Communication

[Link]

Bayesian Machine Learning in Metamaterial Design: Fragile

Becomes Supercompressible
Miguel A. Bessa,* Piotr Glowacki, and Michael Houlder

data-driven approach to explore new

Designing future-proof materials goes beyond a quest for the best. The next design possibilities, while reducing experi-
generation of materials needs to be adaptive, multipurpose, and tunable. This mentation to a minimum (validation).
is not possible by following the traditional experimentally guided trial-and- Without loss of generality, we illustrate
error process, as this limits the search for untapped regions of the solution this paradigm shift by focusing on a new
concept for a low density mechanical met-
space. Here, a computational data-driven approach is followed for exploring
amaterial (Figure  1). The aim is to addi-
a new metamaterial concept and adapting it to different target proper- tively manufacture a building block that
ties, choice of base materials, length scales, and manufacturing processes. achieves recoverable supercompressibility
Guided by Bayesian machine learning, two designs are fabricated at different while maintaining high strength and stiff-
length scales that transform brittle polymers into lightweight, recoverable, ness. This concept results from a combina-
tion of a deployable mast[9,10] developed for
and supercompressible metamaterials. The macroscale design is tuned for
highly deformable space structures, and a
maximum compressibility, achieving strains beyond 94% and recoverable thin-walled conical frustum[11] common in
strengths around 0.1 kPa, while the microscale design reaches recoverable impact absorption applications.
strengths beyond 100 kPa and strains around 80%. The data-driven code is Remarkably, even without inverting
available to facilitate future design and analysis of metamaterials and struc- the design process, additively manufac-
tures ([Link] tured mechanical metamaterials with low
densities (ρ ≪  1 g cm−3) have achieved
compressive strengths above 100 MPa
while exhibiting brittle fracture using
Structure-dominated materials (metamaterials) are pushing the carbon lattices,[12] on the order of 1 MPa with partially recover-
envelope of known electromagnetic,[1] classic,[2] and quantum[3] able compressive strains in excess of 50% using microlattices of
mechanical properties by exploring new geometries. Additive alumina[13] or graphene aerogels,[14] and on the order of 0.5 kPa
manufacturing has been a major driving force in this explora- and exceeding 80% strains for another graphene aerogel.[15]
tion since virtually any topology can be obtained to probe the These investigations demonstrate that tuning the geometry of
vast design space created by geometric changes in the mate- the metamaterial to explore symmetry-breaking instabilities
rial structure. This has led to discoveries of materials with new (buckling) is the key to achieve large effective compressibility
properties and functionality, for example exhibiting negative while causing small deformation of a high strength base mate-
linear compressibility,[4] tunable negative stiffness,[5] controllable rial. Yet, the process for designing these metamaterials and
reconfigurability via programmable surfaces[6] or via origami,[7] assessing their mechanical limits is cumbersome and time-con-
and symmetry breaking elastic surface patterns.[8] However, suming due to the absence of general design principles arising
metamaterial design has relied on extensive experimentation from highly nonlinear, unstable, and imperfection sensitive
and a trial-and-error approach where analytical or computa- responses that depend on many geometric parameters. More
tional models only provide a posteriori explanations. Here, we importantly, even when successful designs are found,[12–15]
argue in favor of inverting the process by using a computational tuning them for new applications requiring different properties
and functionality is not trivial. In this context, machine learning
can provide significant advantages to the design process.
Prof. M. A. Bessa, P. Glowacki, M. Houlder Machine learning and deep learning are already able to sur-
Department of Materials Science and Engineering pass the limits of a human mind in specific tasks. Algorithms
Delft University of Technology are capable of winning against human champions in increasingly
2628 CD Delft, The Netherlands
E-mail: [Link]@[Link]
difficult games,[16–18] they recognize faces with near human-level
accuracy,[19] and predict ratings of unseen movies based on past
The ORCID identification number(s) for the author(s) of this article
can be found under [Link] user selection.[20] These achievements are permeating to different
scientific disciplines such as materials science[21,22] and chemi­
© 2019 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, stry,[23] and recently they are reaching the design of electromag-
Weinheim. This is an open access article under the terms of the Creative netic metamaterials via nonprobabilistic machine[24] and deep
Commons Attribution License, which permits use, distribution and repro-
duction in any medium, provided the original work is properly cited. learning[25–28] methods, and even generative deep learning.[29]
However, metamaterials usually derive their unprecedented prop-
DOI: 10.1002/adma.201904845 erties from exploring imperfection-sensitive behavior because

