Professional Documents
Culture Documents
Full download Graph neural networks for efficient learning of mechanical properties of polycrystals Jonathan M. Hestroffer file pdf all chapter on 2024
Full download Graph neural networks for efficient learning of mechanical properties of polycrystals Jonathan M. Hestroffer file pdf all chapter on 2024
https://ebookmass.com/product/accelerators-for-convolutional-
neural-networks-arslan-munir/
https://ebookmass.com/product/physics-informed-neural-networks-m-
raissi-p-perdikaris-g-e-karniadakis/
https://ebookmass.com/product/analysis-and-visualization-of-
discrete-data-using-neural-networks-koji-koyamada/
https://ebookmass.com/product/python-deep-learning-understand-
how-deep-neural-networks-work-and-apply-them-to-real-world-
tasks-3rd-edition-vasilev/
Federated Learning for Future Intelligent Wireless
Networks Yao Sun
https://ebookmass.com/product/federated-learning-for-future-
intelligent-wireless-networks-yao-sun/
https://ebookmass.com/product/ai-applications-to-communications-
and-information-technologies-the-role-of-ultra-deep-neural-
networks-daniel-minoli/
https://ebookmass.com/product/building-computer-vision-
applications-using-artificial-neural-networks-2nd-edition-
shamshad-ansari/
https://ebookmass.com/product/graph-database-and-graph-computing-
for-power-system-analysis-renchang-dai/
https://ebookmass.com/product/the-mechanics-of-hydrogels-
mechanical-properties-testing-and-applications-elsevier-series-
in-mechanics-of-advanced-materials-1st-edition-hua-li-editor/
Computational Materials Science 217 (2023) 111894
Dataset link: https://github.com/jonathanhestr We present graph neural networks (GNNs) as an efficient and accurate machine learning approach to predict
offer/PolyGRAPH mechanical properties of polycrystalline materials. Here, a GNN was developed based on graph representation
of polycrystals incorporating only fundamental features of grains including their crystallographic orientation,
Keywords:
Titanium
size, and grain neighbor connectivity information. We tested our method on modeling stiffness and yield
Texture strength of 𝛼-Ti microstructures, varying in their crystallographic texture. We find the GNN predicts both
Graphs properties with high accuracy with mean relative errors of ∼ 1% for unseen microstructures from a given set
Deep learning of textures and <2% for microstructures of unseen texture, even when presented with limited training data. This
Strength accuracy is comparable to methods that require high-resolution three-dimensional (3D) microstructure data,
Stiffness such as 3D convolutional neural networks (3D-CNNs) and models that depend on the computation of spatial
statistics. The present results show that graph-based deep learning is a promising framework for property
prediction, especially considering the high cost associated with obtaining high-resolution 3D microstructure
data and the general scarcity of experimental materials datasets.
∗ Corresponding authors.
E-mail addresses: jonathanhestroffer@ucsb.edu (J.M. Hestroffer), latmarat@arizona.edu (M.I. Latypov).
https://doi.org/10.1016/j.commatsci.2022.111894
Received 28 June 2022; Received in revised form 8 October 2022; Accepted 31 October 2022
Available online 12 November 2022
0927-0256/© 2022 Elsevier B.V. All rights reserved.
J.M. Hestroffer et al. Computational Materials Science 217 (2023) 111894
2
J.M. Hestroffer et al. Computational Materials Science 217 (2023) 111894
Fig. 2. GNN architecture consisting of three fully-connected layers (FLs) and two
message-passing layers (MPLs).
3. Results
Fig. 3. Mean training and validation loss curves from 10-fold cross-validation for (a)
We trained and critically assessed our GNN model using a synthetic stiffness and (b) strength.
3
J.M. Hestroffer et al. Computational Materials Science 217 (2023) 111894
Table 1
Regression metrics for stiffness and strength inference of textures A through G. Results of the 10-fold cross-validation (valid.)
and test set (test) reported for the GNN, ANN, and dummy regressor model.
Model Stiffness Strength
MeanARE MaxARE R2 MeanARE MaxARE R2
(%) (%) (%) (%)
GNN (valid.) 0.25 1.05 0.998 0.92 5.44 0.994
GNN (test) 0.27 0.95 0.998 1.20 3.32 0.990
ANN (valid.) 0.59 3.00 0.985 1.35 9.59 0.985
ANN (test) 0.71 5.72 0.982 1.65 7.20 0.978
Dummy (valid.) 5.40 14.65 0 10.16 25.89 0
Dummy (test) 5.49 14.32 0 10.20 25.22 0
Table 2
Regression metrics for stiffness and strength inference of textures H through L. Values reported for GNN, ANN, and dummy
regressor models with entire and reduced (red.) training data.
