compression
employing an integrated approach combining in-situ EBSD technique This work was financially supported by the National MCF (Magnetic
and machine learning algorithms. The factors influencing twin nucle­ Confinement Fusion) Energy R&D Program of China (with Grant no.
ation are discussed. The following key findings are obtained. 2018YFE0306102), National Natural Science Foundation of China
(52104372, 52001060). Authors are grateful to R.D.K. Misra for a
(1) Microstructural analysis after deformation reveals that deforma­ careful review and critical analysis of the study.
tion twins occurred in 32 of 235 grains. Nine attributes with the
potential to impact twin nucleation are identified and ranked in
Appendix A. Supporting information
terms of their importance. Specifically, grain area (A), average
image quality (BC), number of neighboring grains (NG), and the
Supplementary data associated with this article can be found in the
maximum value of the Schmid factor for twinning systems
online version at doi:10.1016/j.mtcomm.2024.108477.
(TwSF) are found to have significant influences on twin
nucleation.
References
(2) Predictive models, including the decision tree and the XGBoost
model, achieved high accuracies of 94% and 96% on the test set, [1] S. Guruswamy, N. Srisukhumbowornchai, A.E. Clark, J.B. Restorff, M. Wun-Fogle,
respectively, demonstrating strong generalization capability. The Strong, ductile, and low-field-magnetostrictive alloys based on Fe-Ga, Acta Mater.
decision tree model offers three criteria for different grain sizes to 43 (2000) 239–244.
[2] S.M. Na, A.B. Flatau, Deformation behavior and magnetostriction of polycrystalline
determine which grains undergo twinning after deformation. Fe–Ga–X(X = B, C, Mn, Mo, Nb, NbC) alloys, J. Appl. Phys. 103 (2008) 07D304.
Specifically, for smaller grain sizes, the conditions include A ≤ [3] Y. Xia, Y. Zhang, T. Wu, N. Wang, R. Ran, Y. Wang, F. Fang, G. Wang, Deformation
7640, MAD ≤ 0.955, TwSF ≥ 0.49, and BC > 146. Moderate-sized twinning caused by warm rolling and secondary recrystallization in twin-roll strip
casting Fe81Ga19 alloy, J. Alloy. Compd. 922 (2022) 166039.
grains necessitate compliance with 7640 ≤ A ≤ 8724 and NG > 6, [4] J.W. Christian, S. Mahajan, Deformation twinning, Prog. Mater. Sci. 39 (1995)
while larger grains require satisfaction of 8724 ≤ A ≤ 28098 and 1–157.
BC ≤ 139. These insights reveal a deeper understanding hidden [5] T. Smida, J. Bosansky, Deformation twinning and its possible influence on the
ductile brittle transition temperature of ferritic steels, Mater. Sci. Eng., A. 287
within the details of twin nucleation event in Fe81Ga19 alloy. (2000) 107–115.
Incorporating this knowledge into crystal plasticity modeling is [6] F. Sorbello, P.E.J. Flewitt, G. Smith, A.G. Crocker, The role of deformation twins in
expected to enhance the predictive performance and interpret­ brittle crack propagation in iron–silicon steel, Acta Mater. 57 (2009) 2646–2656.
[7] G.E. LAKSO, M.J. MARCINKOWSKI, Plastic deformation in Fe-Si Alloys, Metall.
ability of the model.
Mater. Trans. 5 (1974) 839–845.
(3) SVM and ANN models without considering the underlying [8] R. Groger, J. Holzer, T. Kruml, Twinning and antitwinning in body-centered cubic
framework exhibit excellent predictive performance on both the metals, Comput. Mater. Sci. 216 (2023) 111874.
training and test set by optimizing hyperparameters and [9] X. Wang, J. Wang, Y. He, C. Wang, L. Zhong, S.X. Mao, Unstable twin in body-
centered cubic tungsten nanocrystals, Nat. Commun. 11 (2020) 2497.
capturing potential attribute relationships. The ANN model ach­ [10] J. Wang, A.H.M. Faisal, X. Li, Y. Hong, Q. Zhu, H. Bei, Z. Zhang, S.X. Mao, C.
ieves an accuracy of 97% and an F1 score of 0.91 on the test set, R. Weinberger, Discrete twinning dynamics and size-dependent dislocation-to twin

14
Z. Ma et al. Materials Today Communications 38 (2024) 108477

transition in body-centred cubic tungsten, J. Mater. Sci. Technol. 106 (2022) [31] T. Chen, C. Guestrin, XGBoost: a scalable tree boosting system, Proc. 22nd ACM
33–40. SIGKDD Int. Conf. Knowl. Discov. Data Min. (2016) 785–794.