Adv. Mater. 2019, 1904845 1904845  (1 of 6) © 2019 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
[Link] [Link]

exploration and exploitation. The Supporting Information

includes two sections (“Why Bayesian machine learning?” and
“Bayesian machine learning vs Bayesian optimization”) with addi-
tional references and considerations for the interested reader.
Our data-driven exploration follows a recent framework
created to design materials and structures.[32,33] This modular
framework integrates: 1) design of experiments to sample the
input variables, 2) efficient predictive analyses to generate the
output database, 3) machine learning to establish input–output
relationships from the database, and 4) optimization to deter-
mine optimum designs from the machine learning model.
Different methods can be chosen for each module depending
on the problem’s dimensionality, size of database needed for
the learning process, and whether the phenomena of interest
is probabilistic or deterministic. Our mechanical metamaterial
undergoes probabilistic responses due to buckling/postbuck-
ling mechanisms, and involves a moderate number of design
parameters that is expected to lead to large training datasets.
Recent Bayesian machine learning methods called sparse
Gaussian processes[30] have been developed to fulfill these
requirements while being sufficiently scalable. We considered
several algorithms, as detailed in the Supporting Information
(“Sparse Gaussian processes: scalability & accuracy”), and con-
cluded that the sparse Gaussian process regression (SGPR)
algorithm[34] was the most adequate for the regression tasks,
while the scalable variational Gaussian process classification
(SVGP)[35] was the most adequate for classification.
Data-driven design of the metamaterial concept is then
summarized in Figure  2. First, the building block (Figure 1) is
parameterized according to the design variables. The geometry is
defined by the top and bottom base diameters, D1 and D2, height
P, and four parameters that define the cross-section of the ver-
tical elements (longerons): the cross-sectional area A, moments
of inertia Ix and Iy, and torsional constant Jτ. In addition, since
the metamaterial is targeted to be reversibly deformable, any
chosen base material is defined by its elastic constants: the
Young’s modulus E and the shear modulus G. Given that this
is a nonlinear elastic mechanics problem, we know a priori that
the geometric variables can be scaled by one of the dimensions,
chosen here as D1, and that we can consider the ratio of elastic
Figure 1.  Supercompressible metamaterial building block with general- constants G/E. Finally, since no contact between the deformed
ized cross-section for the longerons (x is the radial direction; y is the
longerons is intended, their number is not a variable due to
tangential direction).
the principle of superposition, i.e., increasing the number of
longerons leads to a proportional increase of the compressive
they operate at extreme regimes. In these cases, Bayesian methods strength. All seven input variables and respective bounds of the
become important because they deal with noisy observations by design space are defined in Table S1, Supporting Information.
inferring a smooth average response and quantifying uncertainty. Different designs are then generated by sampling this
The key novelty of our work is the consideration of these methods bounded space according to a Sobol sequence[36] (Figure 2A),
to design and analyze a new mechanical metamaterial. since this design of experiments method facilitates the subse-
Broadly, there are two classes of Bayesian methods that can quent learning process.[32] The response of each design is then
be useful to analyze and design metamaterials: 1) Bayesian predicted via nonlinear finite element analyses (Figure 2B)
machine learning;[30] 2) Bayesian optimization.[31] Here, we that provide the complete buckling and postbuckling behavior
focus on Bayesian machine learning because our goal is to (Figure 2C) by the arc-length method.[33] Each simulation of
create a map of the solution space that can be used later for the virtual design predicts the effective compressive strains
robust optimization of different applications. However, if the and stresses of the metamaterial, from which two quantities
goal was to find only one optimum metamaterial for a specific of interest can be computed: critical buckling stress σcrit and
application, then Bayesian optimization methods are likely to elastic energy absorption Eabs. The critical buckling stress σcrit
be more effective because they seek the optimum solution by represents the point at which the instability starts, from which
adaptively sampling the solution space in a trade-off between the structure no longer behaves as linearly elastic. The elastic

Adv. Mater. 2019, 1904845 1904845  (2 of 6) © 2019 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
[Link] [Link]