Model Stiffness Strength
MeanARE MaxARE R2 MeanARE MaxARE R2
(%) (%) (%) (%)
GNN 0.43 1.87 0.986 1.36 5.91 0.950
GNN (red.) 0.65 3.29 0.967 1.67 7.70 0.937
ANN 1.24 5.08 0.898 2.15 14.31 0.911
ANN (red.) 1.03 5.61 0.910 2.19 11.71 0.901
Dummy 3.81 11.89 0 7.15 23.70 0
the typical evaluation scenario where training and test datasets are
drawn from the same distribution, in this case from the same tex-
ture groups. More impressive however is the model’s generalization
performance for new microstructures of unseen textures, even with
reduced training data. Such tests do not reach the level of extrapolation,
given the test textures H through L represent various combinations of
texture groups A through G, however they do represent more difficult
and probing tests of interpolation than what has been performed in
prior studies involving GNNs. The accuracy achieved by the ANN
suggests that the effective properties of MVEs can largely be predicted
through non-linear relationships of orientation and grain size, however
the performance of the GNN indicates that incorporating grain–grain
interactions results in consistent improvement of model accuracy across
all assessments.
To contextualize our model’s accuracy, valuable comparisons can
be made against work done using 3D-CNNs as well as statistics-based
models. The value comes from the very distinct representations of
microstructure used, with the GNN developed here containing only
rudimentary descriptors of grains and the only spatial information
available is the connectivity of grains, while both CNNs and statistics-
based models require full-field discretization of the microstructure.
These various methods have largely been used for predicting effective
properties of high-contrast composite materials reporting MeanAREs of
Fig. 5. Ground truth versus GNN predicted values of stiffness and strength for MVEs approximately 0.4% to 3.0% for both stiffness and strength inference
of textures H through L. Results when all training data is available (a, b), and those tasks [7,13,18]. The results of our GNN sit comfortably within this
for the limited data case (c, d). range; however, as these studies are not for polycrystalline materials,
the results are not directly comparable.
The best comparison that can be made is with work by Paulson and
in the previous assessments for seen textures, more substantial gains coauthors [11] that used a statistics-based model formulated on the
in accuracy are made with the GNN compared to the ANN for the computation of 2-point statistics to predict stiffness and yield strength
unseen textures in terms of the maximum prediction error indicating for the same dataset presented here. There, they conducted similar
that the treatment of grain–grain interactions is particularly beneficial experiments, training and testing their model on texture groups A
for reducing prediction outliers and increasing model precision. through G as well as testing on unseen textures H through L. When
performing calibration/validation of their model on texture groups
4. Discussion A through G, they reported very low MeanAREs of approximately
0.2% and 1.5% respectively for stiffness and strength compared to our
4.1. Generalization performance of the GNN GNN’s 0.25% and 0.92%. When testing on unseen textures H through
L they reported MaxAREs of 1% and 5% respectively for stiffness and
The GNN developed in this work achieves very high accuracy for strength, compared to our model’s MaxAREs of 1.87% and 5.91%. The
both stiffness and strength regression tasks with fixed architecture and accuracy achieved by the GNN model suggests that grain connectivity
hyper-parameter configuration. Results of the 10-fold cross-validation suffices to represent essential microstructural information for modeling
and corresponding testing indicate great generalization capability in effective properties of polycrystalline materials without the need of
4
J.M. Hestroffer et al. Computational Materials Science 217 (2023) 111894
Table 3
Regression metrics for stiffness and strength of GNN variants. Values reported for 10-fold cross-validation of textures A through
G.
Model Stiffness Strength
MeanARE MaxARE R2 MeanARE MaxARE R2
(%) (%) (%) (%)
GNN (base) 0.25 1.05 0.998 0.92 5.44 0.994
Euler angles 0.23 1.35 0.998 0.82 3.93 0.995
PReLU 0.25 1.34 0.998 1.14 6.96 0.991
Max aggregator 0.34 1.50 0.996 1.18 6.05 0.990
high-resolution spatial data. This is a desirable feature of our GNN, Additionally, the specific microstructure graph design method and
since such high-resolution data is very time consuming and expensive GNN architecture presented in this work is strongly suited for learn-
to acquire. ing multiscale properties of polycrystalline materials. The same node
embedding vectors, 𝐡(𝑘)𝑣 , used here to perform graph-level inference,
4.2. Effects of feature selection and intra-layer GNN design can be leveraged to perform node-level inference as done by Pagan
et al. [39]. This versatility allows GNNs to potentially address both
While many hyper-parameters of the GNN were optimized, others homogenization and localization problems with a single architecture.
