Professional Documents
Culture Documents
1 s2.0 S2590150823000030 Main
1 s2.0 S2590150823000030 Main
Materials Letters: X
journal homepage: www.sciencedirect.com/journal/materials-letters-x
A R T I C L E I N F O A B S T R A C T
Keywords: FeCrAl alloys are among the most promising candidates for accident-tolerant fuel cladding material in light water
FeCrAl nuclear reactors. Despite their high-temperature oxidation resistance in corrosive environments coupled with
Machine Learning their hydrothermal corrosion resistance, a key challenge remains in optimizing the composition of the alloy that
Oxidation
can be achieved through statistical analysis. However, the current literature on FeCrAl alloy design lack studies
for designing alloys based on oxidation resistance. This study addresses that gap by developing a predictive
model for the oxidation of FeCrAl alloys based on an experimental dataset, which lays the groundwork for model-
based optimization for alloy composition.
* Corresponding author.
E-mail address: andrew.hoffman@ge.com (A. Hoffman).
https://doi.org/10.1016/j.mlblux.2023.100183
Received 3 June 2022; Received in revised form 3 January 2023; Accepted 15 January 2023
Available online 21 January 2023
2590-1508/© 2023 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
I. Roy et al. Materials Letters: X 17 (2023) 100183
Fig. 1. The performance of (a) RF and (b) BHM for predicting mass change in 20% random split of the dataset; (c)Splitting of data into train and test set for ensemble
and separate models.
Table 1 √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
}̅
√{
Prediction errors of Random Forest (RF) and BHM. √ 1 ∑ N
RMSE = √ (yi − ŷi ) 2
(1)
Normalized RMSE N i=1
Conditions No. of test cases RF BHM
RMSE
NRMSE = (2)
Steam 18 1.80 3.20 ymax − ymin
Air 6 0.20 0.45
BWR-HWC 6 0.71 3.5 Here N is the size of test set, yi is the experimental mass change and ŷi
BWR-NWC 5 10.77 13.5 is the predicted mass change for ith data entry. Fig. 1 shows the per
Ensemble 35 1.13 2.26 formance comparisons of the unified or ensemble models for (a)RF and
(b)BHM. The distribution of the training and test data is shown in Fig. 1
(c). Along with training RF and BHM with the whole dataset, we trained
and tested for four separate experimental conditions, e.g., steam, air,
libraries are used for development of BHM models.
BWR-HWC and BWR-NWC and the normalized RMSE of the four cases
Additionally, both modeling approaches are implemented in two
and listed in Table 1.
different ways. RF and BHM models are developed as a unified or
The importance of various features in predicting sp. mass change are
ensemble model first where they are trained for all oxidation conditions,
ranked and plotted (Fig. 2). Fig. 2(a) shows that for the former model Fe
e.g., steam, air, and prototypic boiling water reactor (BWR)-[hydrogen
concentration has the most impact in mass change prediction followed
water chemistry (HWC), and normal water chemistry (NWC)]. The
by the temperature. From the material science point of view, however,
different oxidation conditions are one hot encoded to the models. Then,
Fe is used as strengthening element for majority of Fe-containing alloys,
another two models of RF and BHM are developed for each individual
including steel, and thus should not be the most important factor from
conditions.
oxidation perspective. The specific mass change is sensitive to concen
tration of Al and Cr individually. Imagine two FeCrAl alloys where Fe
Result and discussion concentration is the same because the sum of Al and Cr concentration
remains fixed. But those two alloys are expected to have different
To compare the predictive performance of the different models oxidation chemistry, therefore, we reconstructed a model without Fe
quantitatively, Normalized Root Mean Square Error (NRMSE) for the (filler element) concentration as an input feature. In that case, Fig. 2(b)
predictions are defined by the following equations:
2
I. Roy et al. Materials Letters: X 17 (2023) 100183
Fig. 2. Features importance in random forest for predicting mass change in FeCrAl alloy during steam test (a) with and (b) without wt. % of Fe as a feature.
shows Al and Cr concentration to have the most impact in mass change Conclusions
prediction among the constituents. This emphasizes the need to bring in
the available domain knowledge in constructing the machine learning The work described in this article is the first step towards developing
models to generate meaningful outputs that should always be cross a systematic optimization framework for FeCrAl composition design,
validated by expert knowledge. which is of immediate importance in the nuclear power industry. A
Using both RF and BHM models, we analyze the effect of varying Cr predictive model forms the backbone of such optimization work. As
(12 to 25 wt%) and Al (0 to 10 wt%) concentration on sp. mass change complicated physics precludes the possibility of developing a physics-
prediction of FeCrAl. For high temperature oxidation for 2 h, we choose, based model of suitable fidelity, data-driven modeling remains the
(1) 1000 ◦ C and (2) 1200 ◦ C, to demonstrate the ability of predictive most obvious choice. Although the skewness in the dataset makes it
modeling. We have also plotted the confidence of each of these models in difficult for the predictive models to perform well in all ranges, the
a surface plot (Fig. 3). Although the results from BHM and RF are benefits of ML in conjunction with domain expert knowledge are
similar, the trend of the heatmap of sp. mass change is more continuous demonstrated, nevertheless. To summarize, a cumulative approach of
for BHM than RF because of the inherent smooth nature of the gaussian predictive modeling, optimization, and explainable artificial intelli
process which is the underlying model in the in BHM framework. gence needs to be integrated with conventional metallurgic theoretical
The eight subplots in Fig. 3 presents the response surface of specific and experimental, computational tools for faster convergence of optimal
mass change across different Al and Cr wt. % using the BHM (Fig. 3 composition for alloy design, and this work is an early step towards that
(a–d)) and RF (Fig. 3(e–h)) models. Fig. 3(a, c, e, g) represent the 2D vision.
