Aoje 1 011021
Aoje 1 011021
Aoje 1 011021
The architecture of the DLANN model chosen in this work has temperature and concentrations of 12 alloying elements. For
been listed in Table 2. The Mean Absolute Error (MAE) in MAE at 0.04, it will be rescaled to 0.004. MAE is about 4%
Table 2 appears to be large at first glance. As mentioned in Sec. when compared with the maximum value for AL2CU_C16 (θ),
2.3, data for all the design variables and objectives were scaled which is about 0.1. Thus, although MAE and MSE appear large
between 0 and 1. DLANN models were developed for all the for a few cases, they are still between 1% and 5% of the
stable and metastable phases as a function of composition and tem- maximum value, which is still quite small for a complex problem
perature. From phase stability data, it could be observed that the like this.
maximum amount (mole fraction) of any of these phases was One can observe that the chosen model has a different structure,
numerically low (about 0.1). Mole fraction value for any of these where 50–100–100–100 represents the number of neurons in the
phases varied between 0 and 0.1. four hidden layers of the DLANN. Based on our experience,
Regarding the error metrics in Table 2, apart from AL2CU_C16 DLANN models with acceptable performance metrics MSE and
(θ), mean square errors (MSE) for all the phases are quite low, on MAE, calculated over the validation/testing set, were chosen for
the order of 1e-3 to 1e-5. AL2CU_C16 (θ) and V_PHASE further study. A reader needs to take into account that there were
(η-MgZn2) have a mean absolute error (MAE) of about 0.4. Most several cases where the amount of phase is zero, and the alloy
of the MAE values for other phases are around 0.01, while MAE system contains 9–12 alloying elements. From the data sheet,
for AL3X for 7XXX series is extremely small, of the order of 67% of data were assigned to the training set and 33% of data
1e-3. MAE of around 0.01 still seems large. However, when were assigned to the testing set, mainly to avoid overfitting.
these models will be used as a predictive tool, these results will Hence, there is a possibility to improve the performance metrics
be rescaled back to provide a practical estimate of the amount of of the models by changing these parameters.
all these phases. For example, the initial range (0–0.1) was scaled
to a new range (0–1) for model development. After these models
3.2 Self-organizing Maps Analysis. Table 3 summarizes the
are used for prediction, these data will be rescaled, or the range
error metrics of SOM analysis for 2XXX, 6XXX, and 7XXX
(0–1) will be scaled back to the range (0–0.1). Thus, even if consid-
alloys. It can be observed in Table 3, that topological error is
ering MAE at 0.01, after rescaling it will become 0.001. In a
low, while quantization error is comparatively higher. In this
complex problem with 10–12 alloying elements and a large data-
work, predictive models were developed through TensorFlow/
base with several missing values, an error of 0.001 can be accept-
Keras libraries in PYTHON [16,17] with better prediction accuracy
able, though there is still room for improvement. Again, it should
(Table 2). Thus, SOM was used for understanding patterns in the
be recalled that in this work artificial neural network (ANN)
dataset, while for prediction, the Deep Learning Artificial Neural
model with four hidden layers was developed through Deep Learn-
Network (DLANN) models were used. As mentioned before, a
ing approach. ANN models are prone to overfitting. Thus, one must
commercial software Thermo-Calc was used so that phase stability
be careful while choosing a model and using it as a predictive tool.
calculations were performed within the framework of the
As mentioned before, priority was given to physical metallurgy of
CALPHAD approach.