[11] B. Jiang, A. Tu, H. Wang, H. Duan, S. He, H. Ye, K. Du, Direct observation of [32] Y. Liu, X. Zou, S. Ma, M. Avdeev, S. Shi, Feature selection method reducing
deformation twinning under stress gradient in body-centered cubic metals, Acta correlations among features by embedding domain knowledge, Acta Mater. 238
Mater. 155 (2018) 56–68. (2022) 118195.
[12] J. Liu, H. Chen, Nucleation and grain boundary evolution in dynamic [33] J. Schmidhuber, Deep learning in neural networks: an overview, Neural Netw. 61
recrystallization of 316LN steel during hot deformation, Front. Mater. 6 (2019) (2015) 85–117.
209. [34] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (2015) 436–444.
[13] F. Zhang, Z. Liu, M. Yang, G. Su, R. Zhao, P. Mao, F. Wang, S. Sun, Microscopic [35] A. Jain, K. Nandakumar, A. Ross, Score normalization in multimodal biometric
mechanism exploration and constitutive equation construction for compression systems, Pattern Recognit. 38 (2005) 2270–2285.
characteristics of AZ31-TD magnesium alloy at high strain rate, Mater. Sci. Eng., A. [36] V. CHERKASSKY, Model complexity control and statistical learning theory, Nat.
771 (2020) 138571. Comput. 1 (2002) 109–133.
[14] Z. Liu, H. Wang, M.J.R. Haché, X. Chu, E. Irissou, Y. Zou, Prediction of [37] W.Z. Han, S.D. Wu, S.X. Li, Z.F. Zhang, Origin of deformation twinning from grain
heterogeneous microstructural evolution in cold sprayed copper coatings using boundary in copper, Appl. Phys. Lett. 92 (2008) 1–3.
local Zener-Hollomon parameter and strain, Acta Mater. 193 (2020) 191–201. [38] M. Muzyk, Z. Pakiela, K.J. Kurzydlowski, Ab initio calculations of the generalized
[15] A.H. Cottrell, B.A. Bilby, LX. A mechanism for the growth of deformation twins in stacking fault energy in aluminium alloys, Scr. Mater. 64 (2011) 916–918.
crystals, Philos. Mag. 42 (2010) 573–581. [39] Y.X. Zhang, M.F. Lan, Y. Wang, F. Fang, X. Lu, G. Yuan, R.D.K. Misra, G.D. Wang,
[16] K. Ogawa, Edge dislocations dissociated in {112} planes and twinning mechanism Microstructure and texture evolution of thin-gauge non-oriented silicon steel with
of b.c.c. metals, Philos. Mag. 11 (1965) 217–233. high permeability produced by twin-roll strip casting, Mater. Charact. 150 (2019)
[17] R. Priestner, W.C. Leslie, Nucleation of deformation twins at slip plane 118–127.
intersections in B.C.C. metals, Philos. Mag. 11 (1965) 895–916. [40] J. Wang, Z. Zeng, M. Wen, Q. Wang, D. Chen, Y. Zhang, P. Wang, Ht Wang,
[18] K.P.D. Lagerlof, On deformation twinning in b.c.c. metals, Acta Metall. Mater. 41 Z. Zhang, S.X. Mao, T. Zhu, Anti-twinning in nanoscale tungsten, Sci. Adv. 6 (2020)
(1993) 2143–2151. 2792.
[19] G.L.W. Hart, T. Mueller, C. Toher, S. Curtarolo, Machine learning for alloys, Nat. [41] A.J. SMOLA, B. SCHOLKOPF, A tutorial on support vector regression, Stat. Comput.
Rev. Mater. 6 (2021) 730–755. 14 (2004) 199–222.
[20] K.T. Butler, D.W. Davies, H. Cartwright, O. Isayev, A. Walsh, Machine learning for [42] O. Muránsky, M.R. Barnett, D.G. Carr, S.C. Vogel, E.C. Oliver, Investigation of
molecular and materials science, Nature 559 (2018) 547–555. deformation twinning in a fine-grained and coarse-grained ZM20 Mg alloy:
[21] K. Choudhary, B. DeCost, C. Chen, A. Jain, R.C. Francesca Tavazza1, C. WooPark, combined in situ neutron diffraction and acoustic emission, Acta Mater. 58 (2010)
A. Choudhary, A. Agrawal, S.J.L. Billinge, E. Holm, S.P. Ong, C. Wolverton, Recent 1503–1517.
advances and applications of deep learning methods in materials science, npj [43] M.J. MARCINKOWSKI, H.A. LIPSITT, The plastic deformation of chromium at low
Comput. Mater. 8 (2022) 59. temperatures, Acta Met. 10 (1962) 95–111.