were not tuned by authors. Notably, these are the chosen representation One limitation of the GNN developed here, is its ability to generalize
of crystallographic orientation and intra-layer hyper-parameters of the to new load directions. Currently, additional training would be required
GNN, including the nonlinear activation function, and neighborhood for the GNN to predict effective properties of the MVEs in either the
aggregator function of the message-passing layers. It is worthwhile to y- or 𝑧-direction as both the graphs and input graph features would
analyze the model’s sensitivity to these parameters as their effects on be identical between inference tasks, with only the target variables
model accuracy are not determinable a priori and can guide future changing. This limitation is not unique to our GNN as inference in a
GNN design. For this, the same 10-fold cross-validation experiment was new load direction would present a very difficult extrapolation problem
performed for different GNN variants, each variant representing the for any type of neural network. While the GNN can be trained for multi-
base GNN model with only one feature or hyper-parameter difference. output regression, predicting effective properties in all three directions
The three variants tested included using Bunge Euler angles instead simultaneously, the network lacks prior physical domain knowledge
of quaternions, parametric ReLU (PReLU) in place of ReLU activation of the material’s elastic anisotropy to generalize well to new load
function, and max rather than mean aggregator function. Overall, the directions.
results presented in Table 3 indicate only minor changes occur in Another limitation worth mentioning is that while the GNN pre-
model performance in terms of MeanARE and MaxARE for stiffness and sented here is well-suited for polycrystalline materials, it is not easily
strength regression across the different variants tested. The marginal employable for every type of microstructure. In particular, training a
performance differences suggest the GNN developed here is relatively similar regression model for a heterogeneous or two-phase microstruc-
insensitive to these hyper-parameters. ture poses certain challenges. For these kinds of microstructures, a
Choice of crystallographic orientation representation and its effect direct application of our graph construction method would result in
on model performance was of special interest. In the literature, po- a graph with only two nodes, one for each phase, connected by a
tential challenges in training deep learning models with orientation single edge. It is unknown how such a graph and accompanying GNN
space representations like Euler angles and quaternions have been would perform in predicting effective mechanical properties, even with
discussed as they exhibit multiple degeneracies due to inherent crystal added node features. What is certain is that the application of multiple
symmetries with generalized spherical harmonics as an alternative convolutional layers would be meaningless for such a graph. One
representation of orientations [53]. In this work, all grain orientations way around this might be to discretize the two-phase microstructure
were reduced to the hexagonal fundamental region to avoid symmetry into connected subregions. This however calls into question whether
equivalence issues. Our findings show the GNN exhibited high accu- 3D-CNNs should be used instead since they generalize well to any
racy in cross-validation experiments for both stiffness and strength regular grid 3D image data, including two-phase microstructures. That
regardless of the orientation space representation. being said, non-Euclidean discretizations could serve as the basis of
a microstructure graph, such as an unstructured mesh with variable
4.3. Extensions and limitations element size [54]. Such a scenario might support the usage of GNNs
over 3D-CNNs depending on how compact the input graph is compared
Polycrystalline microstructures lend themselves to intuitive graph to the regular grid representation of the microstructure [55].
design as grains and the physical boundaries they share with their The major benefit of GNNs compared to CNNs is that the computa-
neighbors can be described naturally as a multi-relational network. tional cost of many spatial convolution algorithms scale linearly with
Intuition can also be applied when assigning features to the graph. For the number of edges in the graph. For graphs based on polycrystalline
the inference tasks presented here, grain orientation and size proved microstructure, this scaling is advantageous given that the number
as sufficient node features. However, one can introduce more detailed of edges is largely independent of microstructure resolution [32,56].
information as needed, such as grain shape or even positional informa- This contrasts greatly with 3D-CNNs where computational cost scales
tion in the form of x-y-z coordinates. Characteristics of grain boundaries cubically with microstructure volume and dataset resolution. This scal-
can also be incorporated into the graph. As an example, it might ing benefit, when combined with the streamlined representation of
be desirable to weigh the individual messages passed by neighboring microstructure offered by graphs, results in significant reductions in
nodes according to the physical area two grains share. In this example, training time and computational resource dependence compared to
edge-weights, 𝛼𝑢,𝑣 , proportional to the boundary area, can be assigned CNNs. Recent studies reported GNN training times were 35 times
to the microstructure graph and affect the calculation of the aggregated faster than comparative CNNs for microstructure volumes of moder-
neighborhood representation in Eq. (B.3) in the following manner, ate size (1203 voxels, containing 300 grains) [37]. This makes GNNs
( ) particularly attractive for deep learning on very large microstructures
(𝑘)
∑ containing tens or hundreds of millions of voxels and thousands of
(𝑘−1)
𝐡 (𝑣) ← mean 𝛼𝑢,𝑣 ⋅ 𝐡𝑢 (1)
𝑢∈ (𝑣)
grains [57].
