Professional Documents
Culture Documents
Graph Neural Networks For Efficient Learning of Mechanical Properties of Polycrystals Jonathan M Hestroffer Full Chapter
Graph Neural Networks For Efficient Learning of Mechanical Properties of Polycrystals Jonathan M Hestroffer Full Chapter
Graph Neural Networks For Efficient Learning of Mechanical Properties of Polycrystals Jonathan M Hestroffer Full Chapter
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:
CONTINUOUS PROGRESS.
Already the new bakery was becoming too small for the work, and
in August 1908 it was agreed to complete the central portion of the
building by the addition of other two storeys at a cost of £2,263. Just
prior to this time, however, the committee felt compelled to dispense
with the services of their foreman baker, and Mr W. H. Bell, who had
been his assistant, was appointed in his place. For several years
Belfast Society used the stable at the bakery for stabling their horses,
while they were also permitted to erect a cart shed and a shoeing
forge. Later this society erected stables, etc., for themselves on
property adjoining the bakery, and the ground which had been
occupied by them was utilised by the Federation for necessary
extensions to the bakery.
By March 1909 the extended premises were ready for occupation.
It was not long ere the directors were in the mortar tub again,
however, for another extension costing almost £4,000 was entered
on in the autumn of 1910; and not long after it was completed there
came, in December 1911, an urgent request from the advisory
committee for more ovens, and another extension, at a total cost of
almost £5,000, was entered on and completed in 1912. These various
extensions practically completed the bakery as it stands at present,
except for alterations, minor in themselves, which were made from
time to time during the war period with the object of increasing the
working facilities.
In July 1909 the committee recorded their satisfaction that the
average weekly sales from the bakery amounted to £1,219, while in
March 1910 the weekly turnover had reached 500 sacks; and to meet
the increasing demand three new ovens had to be erected. The
Bakery continued to win prizes at the Agricultural Hall and other
exhibitions, thus proving that Co-operative bread baking on a large
scale was equal to producing bread of the finest quality.
Early in 1910 the directors were saddened by the news that one of
the members of the first advisory committee, Mr Crook, of Lisburn,
had passed away.
Although most of the societies in the North were doing well, there
were one or two which were in a bad way. In 1908 Lurgan Society
had to close its doors. The Federation were creditors to the extent of
£114, and when settling day came it was found that the assets of the
dead society were only capable of returning 1/6 in the pound. In 1910
Newry Society went the same way. The Federation were creditors to
the extent of £200, and it was expected that the assets would realise
10/ per pound. In Newry the Federation made temporary
arrangements to carry on the bread trade, as had been done in
Banbridge, but after some time this course was abandoned.
Shortly after the new bakery was opened the two Dublin societies
were in consultation with the committee of the Federation about the
erection of a bakery, and the committee agreed to assist them. The
two societies were unable to agree, however, and the idea of a federal
bakery for Dublin was departed from. In 1910 Dublin Industrial
Society erected a bakery for themselves, and the opening of this
bakery was followed in a short time by the amalgamation of the two
societies. The Dublin Society, however, experienced considerable
difficulty in acquiring the knack of baking good bread—their position
in this respect recalls some of the earlier experiences of the U.C.B.S.
—and the Federation readily consented to Mr Bell or his assistant
visiting Dublin to put them on right lines. In 1913, when the distress
due to the strike was at its height in Dublin and the Trade Unions
Congress was coming to the rescue with financial support, the
Industrial Society received a contract for the supply of from 3,000 to
5,000 loaves daily. As they were unable to handle the contract in
their own bakery they secured the assistance of the U.C.B.S. bakery
in Belfast, which supplied them with the needed quantity of bread
during the period covered by the contract. Later, during the period of
the war, the Industrial Society was in considerable difficulty for a
time, and those responsible for its management were exceedingly
anxious that the U.C.B.S. should take over the bakery. This was not
done, however, and fortunately the society was able to maintain and
even to improve its position.
GOVERNMENT CONTRACTS.
At the beginning of 1909 the Bakery secured a contract for the
supply of bread and flour to Newtownards Barracks, and since then
have supplied from time to time that Barracks, as well as those of
Holywood, Belfast, and Kilroot, while, either directly or through local
societies, they have been successful in securing contracts for various
local institutions, one society during the war securing the contract
for the Admiralty.
In 1912 the Society had loaned some motor lorries for an
excursion, and during the day a painful accident took place whereby
one child was killed and three others were severely injured. The
Society’s motorman was completely exonerated from blame for the
accident, but in token of their sympathy the Society paid the doctor’s
fees and granted £55 to the relatives. In 1913 two of the societies
were experiencing difficulty in selling the bread because of the unfair
competition to which they were being subjected in attempts to cause
them to give up the bread trade, but the Federation came to their
assistance by allowing them a little additional discount on their
purchases. In 1911 the Irish Agricultural Wholesale Society was
appointed agent in Ireland for the Federation, and some time later it
was stated that since that federation had become agent the trade had
increased. At the Dublin Congress in 1914 the Federation erected a
scone and oatcake baking plant as part of their exhibit in the
Congress exhibition, and this exhibit was an object of much interest
to the Dublin visitors to the exhibition.
