Graph neural networks for efficient

learning of mechanical properties of

polycrystals Jonathan M. Hestroffer
Visit to download the full and correct content document:
https://ebookmass.com/product/graph-neural-networks-for-efficient-learning-of-mecha
nical-properties-of-polycrystals-jonathan-m-hestroffer/
 Computational Materials Science 217 (2023) 111894

Contents lists available at ScienceDirect

Computational Materials Science

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

Full length article

Graph neural networks for efficient learning of mechanical properties of

polycrystals
Jonathan M. Hestroffer a ,∗, Marie-Agathe Charpagne b , Marat I. Latypov c,d ,∗, Irene J. Beyerlein a,e
a
Materials Department, University of California, Santa Barbara, CA, USA
b
University of Illinois at Urbana-Champaign, Urbana, USA
c
Department of Materials Science and Engineering, University of Arizona, Tucson, AZ 85721, USA
d
Graduate Interdisciplinary Program in Applied Mathematics, University of Arizona, Tucson, AZ 85721, USA
e
Department of Mechanical Engineering, University of California, Santa Barbara Santa Barbara, CA, USA

ARTICLE INFO ABSTRACT

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.

1. Introduction with explicit account for microstructure. Belonging to a class of deep

A crucial element in materials design for advanced structural com-
ponents is the prediction of mechanical properties of materials based
on their microstructure. In the case of polycrystalline materials, many
microstructural features are known to affect their mechanical response,
such as grain size, shape, and crystallographic orientation, to name a
few. Explicit account for the crystallographic orientation, grain mor-
phology, along with grain-neighbor interactions and non-uniform intra-
granular stress fields is possible with full-field, spatially resolved three-
dimensional (3D) polycrystalline techniques such as crystal plasticity
finite element (CPFE) or fast Fourier transform (CPFFT) methods [1].
While powerful, these computational models come at a significant
computational cost [2–5]. The problematic trade-off between explicit
and complete 3D polycrystalline microstructure representation and the
computational costs calls for data-driven surrogate models that can
consider microstructure details and replace computationally expensive
micromechanical simulations.
Currently, 3D convolutional neural networks (3D-CNNs) [6–10] and
statistical learning models (also referred to as materials knowledge sys-
tems, or MKS) [11–15] are prevalent approaches to surrogate modeling
neural networks, 3D-CNNs automatically learn relevant spatial correla-
tions in microstructure through repeatedly applied spatial convolutions.
On the other hand, statistical learning models, which are machine
learning models based on microstructure statistics, depend on compu-
tation of spatial statistics (e.g., 𝑛-point correlations) as features [16].
Both approaches have been increasingly applied to model mechanical
properties, capable of both local and effective property prediction
as demonstrated for heterogeneous (e.g., two-phase composite) and
polycrystalline materials [17–22]. These models have also shown to be
robust and orders of magnitude faster than physics-based simulations.
A promising addition to these state-of-the-art frameworks is graph-
based deep learning [23], a set of techniques that generalize the
operations of structured deep learning models, such as CNNs, to the
non-Euclidean domain of graphs. Graphs present a lightweight, ver-
satile, and highly interpretable data structure for digitizing polycrys-
talline microstructure information. Typically, microstructure graphs
consist of nodes representing individual grains, while graph edges
connecting nodes represent the grain boundaries that they share [24].

∗ 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

Most importantly, through this representation, graphs naturally capture

topological characteristics of microstructures without the need for 3D
high-resolution spatial microstructure data. This makes graphs particu-
larly suitable representations of microstructure when high-resolution
spatial information is not experimentally accessible. Far-field high-
energy X-ray diffraction microscopy (HEDM) is a representative case,
where higher confidence is placed on measurements of grain centroid,
and crystallographic orientation, but less so on exact microstructure
geometry [25–27]. Graphs offer effective representation of microstruc-
ture data consistent with such measurements, which also maintains
information about the grain neighborhood and connectivity. Even sim-
ple structural analyses of microstructure graphs, outside of any ma-
chine learning, have yielded important insight into structure–property
relationships. Examples include analysis of node degree (number of
neighboring nodes) and node eccentricity (maximum distance to any
other node) to predict strain localization and fatigue hotspots in nickel-
base superalloy René 88DT [24] and analyzing various measures of
centrality (node importance) to reveal extended networks of grains that
drive plastic strain in Ti-7Al [28].
Graph neural networks (GNNs) have become a widely applied deep
learning method given the expressive power of graph data with well-
known applications including recommendation systems, traffic predic-
tion, and molecule design [29–32]. Operating much like CNNs, GNNs
adaptively learn multiscale localized spatial features of graph data Fig. 1. Sample MVE of 𝛼-Ti (a), with corresponding (0002) pole figure (b), microstruc-
ture graph (c), and graph with edge-bundling (d). For each graph, nodes represent
through layers of spatial graph convolution operations [32]. GNNs
individual grains with corresponding inverse pole figure coloring, while edges represent
are also very versatile, capable of producing low-dimensional vector boundaries shared between grains.
embeddings at multiple scales for graph, edge, or even node-level infer-
ence tasks with little or no change to model architecture or convolution
algorithm [33–35]. Only recently have GNNs been used for model-
letters, and sets by calligraphic font. We represent the 3D polycrys-
ing microstructure–property relationships with some early successes
talline microstructure of microstructure volume elements (MVEs) as a
being the prediction of stored elastic energy functionals [36], mag-
homogeneous, undirected graph (, ), with nodes, 𝑣 ∈ , constituting
netostriction of Tb0.3Dy0.7Fe2 alloys [37], elongation of magnesium
individual grains connected by edges, 𝑒 ∈ , signifying the boundaries
alloys [38], and grain-scale elasticity in Ni and Ti alloys [39]. These
that are shared between grains, accounting for any periodicity. Addi-
studies however present limited evaluations of model generalization
tionally, features are assigned to the individual nodes, {𝐱𝑣 , 𝑣 ∈ }.
capability with respect to the similarity of training and test set distri-
These features include corresponding grain-average crystallographic
butions, and offer minimal insight into model sensitivity in regard to
feature selection and GNN design. Given how nascent the application of orientation, reduced to the hexagonal fundamental region, and defined
graph-based deep learning is to the micromechanics of polycrystalline by the four-component unit quaternion vector, 𝐪, as well as grain size,
𝑑. This results in a 5-dimensional feature vector for each node, 𝐱𝑣 =
materials, cementing its usage in the field requires further exploration [ ]𝑇
of different inference tasks, model architectures and microstructure 𝑞0 , 𝑞1 , 𝑞2 , 𝑞3 , 𝑑 . A microstructure graph for a sample MVE is shown in
graph design. Fig. 1c, visualized in a 2D force-directed layout [40]. The same graph is
In this work, a GNN is developed for predicting the stiffness and provided in Fig. 1d with edge-bundling [41] applied to alleviate visual
yield strength of 𝛼-Ti based on 3D grain-level microstructure data. edge clutter.
Using crystallographic orientation and grain size as the only grain-
level attributes, the trained model achieves high prediction accuracy for 2.2. GNN model architecture
both regression tasks using a single architecture and hyper-parameter
optimization campaign. Additionally, a comprehensive assessment of
Once graph representation of microstructures was established, the
the GNN’s generalization performance was conducted showing that the
next step was to design a relevant GNN model architecture. As seen
model is capable of generalizing predictions of stiffness and strength
in Fig. 2, the GNN developed here consisted of one fully-connected
to new microstructures generated from texture groups both seen and
layer for feature pre-processing, two message-passing layers, a global
unseen during training. Accuracy of the GNN is comparable to cur-
mean pooling layer (graph-readout), and two fully-connected post-
rent state-of-the-art frameworks including 3D-CNNs and statistics-based
processing layers. This architecture was inspired by the general GNN
methods even in a limited training data setting, which is advantageous
given the current scarcity of experimental materials datasets. Lastly, design space proposed by You et al. [42]. For the message-passing
sensitivity to feature and hyper-parameter selection is analyzed and layers, we employ SAGE convolutional layers as implemented in the
potential extensions of the GNN are discussed. Altogether, our work PyTorch Geometric library [43]. These layers, which are based on the
highlights the importance of grain neighborhood connectivity on the GraphSAGE algorithm [44], train aggregator functions that best com-
effective mechanical response of polycrystals and the utility of their bine feature information of a node’s local neighborhood. This method
graphical representation. enables generalization across graphs with node features of the same
form. Finally, we implement an artificial neural network (ANN) which
2. Methods ignores grain–grain interaction to serve as a point of comparison for
the GNN. The ANN developed shares the same architecture and hyper-
2.1. Microstructure graph design parameters of the GNN except node neighborhood aggregation is not
performed, essentially reducing the message-passing layers into fully-
As a brief remark on notation, scalars are denoted by non-bold low- connected layers. See Appendix for details of the GNN and ANN along
ercase letters, vectors bold lowercase letters, matrices bold uppercase with descriptions of model training, evaluation, and optimization.