5
J.M. Hestroffer et al. Computational Materials Science 217 (2023) 111894
5. Conclusions
6
J.M. Hestroffer et al. Computational Materials Science 217 (2023) 111894
Fig. B.1. Parallel coordinates plot of the 200, 10-fold cross-validation experiments trained on stiffness data of textures A through G. Hyper-parameters listed left to right in
decreasing order of importance, based on the predicted complexity of MSE response relative to perturbations of the hyper-parameter. Optimal hyper-parameters highlighted in
orange.
where 𝑦𝑖 and 𝑦̂𝑖 the denote ground truth and predicted target values re-
spectively of the 𝑖th sample of 𝑛 total samples in a training or validation loss histories to ensure the GNN was not overfitting at the optimal
batch. Evaluation metrics reported to analyze model performance are number of epochs. No overfitting was observed, as shown in Fig. 3,
the mean absolute relative error (MeanARE), the maximum absolute and therefore no early stopping procedure was adopted. Training and
relative error (MaxARE), and the coefficient of determination (R2 ) inference times of the optimized GNN range from 30 to 180 s, and
defined below, 0.004 to 0.024 s respectively, depending on the amount of training data
(10,000 to 100,000 grains), using an NVIDIA RTX 2060 GPU with 6 GB
1 ∑ |𝑦𝑖 − 𝑦̂𝑖 |
𝑛
MeanARE = × 100% (B.7) of memory.
𝑛 𝑖=1 |𝑦𝑖 |
( )
|𝑦𝑖 − 𝑦̂𝑖 |
MaxARE = max𝑖 × 100% (B.8)
|𝑦𝑖 |
References
∑𝑛 ( )2
2 𝑖=1 𝑦𝑖 − 𝑦̂𝑖
R =1− ∑ ( )2 (B.9)
𝑛 [1] J. Segurado, R.A. Lebensohn, J. Llorca, Chapter one - computational homoge-
𝑖=1 𝑦𝑖 − 𝑦 nization of polycrystals, in: M.I. Hussein (Ed.), Advances in Crystals and Elastic
∑
where 𝑦 = 1𝑛 𝑛𝑖=1 𝑦𝑖 . These metrics are reported for 10-fold cross- Metamaterials, Part 1, in: Advances in Applied Mechanics, vol. 51, Elsevier,
2018, pp. 1–114, http://dx.doi.org/10.1016/bs.aams.2018.07.001, URL https:
validation as well as testing. In the case of 10-fold cross-validation //www.sciencedirect.com/science/article/pii/S0065215618300012.
or for the five repeated evaluations of reduced training data, ground [2] I.J. Beyerlein, M. Knezevic, Review of microstructure and micromechanism-based
truth and predicted values of all folds are included in the calculations. constitutive modeling of polycrystals with a low-symmetry crystal structure, J.
These metrics are evaluated for the GNN model as well as for a dummy Mater. Res. 33 (22) (2018) 3711–3738, http://dx.doi.org/10.1557/jmr.2018.333.
regressor model. The dummy model predicts the mean ground truth [3] I.J. Beyerlein, M. Knezevic, Mesoscale, Microstructure-Sensitive Modeling for
Interface-Dominated, Nanostructured Materials, Springer International Publish-
value for a given validation or test set, 𝑦̂ ← 𝑦.
ing, Cham, 2018, pp. 1–42, http://dx.doi.org/10.1007/978-3-319-42913-7_82-
1,
Hyper-parameter Tuning [4] M. Knezevic, I.J. Beyerlein, Multiscale modeling of microstructure-property
relationships of polycrystalline metals during thermo-mechanical deforma-
To maximize the performance of our GNN model, we carried out tion, Adv. Energy Mater. 20 (4) (2018) 1700956, http://dx.doi.org/10.
hyper-parameter optimization, where all model parameters apart from 1002/adem.201700956, URL https://onlinelibrary.wiley.com/doi/abs/10.1002/
adem.201700956, arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1002/adem.