aluminum alloys in selecting DLANN models, while using concepts
The CALPHAD approach utilizes Gibbs Energy minimization as
of statistics and artificial intelligence for guidance. AL2CU_C16 (θ)
a criterion to determine the formation and stability of any particular
data contain lots of missing values, and there are 13 variables:
phase [12,25]. In aluminum alloys (2XXX, 6XXX, and 7XXX), tar-
geted properties are achieved by an optimum combination of stable
and metastable phases. Thermodynamically, these metastable phases
Table 3 SOM Error metrics for 2XXX, 6XXX, and 7XXX series of will be absent in the presence of stable phases, but kinetically these
Al–Sc-based alloys metastable phases have been observed in the microstructure of alu-
minum alloys [2,3]. In order to stabilize a metastable phase, one
Quantization error Topological error
needs to remove the stable phases while performing calculations
2XXX 0.126 0.030 for the stability of metastable phases. In this work, a complex multi-
6XXX 0.109 0.034 component system for 2XXX, 6XXX, and 7XXX series alloys was
7XXX 0.131 0.017 chosen. It can be possible that a certain metastable phase can be
absent for a certain composition and temperature combination
even after the stable phases are removed from the calculation as regime around 200 °C and are unstable at elevated temperatures
several other equilibriums exist in a multicomponent system. Deter- [1]. SOM maps were able to recognize this pattern and correctly posi-
mining these compositions and temperature regimes can be helpful tioned the candidate alloys on the vertices of hexagonal units in a
in screening alloys prior to performing experiments. Thus, SOM way that the predictions can be verified through concepts of physical
plots can be used to determine the composition regime prior to per- metallurgy. Even though the SOM algorithm does not operate on
forming precipitation kinetics simulations for the Al3Sc phase or principles of physical metallurgy, still it was able to mimic these
experiments. vital correlations. This proves the efficacy of the application of
SOM maps in materials design as in this problem the database con-
tains several missing points, which is quite common in materials/
3.3. 6XXX Alloys. Figure 1 shows the component plot for alloy design problems. In Fig. 2, one can additionally observe that
6XXX alloys. In Fig. 1, one can observe that elements Al, Sc, elements aluminum, scandium, and AL3X phase are correlated.
and the phase AL3X are clustered together or are correlated. This An increase in aluminum and scandium results in an increase in
relation can be established as an increase in Al and Sc will lead the AL3X phase, which can be understood as the Al3X phase, in
to an increase in AL3X (Al3Sc) phase. In Fig. 1, MG2SI_C1 (β) this case, is Al3Sc.
and BETA_PRIME (β′ ) phases are clustered together. From the One of the goals of this work is to determine the scope of working
physical metallurgy of aluminum alloys, one can establish this rela- with small concentrations of Sc since Sc is expensive. From Fig. 2,
tion as BETA_PRIME (β′ ) is a metastable phase, which finally it can be observed that it is possible to work with a lower amount
transforms into a stable MG2SI_C1 (β) phase [1]. of Sc (<wt. 2%) at lower temperatures since BETA_PRIME,
Figure 2 shows the distribution of Sc, temperature, and a few BETA_DPRIME, and MG2SI_C1 phases are stable. Heat treatment
critical phases for 6XXX series. SOM maps consist of hexagonal of aluminum alloys in the case of Sc addition is performed in two
cells. Values visible on SOM maps are the average value over a stages. The annealing temperature for precipitating BETA_PRIME,
cell. Hexagonal cells are not visible in Fig. 2 since a significantly BETA_DPRIME, and MG2SI_C1 phases is performed at around
large number of candidate alloys are included in this study; thus, 100–200 °C, while the annealing temperature for precipitating
the SOM map looks pixelated. AL3X is usually above 300 °C. Thus, SOM maps proved to be
In Fig. 2, BETA_PRIME (β′ ) and BETA_DPRIME (β”) exist in helpful in determining candidates for performing solidification
the lower temperature regime (∼180 °C) and is nonexistent at ele- and heat-treatment simulations. Based on the preceding computa-
vated temperature (540 °C). Additionally, MG2SI_C1 (β) phase tional effort, the most promising Al–Sc-based alloy from each of
exists in a larger amount in a lower temperature regime (∼180 °C). the three Al-based series (2XXX, 6XXX, and 7XXX) was selected
From the physical metallurgy of aluminum alloys, BETA_PRIME for heat-treatment simulation as reported in Table 4. Candidates
(β′ ) and BETA_DPRIME (β′′ ) are the metastable phases that reported in Table 4 are also reported in Ref. [15], where simulations
finally transform into the thermodynamically stable MG2SI_C1 were performed for solidification and heat treatment for studying
(β) phase [1]. Additionally, MG2SI_C1 (β), BETA_PRIME (β′ ), the precipitation kinetics of the Al3Sc phase for these candidate
and BETA_DPRIME (β′′ ) phases are stable in the temperature alloys [15]. These compositions were obtained from artificial
intelligence-based algorithms. Practically, it is not possible to grain size of the Al3Sc phase at dislocations, available literature
achieve similar compositions as shown in Table 4. [1] was consulted. Depending on the application, Al3Sc grain size
can vary from 2 to 100 nm. Strengthening in aluminum alloys is
achieved by precipitation hardening and also mechanical working.