THE WAR AND ITS EFFECTS.
In Ireland, as in Scotland, the commanding position of the
U.C.B.S. had a steadying effect on the price of bread at the outbreak
of war. The Society carried good stocks of flour, and by adopting the
fixed policy of regulating the price of bread by the average price of
the flour in stock was able to maintain the price at a lower level than
the current price of flour warranted. In this way, while it was
possible to maintain full stocks of flour, the Bakery was able to keep
the price of bread at a halfpenny per 4–lb. loaf below the price at
which other bakers wished to sell it, and so saved the people of the
North many thousands of pounds.
The difficulties of transport were experienced by the branch,
however, in a much more marked degree than by the parent body. All
coal, flour, sugar, etc., used in the bakery had to be imported, and as
transport costs went up so also did prices, until bread was being sold
in Belfast at one shilling for the 4–lb. loaf. So short did the supply of
coal become that in 1915 Belfast Corporation was unable to supply
the Bakery with coke for firing purposes and a supply had to be sent
over from Glasgow. Flour and sugar also became very scarce, but
notwithstanding those facts the output of the Bakery continued to
increase rapidly until the coming of Government Regulation flour,
with its huge proportion of offal and other nastinesses, created a
distaste for bread amongst the public. At the outbreak of war the
output of the Bakery averaged 766 sacks per week, while by the end
of 1916 this had risen to 892 sacks per week, an increase of 16½ per
cent. in two and a half years. From then it gradually declined, until at
the end of the 63rd quarter of the working life of the branch it was
only 683 sacks per week. From that time and during the next six
months the trade again increased, so that by the end of the period
with which this history deals, although it had not reached the high-
water mark of 1916, all the evidence went to show that that point
would soon be left behind.
In 1918, with the object of relieving the congestion at M‘Neil
Street, it was resolved to establish an oatcake baking department at
Belfast, and for this purpose several hot-plates were transferred to
the Belfast bakery, as well as several girls from the M‘Neil Street
oatcake factory. This department has since proved a valuable adjunct
to the branch, the plant having had to be augmented and the number
of bakers employed increased.
Just at the end of the period Belfast became involved in one of the
most widespread strikes in the history of the Labour movement, and
practically all work except that of bread baking was brought to a
standstill. The electricity workers of the Corporation came out on
strike along with the other workmen, but fortunately for the branch
it had a power-generating plant of its own, and so was able to
continue at work. By this means the trade of the branch was
materially increased. The fact that the Federation has always been
willing to meet the requests of the employees in a reasonable manner
has always enabled the branch to maintain its popularity with the
Belfast workmen, and so impressed were the executives of the
Operative Bakers’ Society with the fairness and even liberality of the
committee that on one occasion an official letter was sent from the
union in which the board of management were thanked for the
assistance which their prompt action in agreeing to the request of the
union for a shorter working week had been in enabling that
concession to be imposed in the baking trade of the city.
When the branch was established in Belfast the distinct
understanding on which the U.C.B.S. took action was that as soon as
convenient it should be taken over by the local societies, but as the
years pass the likelihood of this being done seems to become more
remote. As has already been stated, the branch has been of great
assistance not only to Belfast Society but also to the majority of other
societies in the North. It has fostered the Co-operative spirit and Co-
operative idealism. From a weakling which was very much in need of
the fostering care of the movement in Scotland Co-operation in
Ireland, and especially in the North of Ireland, has grown to be a
strong and healthy organisation, with its centre in Belfast and with
branches scattered all over Ulster. It is probable that the Irish
societies are strong enough now to take over the branch if they were
so disposed, but the farseeing spirits amongst them see that there is
other work lying to their hand to which, relieved of the working of
the bakery, they can turn their undivided attention.
On the other hand, the proposal, adopted in 1917, that the
Federation should proceed, as and when expedient, to plant
branches in other centres of Co-operation in Ireland, will in course of
time cause the U.C.B.S. in Ireland to cease to be an exclusively North
of Ireland concern. It will do much, just as the other federation in
Ireland, the I.A.W.S., is doing much, to weld together into one united
movement the whole of the Co-operative societies in the country. It
will serve for Ireland, as it and its kindred associations in Scotland
and England are serving to-day, as an illustration of the fact that Co-
operation is not merely a principle for adoption by a few farmers or a
few workmen in a given district, but is also a principle on which
these isolated groups can be united into one powerful whole, by
becoming an institution where all can meet on a common platform.
CHAPTER XV.
A NEW PRESIDENT.