the base architecture of the GNN were tuned. We selected 10-fold
201700956.
cross-validation MSE on a portion of the dataset as a loss metric to [5] W. Andreoni, S. Yip (Eds.), Handbook of Materials Modeling, Springer
minimize during hyper-parameter tuning. We utilized the optimization International Publishing, 2020, http://dx.doi.org/10.1007/978-3-319-42913-7.
tool SigOpt [62] to perform the optimization over a grid of possible [6] A. Cecen, H. Dai, Y.C. Yabansu, S.R. Kalidindi, L. Song, Material structure-
values for each hyper-parameter; parameters and possible values are property linkages using three-dimensional convolutional neural networks, Acta
listed in Table B.1. A total of 200 separate 10-fold cross-validation Mater. 146 (2018) 76–84.
experiments were performed each with a unique hyper-parameter con- [7] Z. Yang, Y.C. Yabansu, R. Al-Bahrani, W. keng Liao, A.N. Choudhary, S.R.
Kalidindi, A. Agrawal, Deep learning approaches for mining structure-property
figuration, results of these experiments are provided in Fig. B.1. Once linkages in high contrast composites from simulation datasets, Comput. Mater.
all experiments were completed, the top candidate model was selected Sci. 151 (2018) 278–287, http://dx.doi.org/10.1016/j.commatsci.2018.05.014,
for a more extensive analysis of its respective training and validation URL https://www.sciencedirect.com/science/article/pii/S0927025618303215.
7
J.M. Hestroffer et al. Computational Materials Science 217 (2023) 111894
[8] X. Li, Z. Liu, S. Cui, C. Luo, C.-F. Li, Z. Zhuang, Predicting the effective [29] Y. Koren, R. Bell, C. Volinsky, Matrix factorization techniques for recommender
mechanical property of heterogeneous materials by image based modeling and systems, Computer 42 (8) (2009) 30–37, http://dx.doi.org/10.1109/MC.2009.
deep learning, Comput. Methods Appl. Mech. Engrg. 347 (2019) http://dx.doi. 263.
org/10.1016/j.cma.2019.01.005. [30] W. Jiang, J. Luo, Graph neural network for traffic forecasting: A survey,
[9] C. Herriott, A.D. Spear, Predicting microstructure-dependent mechanical prop- 2021, http://dx.doi.org/10.48550/ARXIV.2101.11174, URL https://arxiv.org/
erties in additively manufactured metals with machine- and deep-learning abs/2101.11174.
methods, Comput. Mater. Sci. 175 (2020) 109599, http://dx.doi.org/10.1016/ [31] D. Duvenaud, D. Maclaurin, J. Aguilera-Iparraguirre, R. Gómez-Bombarelli,
j.commatsci.2020.109599, URL https://www.sciencedirect.com/science/article/ T. Hirzel, A. Aspuru-Guzik, R.P. Adams, Convolutional networks on graphs
pii/S0927025620300902. for learning molecular fingerprints, in: Proceedings of the 28th International
[10] X. Liu, Z. Yan, Z. Zhong, Predicting elastic modulus of porous Conference on Neural Information Processing Systems - Volume 2, NIPS ’15,
La0.6Sr0.4Co0.2Fe0.8O3-𝛿 cathodes from microstructures via FEM MIT Press, Cambridge, MA, USA, 2015, pp. 2224–2232.
and deep learning, Int. J. Hydrogen Energy 46 (42) (2021) 22079– [32] J. Zhou, G. Cui, S. Hu, Z. Zhang, C. Yang, Z. Liu, L. Wang, C. Li, M.
22091, http://dx.doi.org/10.1016/j.ijhydene.2021.04.033, URL https: Sun, Graph neural networks: A review of methods and applications, AI Open
//www.sciencedirect.com/science/article/pii/S0360319921013379. 1 (2020) 57–81, http://dx.doi.org/10.1016/j.aiopen.2021.01.001, URL https://
[11] N.H. Paulson, M.W. Priddy, D.L. McDowell, S.R. Kalidindi, Reduced-order www.sciencedirect.com/science/article/pii/S2666651021000012.
structure-property linkages for polycrystalline microstructures based on 2- [33] P. Veličković, G. Cucurull, A. Casanova, A. Romero, P. Liò, Y. Bengio, Graph
point statistics, Acta Mater. 129 (2017) 428–438, http://dx.doi.org/10.1016/ Attention Networks, in: International Conference on Learning Representations,
j.actamat.2017.03.009, URL https://www.sciencedirect.com/science/article/pii/ 2018, accepted as poster. URL https://openreview.net/forum?id=rJXMpikCZ.