3.3.1 Heat-Treatment Simulations. In Ref. [15], simulations In this work, dislocation density was fixed at 6.0 × 1012 m−3. The
were performed where nucleation sites for precipitation of Al3Sc mechanical treatment introduces many dislocations in the system.
phases were in the bulk. In the current work, precipitation kinetics Thus, studying the precipitation kinetics of Al3Sc crystals with dis-
simulation of the Al3Sc phase was at the nucleation site which is at locations as a nucleation site will be helpful for researchers working
the dislocations. A candidate alloy corresponding to the 6XXX on aluminum alloys. In Ref. [15], the mean radius of the crystals
series from Table 4 was used, and isothermal annealing was per- increases continuously when the nucleation sites were in the bulk.
formed at the dislocations at 300 °C. A system of 10 alloying ele- In this work, nucleation sites are at the dislocations. A detailed anal-
ments, and no reported information on interfacial energy for such ysis at other temperatures will be helpful to quantify the effect of
a complex system. In this work, the reference temperature was time on overall grain size.
chosen as 300 °C. In Ref. [15], temperatures were chosen
between 300 °C and 450 °C, where all the nucleation sites were
in the bulk. Figure 3(a) shows the variation of the mean radius of
the grain over the annealing time, which was set at 1000 h.
4 Discussion
Figure 3(b) shows the variation of volume fraction over annealing
time. From Figs. 3(a) and 3(b), it can be observed that the mean In this work, the research problem was formulated in a way
radius and volume fraction of Al3Sc crystals increase initially and which will be helpful in expanding the domain of the current
saturate at about 67 h. After 67 h, there is minimal change in the state of the art or available literature on aluminum alloys studied
growth of Al3Sc crystals. Interfacial energy was optimized by under the framework of the CALPHAD approach.
trials and in this case, it was set at 0.07 J/m2. For the optimal The novelty of this work can be summarized in a few points as
follows:
Table 4 Compositions (wt%) of three series of Al-based alloys • Literature on aluminum alloys using the CALPHAD approach
chosen for solidification and heat-treatment simulation in a usually focuses on compositions which are around the known
previously published work [15] (reprinted with permission compositions of standard alloys [25]. In this work, a frame-
from Elsevier)
work was developed to predict and test novel compositions
Alloying element 2XXX Series 6XXX Series 7XXX Series
through a thorough investigation based on the CALPHAD
and Artificial Intelligence. Novel compositions can be ana-
Si 0.38828 1.5497 0.28606 lyzed through DLANN models and the equilibrium amount
Fe 0.3919 0.57313 0.21862 of various stable and metastable phases can be estimated in
Cu 4.95088 0.27893 1.51113 2XXX, 6XXX, and 7XXX classes of aluminum alloys
Mn 0.32005 0.63697 0.36188 through this approach.
Mg 0.53321 1.07076 1.05662 • Literature on aluminum alloys using the CALPHAD approach
Cr 0.04353 0.1437 0.12353 usually focuses on alloys with 6–8 alloying elements
Zn 0.19235 0.11882 3.73898 [4–11,25–29]. In this work, 12 alloying elements were consid-
Ti 0.04297 0.13088 0.15506
V 0.09389 0.0 0.0
ered for the 2XXX alloy system, 10 alloying elements for
Zr 0.21774 0.0 0.17829 6XXX, and 11 alloying elements for 7XXX series Al-based
Sc 0.92541 2.60077 0.75953 alloys, which will be helpful for researchers to utilize the pre-
Al 91.89979 92.89525 91.61031 sented approach in their own work, providing room for the
addition of new elements in their existing alloys.