Figure 2.  Data-driven design of supercompressible metamaterial building block using seven design variables.

energy absorption Eabs is defined as the area below the complete input interaction, i.e., if the values are close to 1, then that input
stress–strain response. These quantities of interest are stored strongly influences the output, and if close to zero, its influence
in a large database (over 100 000 simulations) that can then be is negligible. Figure 3A demonstrates that the cross-sectional area
analyzed by machine learning (Figure 2D). A of the longerons is irrelevant for both quantities of interest,
For this problem, both machine learning classification and that the ratio of elastic moduli G/E has negligible influence
and regression are useful. Classification allows to distinguish when compared with the moments of inertia of the longerons.
between relevant and irrelevant regions of the design space, as Also, independently of the cross-section shape, the moments of
many combinations of parameters lead to metamaterials that inertia are complementary because Ix affects more the critical
do not coil when subjected to compression (invalidating super- buckling stress, while Iy affects more the absorbed energy. Inter-
compressibility). For example, global deformation modes in estingly, the torsional constant Jτ has a strong influence on the
bending or local buckling phenomena of the longerons. Based critical buckling stress, but not on the absorbed energy (detailed
on the buckling modes predicted by eigenvalue finite element analyses in the Supporting Information).
analyses, we can classify the designs as: 1) coilable or 2) noncoil­ We can then restrict to manufacturing designs with circular
able. Figure 2D shows a projection of the design space obtained cross-sections for the longerons, reducing Ix, Iy, and Jτ to a
by the machine learning model, where the white region in the single independent variable: the cross-section diameter d. Since
contour plot corresponds to noncoilable designs. Then, we A and G/E are negligible, the remaining variables that compose
can limit the regression task to the coilable regions and pre- the design space are the height P and relative base dia­meters
dict how the quantities of interest are changing (see colored (D1  − D2)/D1. This reduces the design space to only three
region of Figure 2D). The color gradient in the figure represents dimensions, which are easy to visualize (Figure 3B). In addition,
changes of the average critical buckling stress (in kPa), and the when considering printable designs made of brittle base mate-
labeled black isolines show how the average energy absorp- rials (PLA), the local strains at the longerons need to be pre-
tion of the different designs evolves in the same projection of dicted throughout the complete deformation process in order
the design space (in kJ m−3). We refer to Figures S7–S9, Sup- to know if the metamaterial is undergoing plasticity or fracture.
porting Information where different projections of the design We conducted elementary experimental tests to characterize
space are shown, and to Figure S10, Supporting Information the base material which supported the use of a simplified yield
for different stress–strain responses of three different designs, criterion: maximum local strain needs to be below 2%. Clas-
including uncertainty. sifying the design space according to this criterion shows many
Machine learning’s practical relevance is first illustrated by geometric solutions that, despite being coilable (blue region in
targeting a recoverable and highly compressible design manu- Figure 3B), create high local deformations causing plasticity
factured by fused filament fabrication using polylactic acid (PLA) or fracture before the metamaterial is fully compressed. The
(Figure  3). This inexpensive manufacturing process creates inset in Figure 3B represents the corresponding variation of the
parts with defects, so it is challenging to create slender ele- quantities of interest in the reversibly coilable region (yellow
ments. Therefore, simplifying the cross-section of the longer­ons region).
is advantageous. Figure 3A shows the results of a global Sobol Based on the classification and regression results, we
sensitivity analysis[37] obtained by probing the seven input para- manufactured and tested six different designs to validate the
meter machine learning model of Figure 2D. Since the first classification boundaries. Figure 3C,D shows the design with
sensitivity indices of every variable contribute almost entirely highest compressibility, where the final length reduces to the
to the respective total sensitivity, there is minimal high-order thickness of the bottom ring, leading to a total compressive

Adv. Mater. 2019, 1904845 1904845  (3 of 6) © 2019 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
[Link] [Link]

Figure 3.  Design of fully recoverable metamaterial with nearly 100% compressibility printable with hobbyist 3D printer using PLA. Design leads to
local maximum strains below 2%. The experiments were conducted for D1 = 90 mm, D2 = 75 mm, P = 60 mm, d = 1 mm, and 10 longerons. In (C),
the effective compressive strain shown does not consider the thickness of the support rings (measured maximum effective strain was 94% due to
using rings with nonoptimized thickness; note that the maximum value achievable would be 98.3%, where the final configuration height would be
the longerons’ diameter (see Figure S16, Supporting Information for clarity)). In addition, the shaded regions in (C) correspond to the 95% confi-
dence intervals of the experiments and simulations, respectively. The source of uncertainty in the simulations can be seen in Table S3 and Figure S4,
Supporting Information.