S135964541730188X. [34] M. Sun, S. Zhao, C. Gilvary, O. Elemento, J. Zhou, F. Wang, Graph convolutional
[12] A. Gupta, A. Cecen, S. Goyal, A.K. Singh, S.R. Kalidindi, Structure–property link- networks for computational drug development and discovery, Brief. Bioinform.
ages using a data science approach: Application to a non-metallic inclusion/steel 21 (3) (2019) 919–935, http://dx.doi.org/10.1093/bib/bbz042, arXiv:https://
composite system, Acta Mater. 91 (2015) 239–254. academic.oup.com/bib/article-pdf/21/3/919/33227266/bbz042.pdf.
[13] M.I. Latypov, S.R. Kalidindi, Data-driven reduced order models for effective [35] M. Zhang, P. Li, Y. Xia, K. Wang, L. Jin, Revisiting graph neural networks for
yield strength and partitioning of strain in multiphase materials, J. Comput. link prediction, 2021, URL https://openreview.net/forum?id=8q_ca26L1fz.
Phys. 346 (2017) 242–261, http://dx.doi.org/10.1016/j.jcp.2017.06.013, URL [36] N.N. Vlassis, R. Ma, W. Sun, Geometric deep learning for computational
https://www.sciencedirect.com/science/article/pii/S0021999117304588. mechanics part I: Anisotropic hyperelasticity, Comput. Methods Appl. Mech.
[14] M.I. Latypov, L.S. Toth, S.R. Kalidindi, Materials knowledge system for nonlinear Engrg. 371 (2020) 113299, http://dx.doi.org/10.1016/j.cma.2020.113299, URL
composites, Comput. Methods Appl. Mech. Engrg. 346 (2019) 180–196, http: https://www.sciencedirect.com/science/article/pii/S0045782520304849.
//dx.doi.org/10.1016/j.cma.2018.11.034, URL https://www.sciencedirect.com/ [37] M. Dai, M.F. Demirel, Y. Liang, J.-M. Hu, Graph neural networks for an accurate
science/article/pii/S0045782518305930. and interpretable prediction of the properties of polycrystalline materials, Npj
[15] S.R. Kalidindi, A Bayesian framework for materials knowledge systems, MRS Comput. Mater. 7 (1) (2021) 103, http://dx.doi.org/10.1038/s41524-021-00574-
Commun. 9 (2) (2019) 518–531. w.
[16] S.R. Kalidindi, Feature engineering of material structure for AI-based materials [38] C. Shu, J. He, G. Xue, C. Xie, Grain knowledge graph representation learning: A
knowledge systems, J. Appl. Phys. 128 (4) (2020) 041103. new paradigm for microstructure-property prediction, Crystals 12 (2) (2022) http:
[17] Z. Yang, Y.C. Yabansu, D. Jha, W. keng Liao, A.N. Choudhary, S.R. Kalidindi, //dx.doi.org/10.3390/cryst12020280, URL https://www.mdpi.com/2073-4352/
A. Agrawal, Establishing structure-property localization linkages for elastic 12/2/280.
[39] D.C. Pagan, C.R. Pash, A.R. Benson, M.P. Kasemer, Graph neural network
deformation of three-dimensional high contrast composites using deep learn-
modeling of grain-scale anisotropic elastic behavior using simulated and mea-
ing approaches, Acta Mater. 166 (2019) 335–345, http://dx.doi.org/10.1016/
sured microscale data, 2022, http://dx.doi.org/10.48550/ARXIV.2205.06324,
j.actamat.2018.12.045, URL https://www.sciencedirect.com/science/article/pii/
URL https://arxiv.org/abs/2205.06324.
S1359645418309960.
[40] S.G. Kobourov, Spring embedders and force directed graph drawing algorithms,
[18] C. Rao, Y. Liu, Three-dimensional convolutional neural network (3D-
2012, http://dx.doi.org/10.48550/ARXIV.1201.3011, URL https://arxiv.org/abs/
CNN) for heterogeneous material homogenization, Comput. Mater. Sci. 184
1201.3011.
(2020) 109850, http://dx.doi.org/10.1016/j.commatsci.2020.109850, URL https:
[41] H. Zhou, P. Xu, X. Yuan, H. Qu, Edge bundling in information visualization,
//www.sciencedirect.com/science/article/pii/S0927025620303414.