• Scandium is added to aluminum alloys along with zirconium metastable phases for aluminum alloys under the framework of the
[30,31] and has a profound effect on mechanical properties CALPHAD approach. CALPHAD (Thermo-Calc) databases are
in aluminum alloys. Both elements contribute to improvement created from actual experiments and atomistic simulations, while
in strength, hardness, and stress corrosion resistance. [30,31]. DLANN models were developed under the framework of Tensor-
In a previously published work [15], alloys containing Zr Flow/ Keras libraries that are known for determining nonlinear and
were considered for Scheil solidification simulation and pre- complex correlations. The presented approach provides predictive
cipitation kinetic simulation of the Al3Sc phase [15]. The DLANN models that can be used for new chemical compositions
current work is mainly focused on determining novel compo- and temperatures to determine stable and metastable phases in the
sitions for 2XXX, 6XXX, and 7XXX alloys while determining alloys at the same instant, as experimentally these phases coexist,
the scope of Sc addition. Heat-treatment simulation performed but thermodynamically stable and metastable phases cannot
in this work is for the 6XXX alloy in Table 4, which does not coexist. Thereafter, data predicted through DLANN models were
contain any zirconium. This work cannot be compared with utilized to study 2XXX, 6XXX and 7XXX series Al–Sc-based
Refs. [30,31] as Zr containing alloy was not considered for systems and the stable and metastable phases in each of these
heat-treatment simulation. A previously published work [15] systems by utilizing Self-Organizing Maps (SOMs). SOMs are
can be compared to some extent. known for determining complex correlations between design vari-
• Solidification and heat-treatment simulations were performed ables and objectives and also correlations among the objectives.
and have been reported in a previously published work [15]. Models developed through TensorFlow/Keras libraries have low
In the current work, precipitation kinetics simulation results error and SOM analysis of these alloys also shows low topological
with the nucleation site at the dislocation have been included. error. Based on this work, a few of the most promising chemical
Isothermal annealing was performed at 300 °C. For the 6XXX compositions requiring low concentrations of Sc were developed
alloy in Table 4, interfacial energy was optimized, and the for performing heat-treatment simulations.
value is 0.07 J/m2. From Fig. 3, it can be observed that the Solidification and heat-treatment simulations were performed on
grain growth occurs in the beginning till about 67 h. After candidate alloys listed in Table 4 and had been reported [15]. In the
67 h, grain size becomes stable as even up to 1000-h annealing current work, heat-treatment simulations for 6XXX alloy at 300 °C
time there is minimal change in grain size. were performed where the nucleation sites were at dislocations.
From the results, it can be observed that the average grain size
Heat-treatment simulations require optimized values of interfa-
stabilizes at about 67 h, and there is minimal change in grain size
cial energy between precipitate and the matrix phase in order to
afterward even when isothermal simulations were performed for
mimic experimental findings, which is a must from an application
1000 h. In this case, interfacial energy for the Al3Sc phase was opti-
point of view. Thus, our planned future work will focus on optimiz-
mized. The presented work forms the basis for future work on opti-
ing [20] interfacial energy at various nucleation sites for multicom-
mizing interfacial energy for various nucleation sites. Grain-growth
ponent aluminum alloys with scandium addition.
simulations presented in this work are valid for single crystals only.
5 Conclusions Acknowledgment
This study presents a novel computational approach that can be The authors would like to express their sincere gratitude to Pro-
utilized for screening candidate alloys prior to performing experi- fessor Cristian Ciobanu from the Colorado School of Mines for
ments by estimating the equilibrium amount of various stable and providing access to some of the commercial software used in this
metastable phases in aluminum alloys containing scandium. In this paper. Thanks, are also due to Professor Carlo Poloni from
work, Deep Learning Artificial Neural Network (DLANN) was ESTECO Company for approving the free use of ESTECO’s soft-
developed by utilizing a database generated for various stable and ware modeFrontier in this research.
Fig. 6 SOM plot of components for alloying elements, and critical phases in 7XXX series of
Al-based alloys along with Sc and temperature
One of the goals for this work was to explore the possibility of Sc
addition in small amounts (<2 wt%). Through this work, it became ∗ −ΔG∗
possible to find the compositions and temperature regimes for Js = Zβ N0 exp (B3)
kT
which Al3Sc (AL3X) and other stable and metastable phases are
stable in significant amounts. Based on the preceding computational In Eqs. (B2) and (B3), Z which stands for Zeldovich factor, β∗ is
effort, the most promising Al–Sc-based alloy from each of the three the attachment rate of solute atoms to the precipitate (AL3X), N0
Al-based series was selected for heat-treatment simulation as denotes the number of nucleation sites available in the beginning,
reported in Table 4 in the main article and also in a previously pub- ΔG∗ is the Gibbs energy of formation of a precipitate, k is Boltz-
lished work [15]. mann’s constant, and T is absolute temperature [12,13,32–38]. In
Eq. (B2), θ can vary but in Thermo-Calc it is fixed at 2 [13].