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.

𝛼-Ti dataset generated by Priddy et al. [45,46]. The dataset contained

1200 MVEs of polycrystalline 𝛼-Ti microstructures and their corre-
sponding effective mechanical properties. Each MVE was stochastically
generated from one of 12 distinct crystallographic textures (orientation
distributions) with equiaxed grain morphology and log-normal grain
size distribution. These textures, labeled A through L, are provided
in Fig. A.1 in Appendix for reference. Textures A through H were
inspired by literature [47–52], while textures I through L represented
different combinations of texture components present in textures A
through H. Effective mechanical properties of the MVEs were obtained
from CPFE simulations of uniaxial tension in x, y, and 𝑧-directions with
periodic boundary conditions. The simulations were performed under
displacement-controlled loading to a total strain of 1.5% where both
elastic stiffness and yield strength at 0.2% offset strain were determined
for each loading direction.
With this dataset, we first evaluated the accuracy of the GNN for
modeling stiffness and yield strength in the 𝑥-direction for texture
groups A through G. 10 MVEs of each texture group were reserved
for testing and 10-fold cross-validation was performed on the remain-
ing MVEs. Hyper-parameters of the GNN were optimized for stiffness
regression and then directly applied to strength regression, see Ap-
pendix for details. Results of cross-validation and testing for stiffness
and strength can be found in Fig. 4 in the form of parity plots and
respective regression metrics are reported in Table 1. The GNN is able
to accurately predict both stiffness and strength and generalize to new
Fig. 4. Ground truth versus GNN predicted values of stiffness and strength for MVEs
MVEs of texture groups seen during training. This is evident in the of textures A through G. Results of 10-fold cross-validation (a, b), and those for the
low values of MeanARE and MaxARE for both the validation and test test set (c, d).
sets, performing significantly better for both inference tasks than a
baseline dummy model that predicts the mean ground truth value for a
given validation or test set. In addition, the GNN performs marginally As shown in Fig. 5 and Table 2, results of this experiment show that
better than the ANN in terms of MeanARE, but shows more substantial the GNN predicts stiffness and yield strength for MVEs of unseen
improvement in terms of MaxARE. While accuracy of the GNN is high textures with accuracy comparable to the first experiment, with slightly
for both regression tasks, it performs relatively better at predicting increased MeanARE for each regression task, but still performing better
stiffness than strength with lower values of MeanARE and MaxARE than the ANN and dummy model.
respectively. The same insight can be gathered from the reported loss Finally, we assessed how well the GNN performs with limited
curves with ultimately lower mean validation loss achieved for stiffness training data. Especially in settings where obtaining ground truth la-
(Fig. 3a). This is expected partly as the GNN was optimized for stiffness, bels/values from experiment or simulation is time-consuming, machine
but may also indicate that yield strength is a more challenging property learning models which can learn from limited data are very attractive.
to learn. In a manner similar to cross-validation, the GNN was trained for both
Next, we evaluated the predictions of the GNN for unseen tex- stiffness and strength regression using only 20 of the 100 MVEs from
tures. Here, the same exact model, including architecture and hyper- each texture group A through G. The GNN was then tested using all
parameter configuration, was used to predict stiffness and strength for 500 MVE graphs from texture groups H through L. This was repeated
MVEs of textures H through L that were left out during the original five times, each time providing a separate set of 20 MVEs for training.
evaluation. Here, the model received the same set of graphs from Results of this experiment can be found in Fig. 5 and Table 2. Even
textures A through G for training, but a test set of 500 graphs from the with a reduced training set, the GNN still predicts stiffness and strength
never before seen textures H through L. This was a particularly useful with high accuracy. The most noticeable degradation of model perfor-
evaluation of the GNN’s capacity for interpolation and generalization mance is visible when comparing MaxARE values, particularly when
as a large fraction of MVEs from textures H through L exhibit modulus predicting strength (7.70% MaxARE). These increased deviations are
and strength values which are largely absent from the training data also evident in the parity plots, however they appear modest in the
in the ranges 130–145 GPa, and 950–1150 MPa respectively (Fig. 4). context of a near 5-fold reduction in training data. As demonstrated

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

Fig. 5. Ground truth versus GNN predicted values of stiffness and strength for MVEs
of textures H through L. Results when all training data is available (a, b), and those
for the limited data case (c, d).
in the previous assessments for seen textures, more substantial gains
in accuracy are made with the GNN compared to the ANN for the
unseen textures in terms of the maximum prediction error indicating
that the treatment of grain–grain interactions is particularly beneficial
for reducing prediction outliers and increasing model precision.
4. Discussion
4.1. Generalization performance of the GNN
The GNN developed in this work achieves very high accuracy for
both stiffness and strength regression tasks with fixed architecture and
hyper-parameter configuration. Results of the 10-fold cross-validation
and corresponding testing indicate great generalization capability in
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
approximately 0.4% to 3.0% for both stiffness and strength inference
tasks [7,13,18]. The results of our GNN sit comfortably within this
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
coauthors [11] that used a statistics-based model formulated on the
computation of 2-point statistics to predict stiffness and yield strength
for the same dataset presented here. There, they conducted similar
experiments, training and testing their model on texture groups A
through G as well as testing on unseen textures H through L. When
performing calibration/validation of their model on texture groups
A through G, they reported very low MeanAREs of approximately
0.2% and 1.5% respectively for stiffness and strength compared to our
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
strength, compared to our model’s MaxAREs of 1.87% and 5.91%. The
accuracy achieved by the GNN model suggests that grain connectivity
suffices to represent essential microstructural information for modeling
effective properties of polycrystalline materials without the need of