plots where the colormap correspond to the specific mass change and
Fig. 3(b, d, f, h) represent the 3D plots where the Z axis correspond to the CRediT authorship contribution statement
specific mass change and color map correspond to the uncertainty
(standard deviation) of the predictions. From Fig. 3(a), it can be seen Indranil Roy: Writing – original draft, Writing – review & editing,
that low Cr (12–14 wt%) and Al (0–2 wt%) results in high negative Project administration, Supervision. Subhrajit Roychowdhury: Vali
specific mass gain and the model has high confidence in that region as dation, Methodology, Visualization. Bojun Feng: Formal analysis,
well (Fig. 3(b)). It comes as no surprise as without a protective Alumina Validation, Methodology, Visualization. Sandipp Krishnan Ravi:
layer, Chromia tends to vaporize at high temperature (1200 ◦ C). The Formal analysis, Validation, Methodology, Visualization. Sayan Ghosh:
high Al and high Cr in Fig. 3(a, c) regions for both 1000 ◦ C and 1200 ◦ C Formal analysis, Validation, Methodology, Visualization. Rajnikant
show very small mass gain after two hours that may look ideal compo Umretiya: Data curation, Validation. Raul B. Rebak: Funding acqui
sition space for alloy development. Both these regions fall under low sition. Daniel M. Ruscitto: Supervision. Vipul Gupta: Supervision.
confidence regime Fig. 3(b, d), indicating that more experimental data is Andrew Hoffman: Supervision, Data curation, Validation.
necessary. From a metallurgic point of view, we cannot blindly trust ML
alone and disregard regions in the Al-Cr composition space without a
Declaration of Competing Interest
deeper dive into the domain knowledge. Furthermore, specific mass gain
is only one of the many deciding parameters of high temperature
The authors declare that they have no known competing financial
oxidation. However, in this study we show how predictive modeling and
interests or personal relationships that could have appeared to influence
intelligent visualization techniques allows us to focus on regions of in
the work reported in this paper.
terest where the model has high confidence and design future experi
ments is low confidence regime.
Data availability
3
I. Roy et al. Materials Letters: X 17 (2023) 100183
Fig. 3. Specific mass gain is predicted during oxidation after 2 h in steam of FeCrAl alloy with varying Cr and Al concentration (left column) and the uncertainty of
the model (right column). BHM prediction for (a)1200 ◦ C, (c)1000 ◦ C with their respective uncertainty (b and d). RF prediction for (e)1200 ◦ C, (g)1000 ◦ C with their
respective uncertainty (f and h).
4
I. Roy et al. Materials Letters: X 17 (2023) 100183
Acknowledgment [4] V.K. Gupta, N.C. KUMAR, A.J. Vinciquerra, L.C. Dial, V.S. Dheeradhada, T. Hanlon,
L. Salasoo, X. PING, S. Roychowdhury, J.J. GAMBONE, In-Situ Monitoring System
Assisted Material and Parameter Development for Additive Manufacturing,
This work is supported by the US Department of Energy, National US20200298499A1, 2020.
Nuclear Security Administration, under award number DE-NE0009047 [5] A. Roy, I. Roy, L.J. Santodonato, G. Balasubramanian, JOM 74 (2022) 1406–1413.
[6] S. Felix, S. Ray Majumder, H.K. Mathews, M. Lexa, G. Lipsa, X. Ping, S.
Roychowdhury, T. Spears, Sci. Rep. 12 (2022) 8503.
References [7] S. Ghosh, P. Pandita, S. Atkinson, W. Subber, Y. Zhang, N.C. Kumar, S. Chakrabarti,
L. Wang, ASCE-ASME J. Risk Uncert. Engrg Syst. Part B Mech. Eng. 6 (2020).
[1] A.K. Hoffman, R.V. Umretiya, V.K. Gupta, M. Larsen, C. Graff, C. Perlee, P. Brennan, [8] L. Deng, C. Wang, J. Luo, J. Tu, N. Guo, H. Xu, P. He, S. Xia, Z. Yao, Mater. Charact.
R. Rebak, JOM 74 (2022) 1690–1697. 188 (2022), 111894.
[2] E. Aydogan, J.S. Weaver, S.A. Maloy, O. El-Atwani, Y.Q. Wang, N.A. Mara, J. Nucl. [9] D. Morgan, G. Pilania, A. Couet, B.P. Uberuaga, C. Sun, J. Li, Curr. Opin. Solid State
Mater. 503 (2018) 250–262. Mater. Sci. 26 (2022), 100975.
[3] Y. He, J. Liu, S. Qiu, Z. Deng, Y. Yang, A. McLean, Mater. Sci. Eng. A 726 (2018)
56–63.