Tsinghua Sci. Technol. 18 (2) (2013) 145–156, http://dx.doi.org/10.1109/TST.
[19] R. Pokharel, A. Pandey, A. Scheinker, Physics-informed data-driven surrogate
2013.6509098.
modeling for full-field 3D microstructure and micromechanical field evolution
[42] J. You, R. Ying, J. Leskovec, Design space for graph neural networks,
of polycrystalline materials, JOM 73 (2021) http://dx.doi.org/10.1007/s11837-
2020, http://dx.doi.org/10.48550/ARXIV.2011.08843, URL https://arxiv.org/
021-04889-3.
abs/2011.08843.
[20] D. Montes de Oca Zapiain, E. Popova, F. Abdeljawad, J.W. Foulk, S.R. Kalidindi,
[43] M. Fey, J.E. Lenssen, Fast graph representation learning with PyTorch geo-
H. Lim, Reduced-order microstructure-sensitive models for damage initiation
metric, 2019, http://dx.doi.org/10.48550/ARXIV.1903.02428, URL https://arxiv.
in two-phase composites, Integr. Mater. Manuf. Innov. 7 (3) (2018) 97–115,
org/abs/1903.02428.
http://dx.doi.org/10.1007/s40192-018-0112-0. [44] W.L. Hamilton, R. Ying, J. Leskovec, Inductive representation learning on large
[21] A. Marshall, S.R. Kalidindi, Autonomous development of a machine-learning
graphs, NIPS (2017).
model for the plastic response of two-phase composites from micromechanical [45] M.W. Priddy, N. Paulson, Synthetic alpha-ti microstructures and associated elastic
finite element models, JOM 73 (7) (2021) 2085–2095, http://dx.doi.org/10. stiffness and yield strength properties, 2016, URL https://matin.gatech.edu/
1007/s11837-021-04696-w. resources/52. (Accessed Jan 2022).
[22] D.M. de Oca Zapiain, S.R. Kalidindi, Localization models for the plastic response [46] M.W. Priddy, N. Paulson, D. McDowell, S.R. Kalidindi, Synthetic alpha-ti
of polycrystalline materials using the material knowledge systems framework, microstructures and associated elastic stiffness and yield strength properties
Modelling Simulation Mater. Sci. Eng. 27 (7) (2019) 074008, http://dx.doi.org/ - extended, 2017, URL https://matin.gatech.edu/resources/187. (Accessed Jan
10.1088/1361-651x/ab37a5. 2022).
[23] M.M. Bronstein, J. Bruna, Y. LeCun, A. Szlam, P. Vandergheynst, Geometric deep [47] M. Peters, G. Lütjering, G. Ziegler, Control of microstructures of (𝛼+ 𝛽)-titanium
learning: Going beyond Euclidean data, IEEE Signal Process. Mag. 34 (4) (2017) alloys, Int. J. Mater. Res. 74 (5) (1983) 274–282.
18–42, http://dx.doi.org/10.1109/msp.2017.2693418. [48] M. Peters, A. Gysler, G. Lütjering, Influence of texture on fatigue properties of
[24] W.C. Lenthe, M.P. Echlin, J.C. Stinville, M. De Graef, T.M. Pollock, Twin related Ti-6Al-4V, Metall. Mater. Trans. A 15 (8) (1984) 1597–1605, http://dx.doi.org/
domain networks in René 88DT, Mater. Charact. 165 (2020) 110365, http://dx. 10.1007/bf02657799.
doi.org/10.1016/j.matchar.2020.110365, URL https://www.sciencedirect.com/ [49] G. Lütjering, Influence of processing on microstructure and mechanical properties
science/article/pii/S1044580320305647. of (𝛼+ 𝛽) titanium alloys, Mater. Sci. Eng. A 243 (1–2) (1998) 32–45.
[25] H.F. Poulsen, Three-Dimensional X-Ray Diffraction Microscopy: Mapping Poly- [50] Y. Wang, J. Huang, Texture analysis in hexagonal materials, Mater. Chem. Phys.
crystals and their Dynamics, Vol. 205, Springer Science & Business Media, 81 (1) (2003) 11–26, http://dx.doi.org/10.1016/S0254-0584(03)00168-8, URL
2004. https://www.sciencedirect.com/science/article/pii/S0254058403001688.
[26] H.F. Poulsen, An introduction to three-dimensional X-ray diffraction microscopy, [51] G. Lütjering, J. Williams, Titanium, in: Engineering Materials and Processes,
J. Appl. Crystallogr. 45 (6) (2012) 1084–1097, http://dx.doi.org/10.1107/ Springer Berlin Heidelberg, 2007, URL https://books.google.com/books?id=
S0021889812039143. 41EqJFxjA4wC.