strain of 95% and complete recoverability. This is unlike leading to an undesired distortion of the metamaterial even
current additively manufactured solutions reviewed above. before applying a load. We circumvented this challenge by con-
Figure 3C includes a comparison between experiments and ducting the data-driven design on a monolithic concept, i.e., we
finite element predictions taking into account the propagation maintained the unit cell but connected the longerons through
of uncertainty from the input properties to the response of the horizontal elements (Figure 4A). Once again machine learning
metamaterial. The Supporting Information provides additional showed that monolithic designs were possible (Figure S16,
details about input uncertainty, simulations, and experiments Supporting Information), although postcure contraction is not
in other regions of the design space. included in the predictions. Nevertheless, the design charts
Note, however, that the data-driven approach to design relies allowed us to navigate the space in an informed manner, mini-
on the accuracy of the computational predictions (database). mizing our iterations to three until we reached a viable unit cell
The next illustrative example shows that computer models where the longeron diameter was sufficiently large to withstand
often miss important information when confronted with reality, the distorting effects of postcure contraction, while still being
but even in this case, machine learning can be useful. When close to a region of full recoverability.
attempting to downscale the design by nanofabricating the We conducted two different experiments for testing
unit cell, we faced an additional complication not taken into the microscale monolithic metamaterial. In the experi-
account by our finite element predictions: the photoresist resin ment summarized in Figure 4, we considered successively
we used for nanofabrication contracts significantly after curing, longer deformation cycles until we reached maximum

Adv. Mater. 2019, 1904845 1904845  (4 of 6) © 2019 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
[Link] [Link]

Figure 4.  Monolithic metamaterial obtained with two-photon nanolithography with geometry defined by bottom and top support ring diameters,
respectively, D1 = D2 = 200 mm, height P = 120 mm, 21 longerons with diameter d = 7 mm, and horizontal battens with diameter b = 3.5 mm. The white
bars correspond to 50 mm. (D) shows only a few loading cycles for clarity, but the complete experiment is shown in Figure S21, Supporting informa-
tion. We recommend watching Movie S2, Supporting Information of the metamaterial undergoing cycle 14 (after being compressed in stages for 13
times). In addition, Figure S20, Supporting Information refers to a different experiment where the metamaterial is fully compressed at every cycle and
observed with an optical microscope (see also Movie S1, Supporting Information).

compressibility. This experiment demonstrates the recov- ever-increasing role in the quest for the best materials, but it is
erability of every cycle, since each cycle is repeated at least in tuning them for unforeseen scenarios where it may be most
once. Figure S20, Supporting Information shows more data attractive to use it.
about this experiment, including more cycles than showed
in Figure 4D and additional images of the metamaterial (see
also Movie S2, Supporting Information). Furthermore, the Experimental Section
Supporting Information also includes a second experiment Fabrication and Testing of Macroscale Specimens : Based on the classified
where we tested maximum compressibility for every loading– regions shown in Figure 3B, the metamaterial was manufactured with top
unloading cycle (see Figure S19 and Movie S1, Supporting and bottom support rings in aluminum, while 3D printing the longerons
Information). in PLA via fused filament fabrication (Ultimaker 2+). The aluminum
In conclusion, we demonstrate that inverting the design pro- rings minimized friction at the hinges with the longerons because the
cess from experimentally guided to data-driven is possible and finite element analyses did not include the effect of friction. Specimens
were also manufactured completely in PLA and the designs were viable,
advantageous even when computational models are missing but the compressive strength increased due to friction (in this case, the
some information. This approach is widely applicable to design finite element analyses underpredicted strength). Specimens were tested
and analyze different materials and structures. For example, in a standard mechanical testing device.
stretch-dominated and bending-dominated lattice struc- Fabrication and Testing of Microscale Specimens: The microscale
tures[38–40] can be designed and optimized under uncertainty in monolithic specimens were manufactured by two-photon
a similar way. The essential requisites are that “enough” data nanolithography using the Photonic Professional GT system
(Nanoscribe GmbH, Germany) with a high resolution objective (63×)
about the problem of interest is available, and that the data is
and IP-Dip photoresist. Samples were developed in a propylene glycol
sufficiently accurate. This data does not need to be exclusively methyl ether acetate (PGMEA) bath for 30 min followed by a second
computational, as it could also be analytical or experimental. bath in isopropyl alcohol (IPA) for about 2 min. SolidWorks was used for
Therefore, we believe that machine learning will play an the computer-aided design and Describe for compiling the data into the