4
J.M. Hestroffer et al.
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
MeanARE
(%)
GNN (base) 0.25
Euler angles 0.23
PReLU 0.25
Max aggregator 0.34
high-resolution spatial data. This is a desirable feature of our GNN,
since such high-resolution data is very time consuming and expensive
to acquire.
4.2. Effects of feature selection and intra-layer GNN design
While many hyper-parameters of the GNN were optimized, others
were not tuned by authors. Notably, these are the chosen representation
of crystallographic orientation and intra-layer hyper-parameters of the
GNN, including the nonlinear activation function, and neighborhood
aggregator function of the message-passing layers. It is worthwhile to
analyze the model’s sensitivity to these parameters as their effects on
model accuracy are not determinable a priori and can guide future
GNN design. For this, the same 10-fold cross-validation experiment was
performed for different GNN variants, each variant representing the
base GNN model with only one feature or hyper-parameter difference.
The three variants tested included using Bunge Euler angles instead
of quaternions, parametric ReLU (PReLU) in place of ReLU activation
function, and max rather than mean aggregator function. Overall, the
results presented in Table 3 indicate only minor changes occur in
model performance in terms of MeanARE and MaxARE for stiffness and
strength regression across the different variants tested. The marginal
performance differences suggest the GNN developed here is relatively
insensitive to these hyper-parameters.
Choice of crystallographic orientation representation and its effect
on model performance was of special interest. In the literature, po-
tential challenges in training deep learning models with orientation
space representations like Euler angles and quaternions have been
discussed as they exhibit multiple degeneracies due to inherent crystal
symmetries with generalized spherical harmonics as an alternative
representation of orientations [53]. In this work, all grain orientations
were reduced to the hexagonal fundamental region to avoid symmetry
equivalence issues. Our findings show the GNN exhibited high accu-
racy in cross-validation experiments for both stiffness and strength
regardless of the orientation space representation.
4.3. Extensions and limitations
Polycrystalline microstructures lend themselves to intuitive graph
design as grains and the physical boundaries they share with their
neighbors can be described naturally as a multi-relational network.
Intuition can also be applied when assigning features to the graph. For
the inference tasks presented here, grain orientation and size proved
as sufficient node features. However, one can introduce more detailed
information as needed, such as grain shape or even positional informa-
tion in the form of x-y-z coordinates. Characteristics of grain boundaries
can also be incorporated into the graph. As an example, it might
be desirable to weigh the individual messages passed by neighboring
nodes according to the physical area two grains share. In this example,
edge-weights, 𝛼𝑢,𝑣 , proportional to the boundary area, can be assigned
to the microstructure graph and affect the calculation of the aggregated
neighborhood representation in Eq. (B.3) in the following manner,
( )
(𝑘)
∑ containing tens or hundreds of millions of voxels and thousands of
(𝑘−1)
𝐡 (𝑣) ← mean 𝛼𝑢,𝑣 ⋅ 𝐡𝑢 (1)
𝑢∈ (𝑣)
Computational Materials Science 217 (2023) 111894
Strength
MaxARE R2 MeanARE MaxARE R2
(%) (%) (%)
1.05 0.998 0.92 5.44 0.994
1.35 0.998 0.82 3.93 0.995
1.34 0.998 1.14 6.96 0.991
1.50 0.996 1.18 6.05 0.990
Additionally, the specific microstructure graph design method and
GNN architecture presented in this work is strongly suited for learn-
ing multiscale properties of polycrystalline materials. The same node
embedding vectors, 𝐡(𝑘)𝑣 , used here to perform graph-level inference,
can be leveraged to perform node-level inference as done by Pagan
et al. [39]. This versatility allows GNNs to potentially address both
homogenization and localization problems with a single architecture.
One limitation of the GNN developed here, is its ability to generalize
to new load directions. Currently, additional training would be required
for the GNN to predict effective properties of the MVEs in either the
y- or 𝑧-direction as both the graphs and input graph features would
be identical between inference tasks, with only the target variables
changing. This limitation is not unique to our GNN as inference in a
new load direction would present a very difficult extrapolation problem
for any type of neural network. While the GNN can be trained for multi-
output regression, predicting effective properties in all three directions
simultaneously, the network lacks prior physical domain knowledge
of the material’s elastic anisotropy to generalize well to new load
directions.
Another limitation worth mentioning is that while the GNN pre-
sented here is well-suited for polycrystalline materials, it is not easily
employable for every type of microstructure. In particular, training a
similar regression model for a heterogeneous or two-phase microstruc-
ture poses certain challenges. For these kinds of microstructures, a
direct application of our graph construction method would result in
a graph with only two nodes, one for each phase, connected by a
single edge. It is unknown how such a graph and accompanying GNN
would perform in predicting effective mechanical properties, even with
added node features. What is certain is that the application of multiple
convolutional layers would be meaningless for such a graph. One
way around this might be to discretize the two-phase microstructure
into connected subregions. This however calls into question whether
3D-CNNs should be used instead since they generalize well to any
regular grid 3D image data, including two-phase microstructures. That
being said, non-Euclidean discretizations could serve as the basis of
a microstructure graph, such as an unstructured mesh with variable
element size [54]. Such a scenario might support the usage of GNNs
over 3D-CNNs depending on how compact the input graph is compared
to the regular grid representation of the microstructure [55].
The major benefit of GNNs compared to CNNs is that the computa-
tional cost of many spatial convolution algorithms scale linearly with
the number of edges in the graph. For graphs based on polycrystalline
microstructure, this scaling is advantageous given that the number
of edges is largely independent of microstructure resolution [32,56].
This contrasts greatly with 3D-CNNs where computational cost scales
cubically with microstructure volume and dataset resolution. This scal-
ing benefit, when combined with the streamlined representation of
microstructure offered by graphs, results in significant reductions in
training time and computational resource dependence compared to
CNNs. Recent studies reported GNN training times were 35 times
faster than comparative CNNs for microstructure volumes of moder-
ate size (1203 voxels, containing 300 grains) [37]. This makes GNNs
particularly attractive for deep learning on very large microstructures
grains [57].
5
J.M. Hestroffer et al. Computational Materials Science 217 (2023) 111894

5. Conclusions

In summary, we demonstrate that GNNs can be used to predict effec-

tive mechanical properties of synthetic 𝛼-Ti microstructures of various
crystallographic textures. The GNN developed was capable of predict-
ing the stiffness and yield strength in a particular loading direction, and
generalizing well to unseen microstructure and textures. Accuracy of
the GNN is comparable to other machine learning frameworks, which
rely on full-discretization of microstructure such as 3D-CNNs or those
founded on explicit computation of 2-point statistics. The framework
developed here can be easily extended to other effective property
inferences without changing the selected design of the microstructure
graph. With that said, the total design space of microstructures, GNN
architectures, and choice of message-passing layers is vast and remains
largely unexplored. This makes graph-based deep learning an area of
great potential research for a wide variety of applications relying on
microstructure–property predictions. The success of our model inspires
a different approach to studying the micromechanics of polycrystals,
transitioning from the perspective of spatially resolved microstructures
to that of a network arrangement of grains. Fig. A.1. Sample MVE (0002) pole figures for each texture class A through L.

CRediT authorship contribution statement

with activation function, 𝜎, and 𝐡(𝑘) (𝑘−1)
𝑣 , 𝐡𝑣 , and 𝐡(𝑘)
 (𝑣)
indicating the
Jonathan M. Hestroffer: Conceptualization, Software, Formal updated node, current node, and aggregated neighborhood represen-
Analysis, Writing – original draft. Marie-Agathe Charpagne: tations respectively at the kth layer. Additionally for each layer, there
Conceptualization, Software, Writing – review & editing. Marat I. are separate trainable weight matrices, 𝐖(𝑘) , 𝐖(𝑘)
(𝑘)
∈ R𝑑 ×𝑑
(𝑘−1)
, and
self neigh
Latypov: Conceptualization, Writing – review & editing. Irene J. (𝑘)
biases 𝐛(𝑘) ∈ R𝑑 , which linearly transform the current node and
Beyerlein: Project administration, Funding acquisition, Writing –
aggregated neighborhood representations respectively. In this work,
review & editing.
neighborhood representations are calculated via element-wise mean
aggregation as,
Declaration of competing interest
( )
(𝑘)
∑
(𝑘−1)
The authors declare that they have no known competing finan- 𝐡 (𝑣) ← mean 𝐡𝑢 (B.3)
𝑢∈ (𝑣)
cial interests or personal relationships that could have appeared to
influence the work reported in this paper. where {𝐡(𝑘−1)
𝑢 , 𝑢 ∈  (𝑣)} represent all immediate node neighbors within
a 1-hop distance. This contrasts with the original GraphSAGE algorithm
Data availability that sub-samples neighbors to reduce training times of exceedingly
large graphs. Sub-sampling is not needed here as the microstructure
Data and codes necessary to reproduce these findings can be ac- graphs are small, averaging only 200 nodes each. This reduces our
cessed via https://github.com/jonathanhestroffer/PolyGRAPH. update equation to a simple extension of the famous self-loop graph
convolutional network (GCN) introduced by Kipf et al. [56]. The exten-
Acknowledgment sion being that the weights applied to a node and its neighbors are no
longer shared as in the classic GCN, rather they are learned separately;
This work is funded by the U.S. Dept. of Energy, Office of Basic this increases the potential expressivity of the GNN [58].
Energy Sciences Program DE-SC0018901. After updating, node representations calculated in Eq. (B.5) undergo
𝓁 2 normalization and the final graph embedding vector of our GNN,
Appendix A (𝐾)
𝐱 ∈ R𝑑 , is calculated during graph-readout at layer depth 𝐾 by
taking the element-wise average of all final node representations,
1 ∑ (𝐾)
Texture classes A through L 𝐱 ← 𝐡 (B.4)
|| 𝑣∈ 𝑣
Appendix B which is then passed through two additional fully-connected layers to
the final output layer. ReLU nonlinearity is applied after every fully-
connected and message-passing layer except for the output. Models for
Details of the GNN both stiffness and yield strength prediction were based on the GNN
model architecture described above.
Let 𝑘 = 1, … , 𝐾 denote the message-passing layer index. Be-
fore graph convolution layers are applied, the fully-connected pre-
Details of the ANN
processing layer transforms input node features, 𝐱𝑣 , into initial node
representations, 𝐡(0)
𝑣 , (i.e. 0th layer representation). With the graphs
now prepared for convolution, the representations of each node are The GNN and ANN differ only in their message-passing layer update
updated at each message-passing layer according to the following equations, with the ANN ignoring neighborhood aggregation. In the
update equation, case of the ANN, the message-passing layer update equation becomes,
( ) ( )
(𝑘)
𝐡(𝑘)
𝑣 ← 𝜎 𝐖self ⋅ 𝐡𝑣
(𝑘−1)
+ 𝐖(𝑘)
neigh
⋅ 𝐡(𝑘)
 (𝑣)
+ 𝐛(𝑘) (B.2) 𝐡(𝑘) (𝑘)
𝑣 ← 𝜎 𝐖self ⋅ 𝐡𝑣
(𝑘−1)
+ 𝐛(𝑘) (B.5)

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.

Training and evaluation Table B.1

Hyper-parameters of the GNN. Listed are sets of possible values for each as well as the
optimal values determined via Bayesian optimization using SigOpt.
All fully-connected and message-passing layers have learnable
Hyper-parameters Possible values Optimal value
weights and biases associated with them that must be trained. Weights
NFL1 [8, 16, 32] 32
of the GNN were initialized using Kaiming uniform initialization [59]
NMPL1 = NMPL2 [16, 32, 64] 64
and biases were initialized from the uniform distribution,  (0, 1). NFL2 [16, 32, 64, 128] 64
Parameters of the model were trained using an adaptive learning rate NFL3 [8, 16, 32, 64] 16
optimization algorithm (Adam) [60], implemented in PyTorch [61] batch size [16, 32, 64] 16
using a mean-squared error (MSE) loss function, learning rate [1, 5, 10, 50] ×10−4 10 × 10−4
weight decay [1, 5, 10, 50] ×10−5 1 × 10−5
1 ∑(
𝑛
)2 number of epochs [100, 150, 200, 300, 400, 600] 600
MSE = 𝑦 − 𝑦̂𝑖 (B.6)
𝑛 𝑖=1 𝑖

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-
figuration, results of these experiments are provided in Fig. B.1. Once
all experiments were completed, the top candidate model was selected
for a more extensive analysis of its respective training and validation
[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
linkages in high contrast composites from simulation datasets, Comput. Mater.
Sci. 151 (2018) 278–287, http://dx.doi.org/10.1016/j.commatsci.2018.05.014,
URL https://www.sciencedirect.com/science/article/pii/S0927025618303215.

7
J.M. Hestroffer et al.
[8] X. Li, Z. Liu, S. Cui, C. Luo, C.-F. Li, Z. Zhuang, Predicting the effective
mechanical property of heterogeneous materials by image based modeling and
deep learning, Comput. Methods Appl. Mech. Engrg. 347 (2019) http://dx.doi.
org/10.1016/j.cma.2019.01.005.
[9] C. Herriott, A.D. Spear, Predicting microstructure-dependent mechanical prop-
erties in additively manufactured metals with machine- and deep-learning
methods, Comput. Mater. Sci. 175 (2020) 109599, http://dx.doi.org/10.1016/
j.commatsci.2020.109599, URL https://www.sciencedirect.com/science/article/
pii/S0927025620300902.
[10] X. Liu, Z. Yan, Z. Zhong, Predicting elastic modulus of porous
La0.6Sr0.4Co0.2Fe0.8O3-𝛿 cathodes from microstructures via FEM
and deep learning, Int. J. Hydrogen Energy 46 (42) (2021) 22079–
22091, http://dx.doi.org/10.1016/j.ijhydene.2021.04.033, URL https:
//www.sciencedirect.com/science/article/pii/S0360319921013379.
[11] N.H. Paulson, M.W. Priddy, D.L. McDowell, S.R. Kalidindi, Reduced-order
structure-property linkages for polycrystalline microstructures based on 2-
point statistics, Acta Mater. 129 (2017) 428–438, http://dx.doi.org/10.1016/
j.actamat.2017.03.009, URL https://www.sciencedirect.com/science/article/pii/
S135964541730188X.
[12] A. Gupta, A. Cecen, S. Goyal, A.K. Singh, S.R. Kalidindi, Structure–property link-
ages using a data science approach: Application to a non-metallic inclusion/steel
composite system, Acta Mater. 91 (2015) 239–254.
[13] M.I. Latypov, S.R. Kalidindi, Data-driven reduced order models for effective
yield strength and partitioning of strain in multiphase materials, J. Comput.
Phys. 346 (2017) 242–261, http://dx.doi.org/10.1016/j.jcp.2017.06.013, URL
https://www.sciencedirect.com/science/article/pii/S0021999117304588.
[14] M.I. Latypov, L.S. Toth, S.R. Kalidindi, Materials knowledge system for nonlinear
composites, Comput. Methods Appl. Mech. Engrg. 346 (2019) 180–196, http:
//dx.doi.org/10.1016/j.cma.2018.11.034, URL https://www.sciencedirect.com/
science/article/pii/S0045782518305930.
[15] S.R. Kalidindi, A Bayesian framework for materials knowledge systems, MRS
Commun. 9 (2) (2019) 518–531.
[16] S.R. Kalidindi, Feature engineering of material structure for AI-based materials
knowledge systems, J. Appl. Phys. 128 (4) (2020) 041103.
[17] Z. Yang, Y.C. Yabansu, D. Jha, W. keng Liao, A.N. Choudhary, S.R. Kalidindi,
A. Agrawal, Establishing structure-property localization linkages for elastic
deformation of three-dimensional high contrast composites using deep learn-
ing approaches, Acta Mater. 166 (2019) 335–345, http://dx.doi.org/10.1016/
j.actamat.2018.12.045, URL https://www.sciencedirect.com/science/article/pii/
S1359645418309960.
[18] C. Rao, Y. Liu, Three-dimensional convolutional neural network (3D-
CNN) for heterogeneous material homogenization, Comput. Mater. Sci. 184
(2020) 109850, http://dx.doi.org/10.1016/j.commatsci.2020.109850, URL https:
//www.sciencedirect.com/science/article/pii/S0927025620303414.
[19] R. Pokharel, A. Pandey, A. Scheinker, Physics-informed data-driven surrogate
modeling for full-field 3D microstructure and micromechanical field evolution
of polycrystalline materials, JOM 73 (2021) http://dx.doi.org/10.1007/s11837-
021-04889-3.
[20] D. Montes de Oca Zapiain, E. Popova, F. Abdeljawad, J.W. Foulk, S.R. Kalidindi,
H. Lim, Reduced-order microstructure-sensitive models for damage initiation
in two-phase composites, Integr. Mater. Manuf. Innov. 7 (3) (2018) 97–115,
http://dx.doi.org/10.1007/s40192-018-0112-0.
[21] A. Marshall, S.R. Kalidindi, Autonomous development of a machine-learning
model for the plastic response of two-phase composites from micromechanical
finite element models, JOM 73 (7) (2021) 2085–2095, http://dx.doi.org/10.
1007/s11837-021-04696-w.
[22] D.M. de Oca Zapiain, S.R. Kalidindi, Localization models for the plastic response
of polycrystalline materials using the material knowledge systems framework,
Modelling Simulation Mater. Sci. Eng. 27 (7) (2019) 074008, http://dx.doi.org/
10.1088/1361-651x/ab37a5.
[23] M.M. Bronstein, J. Bruna, Y. LeCun, A. Szlam, P. Vandergheynst, Geometric deep
learning: Going beyond Euclidean data, IEEE Signal Process. Mag. 34 (4) (2017)
18–42, http://dx.doi.org/10.1109/msp.2017.2693418.
[24] W.C. Lenthe, M.P. Echlin, J.C. Stinville, M. De Graef, T.M. Pollock, Twin related
domain networks in René 88DT, Mater. Charact. 165 (2020) 110365, http://dx.
doi.org/10.1016/j.matchar.2020.110365, URL https://www.sciencedirect.com/
science/article/pii/S1044580320305647.
[25] H.F. Poulsen, Three-Dimensional X-Ray Diffraction Microscopy: Mapping Poly-
crystals and their Dynamics, Vol. 205, Springer Science & Business Media,
2004.
[26] H.F. Poulsen, An introduction to three-dimensional X-ray diffraction microscopy,
J. Appl. Crystallogr. 45 (6) (2012) 1084–1097, http://dx.doi.org/10.1107/
S0021889812039143.
[27] M.P. Miller, M. Obstalecki, E. Fontes, D.C. Pagan, J.P.C. Ruff, A.J. Beaudoin,
Insit@CHESS, a resource for studying structural materials, Synchrotron Radiat.
News 30 (3) (2017) 4–8, http://dx.doi.org/10.1080/08940886.2017.1316124.
[28] D.C. Pagan, K.E. Nygren, M.P. Miller, Analysis of a three-dimensional slip
field in a hexagonal Ti alloy from in-situ high-energy X-ray diffraction mi-
croscopy data, Acta Mater. 221 (2021) 117372, http://dx.doi.org/10.1016/
j.actamat.2021.117372, URL https://www.sciencedirect.com/science/article/pii/
S1359645421007515.
8
Computational Materials Science 217 (2023) 111894
[29] Y. Koren, R. Bell, C. Volinsky, Matrix factorization techniques for recommender
systems, Computer 42 (8) (2009) 30–37, http://dx.doi.org/10.1109/MC.2009.
263.
[30] W. Jiang, J. Luo, Graph neural network for traffic forecasting: A survey,
2021, http://dx.doi.org/10.48550/ARXIV.2101.11174, URL https://arxiv.org/
abs/2101.11174.
[31] D. Duvenaud, D. Maclaurin, J. Aguilera-Iparraguirre, R. Gómez-Bombarelli,
T. Hirzel, A. Aspuru-Guzik, R.P. Adams, Convolutional networks on graphs
for learning molecular fingerprints, in: Proceedings of the 28th International
Conference on Neural Information Processing Systems - Volume 2, NIPS ’15,
MIT Press, Cambridge, MA, USA, 2015, pp. 2224–2232.
[32] J. Zhou, G. Cui, S. Hu, Z. Zhang, C. Yang, Z. Liu, L. Wang, C. Li, M.
Sun, Graph neural networks: A review of methods and applications, AI Open
1 (2020) 57–81, http://dx.doi.org/10.1016/j.aiopen.2021.01.001, URL https://
www.sciencedirect.com/science/article/pii/S2666651021000012.
[33] P. Veličković, G. Cucurull, A. Casanova, A. Romero, P. Liò, Y. Bengio, Graph
Attention Networks, in: International Conference on Learning Representations,
2018, accepted as poster. URL https://openreview.net/forum?id=rJXMpikCZ.
[34] M. Sun, S. Zhao, C. Gilvary, O. Elemento, J. Zhou, F. Wang, Graph convolutional
networks for computational drug development and discovery, Brief. Bioinform.
21 (3) (2019) 919–935, http://dx.doi.org/10.1093/bib/bbz042, arXiv:https://
academic.oup.com/bib/article-pdf/21/3/919/33227266/bbz042.pdf.
[35] M. Zhang, P. Li, Y. Xia, K. Wang, L. Jin, Revisiting graph neural networks for
link prediction, 2021, URL https://openreview.net/forum?id=8q_ca26L1fz.
[36] N.N. Vlassis, R. Ma, W. Sun, Geometric deep learning for computational
mechanics part I: Anisotropic hyperelasticity, Comput. Methods Appl. Mech.
Engrg. 371 (2020) 113299, http://dx.doi.org/10.1016/j.cma.2020.113299, URL
https://www.sciencedirect.com/science/article/pii/S0045782520304849.
[37] M. Dai, M.F. Demirel, Y. Liang, J.-M. Hu, Graph neural networks for an accurate
and interpretable prediction of the properties of polycrystalline materials, Npj
Comput. Mater. 7 (1) (2021) 103, http://dx.doi.org/10.1038/s41524-021-00574-
w.
[38] C. Shu, J. He, G. Xue, C. Xie, Grain knowledge graph representation learning: A
new paradigm for microstructure-property prediction, Crystals 12 (2) (2022) http:
//dx.doi.org/10.3390/cryst12020280, URL https://www.mdpi.com/2073-4352/
12/2/280.
[39] D.C. Pagan, C.R. Pash, A.R. Benson, M.P. Kasemer, Graph neural network
modeling of grain-scale anisotropic elastic behavior using simulated and mea-
sured microscale data, 2022, http://dx.doi.org/10.48550/ARXIV.2205.06324,
URL https://arxiv.org/abs/2205.06324.
[40] S.G. Kobourov, Spring embedders and force directed graph drawing algorithms,
2012, http://dx.doi.org/10.48550/ARXIV.1201.3011, URL https://arxiv.org/abs/
1201.3011.
[41] H. Zhou, P. Xu, X. Yuan, H. Qu, Edge bundling in information visualization,
Tsinghua Sci. Technol. 18 (2) (2013) 145–156, http://dx.doi.org/10.1109/TST.
2013.6509098.
[42] J. You, R. Ying, J. Leskovec, Design space for graph neural networks,
2020, http://dx.doi.org/10.48550/ARXIV.2011.08843, URL https://arxiv.org/
abs/2011.08843.
[43] M. Fey, J.E. Lenssen, Fast graph representation learning with PyTorch geo-
metric, 2019, http://dx.doi.org/10.48550/ARXIV.1903.02428, URL https://arxiv.
org/abs/1903.02428.
[44] W.L. Hamilton, R. Ying, J. Leskovec, Inductive representation learning on large
graphs, NIPS (2017).
[45] M.W. Priddy, N. Paulson, Synthetic alpha-ti microstructures and associated elastic
stiffness and yield strength properties, 2016, URL https://matin.gatech.edu/
resources/52. (Accessed Jan 2022).
[46] M.W. Priddy, N. Paulson, D. McDowell, S.R. Kalidindi, Synthetic alpha-ti
microstructures and associated elastic stiffness and yield strength properties
- extended, 2017, URL https://matin.gatech.edu/resources/187. (Accessed Jan
2022).
[47] M. Peters, G. Lütjering, G. Ziegler, Control of microstructures of (𝛼+ 𝛽)-titanium
alloys, Int. J. Mater. Res. 74 (5) (1983) 274–282.
[48] M. Peters, A. Gysler, G. Lütjering, Influence of texture on fatigue properties of
Ti-6Al-4V, Metall. Mater. Trans. A 15 (8) (1984) 1597–1605, http://dx.doi.org/
10.1007/bf02657799.
[49] G. Lütjering, Influence of processing on microstructure and mechanical properties
of (𝛼+ 𝛽) titanium alloys, Mater. Sci. Eng. A 243 (1–2) (1998) 32–45.
[50] Y. Wang, J. Huang, Texture analysis in hexagonal materials, Mater. Chem. Phys.
81 (1) (2003) 11–26, http://dx.doi.org/10.1016/S0254-0584(03)00168-8, URL
https://www.sciencedirect.com/science/article/pii/S0254058403001688.
[51] G. Lütjering, J. Williams, Titanium, in: Engineering Materials and Processes,
Springer Berlin Heidelberg, 2007, URL https://books.google.com/books?id=
41EqJFxjA4wC.
[52] B.D. Smith, Microstructure-Sensitive Plasticity and Fatigue of Three Titanium
Alloy Microstructures (Ph.D. thesis), Georgia Institute of Technology, 2013.
[53] D. Montes de Oca Zapiain, A. Shanker, S.R. Kalidindi, Convolutional
Neural Networks for the Localization of Plastic Velocity Gradient
Tensor in Polycrystalline Microstructures, J. Eng. Mater. Technol. 144
(1) (2021) http://dx.doi.org/10.1115/1.4051085, 011004, arXiv:https:
//asmedigitalcollection.asme.org/materialstechnology/article-pdf/144/1/
011004/6701898/mats_144_1_011004.pdf.
J.M. Hestroffer et al.
[54] A.L. Frankel, C. Safta, C. Alleman, R. Jones, Mesh-based graph convolutional
neural networks for modeling materials with microstructure, J. Mach. Learn.
Model. Comput. 3 (1) (2022) 1–30.
[55] R. Hanocka, A. Hertz, N. Fish, R. Giryes, S. Fleishman, D. Cohen-Or, MeshCNN,
ACM Trans. Graphics 38 (4) (2019) 1–12, http://dx.doi.org/10.1145/3306346.
3322959.
[56] T.N. Kipf, M. Welling, Semi-supervised classification with graph convolutional
networks, 2016, http://dx.doi.org/10.48550/ARXIV.1609.02907, URL https://
arxiv.org/abs/1609.02907.
[57] J.C. Stinville, J.M. Hestroffer, M.A. Charpagne, A.T. Polonsky, M.P. Echlin, C.J.
Torbet, V. Valle, K.E. Nygren, M.P. Miller, O. Klaas, A. Loghin, I.J. Beyerlein,
T.M. Pollock, Multi-modal dataset of a polycrystalline metallic material: 3D
microstructure and deformation fields, Sci. Data 9 (1) (2022) 460.
9
Computational Materials Science 217 (2023) 111894
[58] W.L. Hamilton, Graph representation learning, Synth. Lect. Artif. Intell. Mach.
Learn. 14 (3) (2020) 1–159.
[59] K. He, X. Zhang, S. Ren, J. Sun, Delving deep into rectifiers: Surpassing
human-level performance on ImageNet classification, in: 2015 IEEE International
Conference on Computer Vision, ICCV, 2015, pp. 1026–1034.
[60] D. Kingma, J. Ba, Adam: A method for stochastic optimization, in: International
Conference on Learning Representations, 2014.
[61] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin,
N. Gimelshein, L. Antiga, A. Desmaison, A. Köpf, E. Yang, Z. DeVito, M. Raison,
A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, S. Chintala, PyTorch: An
imperative style, high-performance deep learning library, 2019, http://dx.doi.
org/10.48550/ARXIV.1912.01703, URL https://arxiv.org/abs/1912.01703.
[62] S. Clark, P. Hayes, SigOpt web page, 2019, URL https://sigopt.com.
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.

A GREAT DEMONSTRATION—A NEW TYPE OF OVEN—ST

MUNGO HALLS—TEAROOMS BURNED DOWN—MR
M‘CULLOCH RETIRES; MR GERRARD ELECTED
PRESIDENT—AN AMALGAMATION PROPOSAL—PURVEY
AND CONTRACT DEPARTMENT—THE FIRST MOTOR VAN
—A BREAD EXPERIMENT—PROPAGANDA AND PUBLICITY
—UNEMPLOYMENT—THE EFFECT ON CO-OPERATION—
CO-OPERATION THE REMEDY—SOME BUSINESS ITEMS—
THE SOCIETY’S POSITION.

In tracing the history of the two branches at Clydebank and Belfast

we have wandered far ahead of the general development of the
Society, and must now retrace our steps to the main road of our story
at the point at which we left it. In the period which began with the
beginning of the thirty-third year the directors were kept very busy
with business which was connected with the development of the two
new branches, but the parent institution was not allowed to suffer
from neglect. The great demonstration which was held in the
summer of 1902 to celebrate the attainment of an average output of
3,000 sacks per week proved a great success. The demonstration was
in two parts. In the first place, the employees and their families and
friends, to a number which filled three special trains, went picnicing
in the morning to Milngavie. In the second place, a gigantic vehicular
procession, which included about 100 vehicles and 150 horses,
paraded through the city and out the Milngavie Road as far as
Canniesburn Toll, returning to M‘Neil Street by another route. This
procession was headed by a charabanc containing the members of
the Society’s band, and the gaily decorated lorries and vans attracted
much attention as they passed through the streets. All the mottoes
displayed in the procession had some reference to Co-operation.
But, while it is good to demonstrate and advertisement has
certainly its uses, bread must also be baked, and the object of a Co-
operative society is, or ought to be, to produce the best possible
bread at a minimum cost. This was a fact of which there was not
much danger that the directors would lose sight, and even if they did
the representatives of the societies would very quickly remind them
of the lapse. They were about to build a new bakery; and, with the
view of testing the efficiency of a type of oven then practically
unknown in Scotland, they decided to erect two draw-plate ovens at
M‘Neil Street. The tests seem to have been satisfactory, for in both
Clydebank and Belfast branches this type of oven formed the
majority of those installed.
 ST MUNGO HALLS.
Meantime other matters were engaging their attention. When the
new stables had been erected on the land purchased at the south
corner of South York Street and Govan Street a considerable portion
of the ground—more than half indeed—remained unbuilt on, and
early in 1902 the educational committee of the Society came forward
with the recommendation that in any further building which might
be erected a hall which could be used for educational purposes
should be included. The committee also were desirous that the
Society should have a hall of their own, and later in the same year the
architects were instructed to prepare plans for the utilisation of this
vacant land which should include halls and accommodation for the
headquarters of the purvey department. At the quarterly meeting in
March 1903 power was given to proceed with the erection of the
buildings. These were to consist of five halls, containing
accommodation for from 200 to 1,500 people, with the necessary
siderooms, etc., and with ample accommodation for the purvey
department. Permission was also given for the erection of three
tenements of dwellinghouses, with shops on the ground floor; the
total cost to be from £14,000 to £15,000 for the halls and purvey
department buildings, and £6,400 for the tenements. Eleven months
passed, however, before the building of the halls was begun, and they
were not completed until 1906.
Toward the end of 1903 the Paisley Road tearooms were destroyed
by fire, and in restoring them the committee decided to add another
storey. The landlord agreed to bear a proportion of the cost, and on
the reconstruction being completed they were named the
“Wheatsheaf” tearooms.
 MR M‘CULLOCH RETIRES.
For several years the chairman had been desirous of retiring, but
had been induced to remain in office until the schemes on which he
had set his heart—the erection of Clydebank and Belfast branches—
were well on the way. In 1904, however, he definitely decided to end
his official connection with the Society. For fifteen years he had acted
as president. He was first appointed to the board by his society in
June 1888, and had continued to act on the board until his retiral.
Mr Daniel H. Gerrard was elected president in succession to Mr
M‘Culloch. For a short time he had represented St George Society on
the board prior to his election, while the active part which he had
taken, as a member of the Scottish Sectional Board, in combating the
boycott in the later “’nineties” and his position as chairman of St
George Society had made him well known to the majority of Scottish
Co-operators.
 AN AMALGAMATION PROPOSAL.
In the autumn of 1903 an interesting proposal was made by the
directors of Hamilton Baking Society. This was to the effect that the
two societies should become amalgamated. During the following year
several meetings took place between representatives of the two
societies, but ultimately negotiations were broken off, the members
of Hamilton Baking Society having expressed disapproval of the
proposal. If the scheme could have been carried through it would
have left the Baking Society in the position of being the only
federated society producing bread in the West, with the exception of
Chapelhall, and might have paved the way for that society to come in
also.
 PURVEY AND CONTRACT DEPARTMENT.
Meantime the work of the purvey department was growing. For the
two years 1902 and 1903 the department secured contracts from
Glasgow Corporation for the supply of bags of eatables on Children’s
Day. In the aggregate the bags supplied numbered 190,000, of a total
value of about £1,800, and requiring somewhere about twenty-five
tons of flour for their manufacture. When the “Wheatsheaf”
tearooms were reopened after the fire an “at home” was held to
which the customers were invited. The purvey department was also
an offerer on most occasions when purveys on a large scale had to be
carried through, often with success. The department was successful
in securing the purveying contract for the tearooms in the East-End
Exhibition. It was also a successful offerer in 1904 for the supply of
bread to Gailes Territorial Camp, and in the following year
Jamestown Camp was supplied from Clydebank.
The tearooms continued to give cause for a considerable amount of
anxiety on the part of the committee. Sometimes a small profit was
made on the working and sometimes a loss resulted, but there never
was that increase in trade for which the board thought they had a
right to look. With the erection of St Mungo Hall and the
transference of the catering headquarters to South York Street the
need for Main Street as a depot for the catering department
disappeared, but for some time it was carried on as a workmen’s
tearoom, always without any signs of assured success, however.
Finally, in 1904, it was given up altogether. Nor were the London
Street halls or the Union rooms much more successful. Several
experiments were made at London Street, with the object of making
the place more popular. The whole of the three upper flats were
taken by the Society, and several trade union and friendly society
branches made the rooms their headquarters, but the place was
never really popular. An attempt which was made to run the second
flat as workmen’s dining and tea rooms did not meet with much
success, and it seemed as if anything the board could do was not of
much use in popularising the place. Nor were the Union rooms much
more successful. They just managed to keep going, but they did not
become, as had been hoped, a rendezvous for the men and women of
the Co-operative movement. Nevertheless the committee did not
despair. They always kept on hoping that the tide would turn and
renewed the lease time after time, still looking for the Co-operative
patronage which never came in sufficient quantity to make the place
a success.
 THE FIRST MOTOR VAN.
For practically the whole of the nineteenth century the power-
driven road vehicle had to struggle against the bigotry of the people
and the interested opposition of the proprietors of other systems of
locomotion. Steam-driven road cars and wagons were in use long
before Stephenson had designed and completed his first railway, but
they met with strong and unreasoning opposition on the part of
many people instigated by the owners of post and passenger coaches,
who saw in the new method of locomotion a menace to their welfare,
and also by other horse owners, who found these steam-driven, noisy
vehicles intolerable nuisances which frightened into panic the high-
strung horses unfortunate enough to meet them on the roads. This
opposition translated itself into Acts of Parliament which imposed
heavy taxation on the newer mode of locomotion, and culminated in
1836 in the famous “man with the red flag” provision in the Act of
that year, which was the means of clearing practically every “faster
than walking pace” power-driven vehicle off the roads. This Act,
passed at the instance and in the interests of the railways, had a most
detrimental effect on the development of road and also of canal
traffic, and left the country completely at the mercy of the railway
companies until the “red flag” restriction was practically laughed out
of existence in the last decade of the century by the development of
petrol-driven motors, and the prohibition was removed in 1896.
After 1896 the development of power-driven vehicles for road
traffic made rapid strides, but it was not until some seven years later
that the U.C.B.S., on the instigation of the makers, put a bread motor
van on the road for a trial. After working for some months, the
committee decided to send a note of their experience of its working
to the makers, and at the same time to point out some defects which
they had discovered. Gradually, however, the new means of delivery
superseded the old, until all the Society’s long-distance work was
done by means of motor vans. For many years, however, the
construction of the motor engines placed a serious obstacle in the
way of the adoption of motor vans for short journeys or for journeys
which entailed frequent stops. The engine, from its nature, requires
that the cylinder should be charged and the charge compressed
before ignition can take place. Drivers were therefore faced with the
alternatives of putting the engine out of gear and leaving it running—
at a considerable expenditure for fuel which gave no return in work
done—or of having much laborious cranking for the purpose of
charging the cylinder and compressing and igniting the gas, the latter
plan, particularly in cold weather, often entailing considerable
expenditure of time as well as of energy. Nowadays, however, most
up-to-date motor engines have a small electric engine for doing this
work.
 A BREAD EXPERIMENT.
For a long time it had been the wish of the manager and directors
of the Baking Society to introduce a natural working day into the
bread trade, and they had done their best some years earlier to have
the matter taken up by the societies and the public, but without
success. Now, in the autumn of 1905, another attempt was made, a
squad of bakers beginning work at 9 a.m.; the bread thus baked
being delivered on the following morning. A number of the societies
in Glasgow and neighbourhood were induced to take up the sale of
this bread, with the result that the sales speedily rose; but after a trial
which lasted several months the scheme was abandoned, as the
directors found that instead of helping the general work of delivery it
was proving a hindrance to that work.
This decision of the committee met with a considerable amount of
opposition from delegates to the quarterly meeting, but the directors
were not in a position to do other than they had done. The real
obstacle was to be found in the fact that the public insisted on having
new bread, and with a number of the bakers beginning work late in
the day for the production of “natural working day” bread, as it had
been named, there was not labour enough in the early morning to
meet the demand for new bread by ten o’clock. Thus the second
attempt of the Baking Society directors to introduce a natural
working day into the baking trade came to naught, through no fault
of theirs, but because of a public who would insist on having bread
steaming from the ovens.
 PROPAGANDA AND PUBLICITY.
During these years the directors continued to adopt every
reasonable method of keeping the Federation and its productions
before the public. They took advantage of every opportunity offered
by exhibitions to show the quality of goods which were produced by
the Society; they exhibited also at shows, and were occasionally
successful in securing prizes, although the rush conditions under
which their bread had to be produced in order to cope with the ever-
increasing demands of the societies did not provide the conditions
necessary for producing bread of show quality.
At the same time the entertaining of women’s guilds, conference
associations, and other Co-operative organisations was further
developed, and many Co-operators who before had but a very hazy
idea of the size and importance of the U.C.B.S. were enlightened as a
result of these visits. The cake shows, also, held annually, were
excellent propaganda. Here, inside one hall, were to be found
specimens of all the productions of the three bakeries; and these
productions came latterly to include many novelties which were not
to be found in everyday use, as well as the more common specimens
of cakes, buns, bread, etc., and a large variety of biscuits. Thus the
cake show came to be regarded as the annual exhibition of Bakery
productions, which indeed it was, and it was one of the most eagerly
anticipated Co-operative functions of the year.
 UNEMPLOYMENT.
Unfortunately, during the years 1904–1905 there was a very severe
economic crisis which was the cause of widespread unemployment,
and Glasgow and the West suffered severely. Such economic crises
had been periodic during the past century, and differed materially
from those which occurred in earlier centuries. Until the beginning
of the nineteenth century economic crises were really due to actual
scarcity of foodstuffs, and people went hungry as much because there
was not enough food to go round as because they had no money with
which to buy their share of what there was. Thus, in Scotland, until
about the close of the Napoleonic wars, the periods of comparative
plenty or of scarcity depended largely on the character of the
seasons. While it is true that the common people were always in
want, in bad seasons they died of starvation.
With the coming of the industrial era, however, there took place a
gradual change. While prices still depended on the seasons, and were
moderately or extravagantly high as the seasons were good or bad,
there entered into the problem a new factor, and the people became
poor and were unable to purchase because they had produced too
much and there was not an effective market for the goods. As it has
been tersely put: “The shoemaker’s children went barefoot because
their father had produced too many boots; and the tailor’s, naked,
because he had made too many clothes.”
This was a new phenomenon for which history provides no
parallel, and it has persisted, ever increasing in intensity, until
towards the end of the first decade of the twentieth century as many
as 42 per cent. of the breadwinners in a respectable working-class
district in the East-End of Glasgow have been found to be
unemployed at the same time.
These periods of unemployment seem to be in the form of more or
less regularly recurring cycles. There is first a gradually increasing
inflation of the volume of trade. New works are started, old works are
enlarged, and everywhere there is a boom, until the zenith is
reached. Then comes a gradual slackening off. The supply of goods
has outstripped the limits of effective demand and sales gradually
decline as warehouses become full. The rate of slackening increases;
statisticians begin to watch the rapidly ascending unemployment
curve, which in itself does not record the full slackening, as many
workers are on short time. This increase of unemployment still
further weakens the effective demand for goods and still further
accelerates the growth of unemployment, with the result that in
about another year the unemployment crisis is reached, soup
kitchens and relief works open; thousands of hitherto steady
workmen become derelicts—and the huge commercial and industrial
concerns take advantage of the slackness to squeeze small rivals out
of business or to swallow them up, while at the same time improving
their own machinery of production.
Another feature of these industrial phenomena is the acceleration
of their periodicity. During the decades which marked the end of the
nineteenth century and the beginning of the twentieth, the number
of years which elapsed between one crisis and the next was becoming
smaller and the periods were appreciably shorter than those of the
earlier cycles. Fourteen, twelve, or ten years used to be the intervals
between crises, but in the last forty years this interval has gradually
shortened, until only three years intervened between the end of one
crisis and the beginning of the next. This is in strict accordance with
what we would expect to find. The means of supplying the effective
demand of the people for goods has increased out of all proportion to
the increase in that effective demand, with the natural result that the
market gets choked full to overflowing in an ever-shortening time.
When to this is added the continuous perfecting of processes and
equally continuous speeding up of production, we see that under the
conditions which obtain it necessarily follows that the periods which
must elapse before the markets are choked up with goods must get
progressively shorter, while the periods necessary to relieve the glut
again grow gradually longer. Thus, we had a crisis in the late
“’seventies,” another in 1888–89; another in 1893, of a milder type;
another in 1898; another, very acute, in 1904–05; and another, the
worst the industrial world in Britain has known, in 1908–09–10;
while evidences were not wanting when war broke out that we were
again on the downward sweep of the cycle.
 THE EFFECT ON CO-OPERATION.
The Co-operative movement in general is profoundly influenced by
conditions of unemployment, which is our excuse for what has been
written above. The vast majority of Co-operators depend for their
incomes on their being employed in ordinary industrial
undertakings, and when the machinery of industry breaks down their
spending power is affected. Those in fortunate positions can carry on
economically for a longer or shorter period on their investments in
their societies, but those fortunate individuals constitute but a small
proportion of the whole membership. When an unemployment crisis
comes it means distress to a large number of Co-operators, for those
who have large savings to fall back on are usually in more or less
important positions in the concerns by which they are employed and
are amongst the last to be dismissed—are, in fact, rarely out of
employment at all. Sales of Co-operative societies go down, while
working expenses continue. In this respect, however, the Baking
Society has always been the last to suffer, for people continue to buy
bread when they have almost ceased to be able to purchase anything
else. But there is always during such crises a proportion of Co-
operators, varying in different districts, whose incomes are barely
sufficient to meet physical needs while they are in constant
employment and who are thrown on their beam ends by even a
fortnight of unemployment, and the hardship of these crises is that
this is the class of people on whom the curse of unemployment falls
first and on whom it rests longest.
Fortunately, the milk of human kindness is not quite dried up in
Co-operators, and so soon as genuine cases of distress are known
steps are taken to ameliorate the condition. In this process of
amelioration the Baking Society has always played a big part. We
have already seen that during the miners’ and engineers’ strikes and
during periods of distress due to unemployment the Society
distributed thousands of pounds worth of bread, and now, during
1904–05, the same policy was pursued.