[27] M.P. Miller, M. Obstalecki, E. Fontes, D.C. Pagan, J.P.C. Ruff, A.J. Beaudoin, [52] B.D. Smith, Microstructure-Sensitive Plasticity and Fatigue of Three Titanium
Insit@CHESS, a resource for studying structural materials, Synchrotron Radiat. Alloy Microstructures (Ph.D. thesis), Georgia Institute of Technology, 2013.
News 30 (3) (2017) 4–8, http://dx.doi.org/10.1080/08940886.2017.1316124. [53] D. Montes de Oca Zapiain, A. Shanker, S.R. Kalidindi, Convolutional
[28] D.C. Pagan, K.E. Nygren, M.P. Miller, Analysis of a three-dimensional slip Neural Networks for the Localization of Plastic Velocity Gradient
field in a hexagonal Ti alloy from in-situ high-energy X-ray diffraction mi- Tensor in Polycrystalline Microstructures, J. Eng. Mater. Technol. 144
croscopy data, Acta Mater. 221 (2021) 117372, http://dx.doi.org/10.1016/ (1) (2021) http://dx.doi.org/10.1115/1.4051085, 011004, arXiv:https:
j.actamat.2021.117372, URL https://www.sciencedirect.com/science/article/pii/ //asmedigitalcollection.asme.org/materialstechnology/article-pdf/144/1/
S1359645421007515. 011004/6701898/mats_144_1_011004.pdf.
8
J.M. Hestroffer et al. Computational Materials Science 217 (2023) 111894
[54] A.L. Frankel, C. Safta, C. Alleman, R. Jones, Mesh-based graph convolutional [58] W.L. Hamilton, Graph representation learning, Synth. Lect. Artif. Intell. Mach.
neural networks for modeling materials with microstructure, J. Mach. Learn. Learn. 14 (3) (2020) 1–159.
Model. Comput. 3 (1) (2022) 1–30. [59] K. He, X. Zhang, S. Ren, J. Sun, Delving deep into rectifiers: Surpassing
[55] R. Hanocka, A. Hertz, N. Fish, R. Giryes, S. Fleishman, D. Cohen-Or, MeshCNN, human-level performance on ImageNet classification, in: 2015 IEEE International
ACM Trans. Graphics 38 (4) (2019) 1–12, http://dx.doi.org/10.1145/3306346. Conference on Computer Vision, ICCV, 2015, pp. 1026–1034.
3322959. [60] D. Kingma, J. Ba, Adam: A method for stochastic optimization, in: International
[56] T.N. Kipf, M. Welling, Semi-supervised classification with graph convolutional Conference on Learning Representations, 2014.
networks, 2016, http://dx.doi.org/10.48550/ARXIV.1609.02907, URL https:// [61] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin,
arxiv.org/abs/1609.02907. N. Gimelshein, L. Antiga, A. Desmaison, A. Köpf, E. Yang, Z. DeVito, M. Raison,
[57] J.C. Stinville, J.M. Hestroffer, M.A. Charpagne, A.T. Polonsky, M.P. Echlin, C.J. A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, S. Chintala, PyTorch: An
Torbet, V. Valle, K.E. Nygren, M.P. Miller, O. Klaas, A. Loghin, I.J. Beyerlein, imperative style, high-performance deep learning library, 2019, http://dx.doi.
T.M. Pollock, Multi-modal dataset of a polycrystalline metallic material: 3D org/10.48550/ARXIV.1912.01703, URL https://arxiv.org/abs/1912.01703.
microstructure and deformation fields, Sci. Data 9 (1) (2022) 460. [62] S. Clark, P. Hayes, SigOpt web page, 2019, URL https://sigopt.com.
9
Another random document with
no related content on Scribd:
disease, and died June 12th, 1863. The men became very much
attached to him during his brief sojourn with the Regiment.
SERGEANT-MAJOR GEORGE W.
BASSETT
Was born in the town of Barrington, Yates County,
November 6, 1838. When the war broke out, he was a
Law Student at Penn Yan. Enlisting as a private in Co. I,
he was chosen Third Sergeant, and promoted to
Sergeant-Major, May 22nd, 1861. Having borne off
Lieutenant Mix from the battle-field of Antietam, he
returned to the front and was immediately shot through the
head. By his winning ways and zealous attendance to
duties, he had won the esteem of his officers and
commanders, and fell universally regretted.