Adv. Mater. 2019, 1904845 1904845  (5 of 6) © 2019 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
[Link] [Link]

system. A nanomechanical testing device (FemtoTools, Switzerland) was [13] L. R. Meza, S. Das, J. R. Greer, Science 2014, 345, 1322.
used for the compression tests observed by an optical microscope or a [14] C. Zhu, T. Y.-J. Han, E. B. Duoss, A. M. Golobic, J. D. Kuntz,
scanning electron microscope (SEM). C. M. Spadaccini, M. A. Worsley, Nat. Commun. 2015, 6, 6962.
[15] L. Qiu, B. Huang, Z. He, Y. Wang, Z. Tian, J. Z. Liu, K. Wang,
J. Song, T. R. Gengenbach, D. Li, Adv. Mater. 2017, 29, 1701553.
[16] M. Campbell, A. Hoane, F.-h. Hsu, Artif. Intell. 2002, 134, 57.
Supporting Information [17] A. Lally, J. M. Prager, M. C. McCord, B. K. Boguraev, S. Patwardhan,
Supporting Information is available from the Wiley Online Library or J. Fan, P. Fodor, J. Chu-Carroll, IBM J. Res. Dev. 2012, 56, 2:1.
from the author. [18] D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre,
G. van den Driessche, J. Schrittwieser, I. Antonoglou,
V. Panneershelvam, M. Lanctot, S. Dieleman, D. Grewe, J. Nham,
N. Kalchbrenner, I. Sutskever, T. Lillicrap, M. Leach, K. Kavukcuoglu,
Acknowledgements T. Graepel, D. Hassabis, Nature 2016, 529, 484.
M.A.B. would like to acknowledge useful discussions with Ian [19] Y. Taigman, M. Yang, M. Ranzato, L. Wolf, in Proc. of the IEEE
Richardson, Just Herder, Erik Schlangen, and Daniel Fan at Delft Conf. on Computer Vision and Pattern Recognition (CVPR), IEEE,
University of Technology, with Yuchen Wei and Sergio Pellegrino at Piscataway, NJ, USA 2014, pp. 1701–1708.
Caltech, with Beatrice Sartini from Nanoscribe GmbH, and Simon [20] R. Bell, Y. Koren, C. Volinsky, Computer 2009, 42, 30.
Muntwyler at FemtoTools AG. [21] S. Curtarolo, D. Morgan, K. Persson, J. Rodgers, G. Ceder, Phys.
Rev. Lett. 2003, 91, 135503.
[22] R. Gautier, X. Zhang, L. Hu, L. Yu, Y. Lin, T. O. Sunde, D. Chon,
K. R. Poeppelmeier, A. Zunger, Nat. Chem. 2015, 7, 308.
Conflict of Interest [23] D. T. Ahneman, J. G. Estrada, S. Lin, S. D. Dreher, A. G. Doyle,
The authors declare no conflict of interest. Science 2018, 360, 186.
[24] Q. Zhang, C. Liu, X. Wan, L. Zhang, S. Liu, Y. Yang, T. J. Cui, Adv.
Theory Simul. 2019, 2, 1800132.
[25] I. Malkiel, M. Mrejen, A. Nagler, U. Arieli, L. Wolf, H. Suchowski,
Keywords Light: Sci. Appl. 2018, 7, 60.
additive manufacturing, data-driven design, deep learning, machine [26] W. Ma, F. Cheng, Y. Liu, ACS Nano 2018, 12, 6326.
learning, optimization [27] T. Qiu, X. Shi, J. Wang, Y. Li, S. Qu, Q. Cheng, T. Cui, S. Sui, Adv.
Sci. 2019, 6, 1900128.
Received: July 28, 2019 [28] Z. Liu, D. Zhu, S. P. Rodrigues, K.-T. Lee, W. Cai, Nano Lett. 2018,
Revised: September 13, 2019 18, 6570.