The number density of precipitates, in the beginning, can be
shown as in Eq. (B4), while the number density (Nt) at any
Appendix B instant of time (t) can be expressed as in Eq. (B5) [13]
Kampmann–Wagner Numerical Approach. TC-PRISMA
module [13] in Thermo-Calc [12] utilizes the KWN approach for −ΔG∗
N0 exp (B4)
simulating precipitation kinetics of a precipitate which involves kT
nucleation, growth, and coarsening [32–36]. This section explained
the governing equations and lists the parameters used while simulat-
ing precipitation kinetics.
Nt = Jt dt (B5)
In this work, the matrix is the FCC_L12 phase. The precipitate is
AL3X, which is Thermo-Calc notation for Al3Sc. The molar
volume of the precipitate was taken from the Thermo-Calc data- ΔG∗ is the Gibbs energy (or activation energy) of formation of the
base. The grain aspect ratio was fixed at 1.0, while the mobility precipitate can be expressed as in Eq. (B6). ΔGFCC−Al
m
3 Sc
is the molar
enhancement factor was fixed at 5. Gibbs energy of formation of Al3Sc nanocrystals from the FCC
An initial number of nucleation sites (N0) were calculated from matrix and VmAl3 Sc is the molar volume of Al3Sc nanocrystals [34],
dislocation density, which was fixed at 6.0 × 1012/m3. Interfacial while σ is the interfacial energy between the FCC matrix phase
energy was set at 0.07 J/m2. and precipitate Al3Sc phase
Time-dependent nucleation rate (Jt) can be derived from classical
nucleation theory [37] and is expressed as in Eq. (B1) 16π σ 3int
−τ ΔG∗ = 2 (B6)
3 ΔG FCC−Al 3 Sc
Jt = Js exp (B1) m
t VmAl3 Sc
In Eq. (B1), τ is incubation time which can be further simplified
as shown in Eq. (B2), t is the time and Js is the steady-state nucle- Zeldovich factor (Z ) and β∗ can be expressed as shown by
ation rate as shown in Eq. (B3) [37–40] Eqs. (B7) and (B8), respectively [13,39–41]
1 VmAl3 Sc σ
τ= (B2) Z= (B7)
θZ 2 β∗ 2πNA R∗2 kT
2σVmAl3 Sc
R∗ =−
dRAl
t
3 Sc
DFCC Xi,tFCC − XiFCC Fig. 8 JMAK analysis for volume fraction estimations
= i
(B10)
dt ξi,t RAlt
3 Sc
X Al3 Sc − X FCC
i i
FCC
In Eq. (B10), Xi,t is the composition of the supersaturated FCC
matrix phase and XiAl3 Sc is the equilibrium composition of the Al3Sc
FCC
phase. The growth of Al3Sc crystals is dependent Xi,t . Thus, the
mass balance equation can be expressed as shown in Eq. (B11) [40]
FCC
Xi,t = (Xi,0
FCC
− VfAl3 Sc XiAl3 Sc )/(1 − VfAl3 Sc ) (B11)
1 Xi,t FCC
− XiFCC
ξi,t = 1 − λi,t π exp (λ2i,t )erfc(λi,t ) = 2
(B12)
2λi,t X Al3 Sc − X FCC
i i
3
4π(RAl
t
3 Sc
)
V Al
f ,t
3 Sc
= Nt (B17) References
3
[1] Røyset, J., and Ryum, N., 2005, “Scandium in Aluminium Alloys,” Int. Mater.
Rev., 50(1), pp. 19–44.
[2] Dorin, T., Ramajayam, M., Babaniaris, S., Jiang, L., and Langan, T. J., 2019,
“Precipitation Sequence in Al–Mg–Si–Sc–Zr Alloys During Isochronal Aging,”
Appendix C Materialia, 8, p. 100437.
[3] Thermo-Calc Software TCAL5: TCS Aluminium-based Alloys Database v.5,
Johnson–Mehl–Avrami–Kolmogorov Analysis. During nucle- https://www.Thermo-Calc.com/media/56675/TCAL5-1_extended_info.pdf, Accessed
ation and growth, the volume fraction of a precipitate follows the August 30, 2019.