[22] Z. Tong, L. Wang, G. Zhu, X. Zeng, Predicting twin nucleation in a polycrystalline [44] L. Meng, P. Yang, Q. Xie, H. Ding, Z. Tang, Dependence of deformation twinning on
mg alloy using machine learning methods, Metall. Mater. Trans. A 50 (2019) grain orientation in compressed high manganese steels, Scr. Mater. 56 (2007)
5543–5560. 931–934.
[23] Y. Gui, Q. Li, K. Zhu, Y. Xue, A combined machine learning and EBSD approach for [45] C.S. Hong, N.R. Tao, K. Lu, X. Huang, Grain orientation dependence of deformation
the prediction of {10-12} twin nucleation in an Mg-RE alloy, Mater. Today twinning in pure Cu subjected to dynamic plastic deformation, Scr. Mater. 61
Commun. 27 (2021) 102282. (2009) 289–292.
[24] A.D. Orme, I. Chelladurai, T.M. Rampton, D.T. Fullwood, A. Khosravani, M. [46] P. Yang, Q. Xie, L. Meng, H. Ding, Z. Tang, Dependence of deformation twinning on
P. Miles, R.K. Mishra, Insights into twinning in Mg AZ31: a combined EBSD and grain orientation in a high manganese steel, Scr. Mater. 55 (2006) 629–631.
machine learning study, Comput. Mater. Sci. 124 (2016) 353–363. [47] J.O. Stiegler, C.J. McHargue, The effect of impurities on mechanical twinning and
[25] J. Xie, H. Fu, Z. Zhang, Y. Jiang, Deformation twinning feature and its effects on dislocation behavior in body-centered cubic metals, 1963.
significant enhancement of tensile ductility in columnar-grained Fe-6.5 wt%Si [48] G.F. BOLLING, R.H. RICHMA, The effect of solute on slip and mechanical twinning
alloy at intermediate temperatures, Intermetallics 23 (2012) 20–26. in iron alloys, Can. J. Phys. 45 (1967) 541–557.
[26] H. Fu, Z. Zhang, Y. Jiang, J. Xie, Applying the grain orientation dependence of [49] S. Mahajan, D.F. Williams, Deformation twinning in metals and alloys, Int. Mater.
deformation twinning to improve the deformation properties of an Fe-6.5 wt%Si Rev. 18 (2013) 43–61.
alloy, J. Alloy. Compd. 689 (2016) 307–312. [50] N.A. BOUCHER, J.W. CHRISTIAN, The influence of pre-strain on deformation
[27] S.G. Hong, S.H. Park, C.S. Lee, Role of {10–12} twinning characteristics in the twinning in niobium single crystals, Acta Met. 20 (1972) 581–591.
deformation behavior of a polycrystalline magnesium alloy, Acta Mater. 58 (2010) [51] M.A. Meyers, O. Vohringer, V.A. Lubarda, The onset of twinning in metals: a
5873–5885. constitutive description, Acta Mater. 49 (2001) 4025–4039.
[28] J.D. Kelleher, B.M. Namee, A. D’Arcy, Fundamentals of machine learning for [52] J.N. Florando, B.S. El-Dasher, C. Chen, D.C. Swift, N.R. Barton, J.M. McNaney, K.
predictive data analytics, (2015). T. Ramesh, K.J. Hemker, M. Kumar, Effect of strain rate and dislocation density on
[29] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, the twinning behavior in tantalum, AIP Adv. 6 (2016) 045120.
M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, [53] J. Zhang, F. Sun, Z. Chen, Y. Yang, B. Shen, J. Li, F. Prima, Strong and ductile beta
D. Cournapeau, M. Brucher, M. Perrot, E. Duchesnay, Scikit-learn: machine Ti–18Zr–13Mo alloy with multimodal twinning, Mater. Res. Lett. 7 (2019)
learning in python, J. Mach. Learn. Res. 12 (2011) 2825–2830. 251–257.
[30] L. Breiman, J.H. Friedman, R.A. Olshen, C.J. Stone, Classification and regresssion
trees, Chapman & Hall/CRC, 1984.

15