Published online: [29] W. Ma, F. Cheng, Y. Xu, Q. Wen, Y. Liu, Adv. Mater. 2019, 31, 1901111.
[30] Z. Ghahramani, Nature 2015, 521, 452.
[31] J. González, Z. Dai, P. Hennig, N. Lawrence, in Proc. of the 19th
Int. Conf. on Artificial Intelligence and Statistics (Eds: A. Gretton,
[1] N. I. Zheludev, Y. S. Kivshar, Nat. Mater. 2012, 11, 917. C. C. Robert), Microtome Publishing, Brookline, MA, USA 2016,
[2] K. Bertoldi, V. Vitelli, J. Christensen, M. van Hecke, Nat. Rev. Mater. pp. 648–657.
2017, 2, 17066. [32] M. Bessa, R. Bostanabad, Z. Liu, A. Hu, D. W. Apley, C. Brinson,
[3] A. L. Rakhmanov, A. M. Zagoskin, S. Savel’ev, F. Nori, Phys. Rev. B W. Chen, W. Liu, Comput. Methods Appl. Mech. Eng. 2017, 320, 633.
2008, 77, 144507. [33] M. Bessa, S. Pellegrino, Int. J. Solids Struct. 2018, 139–140, 174.
[4] A. B. Cairns, J. Catafesta, C. Levelut, J. Rouquette, A. van der Lee, [34] J. Hensman, A. G. Matthews, M. Filippone, Z. Ghahramani, in
L. Peters, A. L. Thompson, V. Dmitriev, J. Haines, A. L. Goodwin, Advances in Neural Information Processing Systems 28 (Eds: C. Cortes,
Nat. Mater. 2013, 12, 212. N. D. Lawrence, D. D. Lee, M. Sugiyama, R. Garnett), Curran
[5] M. Serra-Garcia, J. Lydon, C. Daraio, Phys. Rev. E 2016, 93, 010901. Associates, Inc., Red Hook, NY, USA 2015, pp. 1648–1656.
[6] C. Coulais, E. Teomy, K. de Reus, Y. Shokef, M. van Hecke, Nature [35] M. Titsias, in Proc. of the 12th Int. Conf. on Artificial Intelligence and
2016, 535, 529. Statistics (Eds: D. van Dyk, M. Welling), Microtome Publishing,
[7] J. T. B. Overvelde, J. C. Weaver, C. Hoberman, K. Bertoldi, Nature Brookline, MA, USA 2009, pp. 567–574.
2017, 541, 347. [36] I. M. Sobol, USSR Comput. Math. Math. Phys. 1976, 16, 236.
[8] N. Stoop, R. Lagrange, D. Terwagne, P. M. Reis, J. Dunkel, Nat. [37] A. Saltelli, P. Annoni, I. Azzini, F. Campolongo, M. Ratto,
Mater. 2015, 14, 337. S. Tarantola, Comput. Phys. Commun. 2010, 181, 259.
[9] S. Pellegrino, Deployable Structures, Springer, Vienna 2001, pp. 1–35 [38] V. Deshpande, M. Ashby, N. Fleck, Acta Mater. 2001, 49, 1035.
[10] L. Puig, A. Barton, N. Rando, Acta Astronaut. 2010, 67, 12. [39] T. A. Schaedler, A. J. Jacobsen, A. Torrents, A. E. Sorensen, J. Lian,
[11] N. Gupta, N. M. Sheriff, R. Velmurugan, Thin-Walled Struct. 2006, J. R. Greer, L. Valdevit, W. B. Carter, Science 2011, 334, 962.
44, 986. [40] X. Zheng, H. Lee, T. H. Weisgraber, M. Shusteff, J. DeOtte,
[12] J. Bauer, A. Schroer, R. Schwaiger, O. Kraft, Nat. Mater. 2016, E. B. Duoss, J. D. Kuntz, M. M. Biener, Q. Ge, J. A. Jackson,
15, 438. S. O. Kucheyev, N. X. Fang, C. M. Spadaccini, Science 2014, 344, 1373.

Adv. Mater. 2019, 1904845 1904845  (6 of 6) © 2019 The Authors. Published by WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim