pubs.acs.

org/macroletters Letter

Inverse Design of Complex Block Copolymers for Exotic Self-

Assembled Structures Based on Bayesian Optimization
Qingshu Dong, Xiangrui Gong, Kangrui Yuan, Ying Jiang,* Liangshun Zhang,* and Weihua Li*
Cite This: ACS Macro Lett. 2023, 12, 401−407 Read Online

ACCESS Metrics & More Article Recommendations *

sı Supporting Information
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

ABSTRACT: Variable chain topologies of multiblock copolymers provide

great opportunities for the formation of numerous self-assembled
nanostructures with promising potential applications. However, the
consequent large parameter space poses new challenges for searching the
Downloaded via FUDAN UNIV on October 13, 2023 at 02:32:24 (UTC).

stable parameter region of desired novel structures. In this Letter, by

combining Bayesian optimization (BO), fast Fourier transform-assisted 3D
convolutional neural network (FFT-3DCNN), and self-consistent field
theory (SCFT), we develop a data-driven and fully automated inverse design
framework to search for the desired novel structures self-assembled by ABC-
type multiblock copolymers. Stable phase regions of three exotic target structures are efficiently identified in high-dimensional
parameter space. Our work advances the new research paradigm of inverse design in the field of block copolymers.

B lock copolymer (BCP) self-assembly provides an efficient

platform for engineering rich ordered nanostructures that
show wide applications.1−4 Because phase-separated structures
about the development of approaches for “inverse design” of
BCPs.22−36 Fredrickson and co-workers have coupled bio-
inspired optimization algorithms (e.g., particle swarm opti-
usually enrich and dictate the functions of polymeric materials, mization) with SCFT to explore some known structures
it is vital to gain deep understanding on the “architecture− efficiently in two-dimensional (2D) and three-dimensional
nanostructure” relationship and thus to realize the inverse (3D) spaces.26,37 Nevertheless, the inverse design of BCPs is
design of BCPs for desired structures, especially novel still in its infancy, and it needs urgently to develop more
structures. Self-consistent field theory (SCFT) is one of the efficient methods.38,39
most successful methods for the study of BCPs.5−7 SCFT is
In this Letter, SCFT incorporated with the Bayesian
widely applied to construct the phase diagrams of various
optimization (BO) algorithm40 is implemented to efficiently
BCPs; meanwhile, it can probe into the underlying self-
assembly mechanisms. SCFT has played an irreplaceable role find the desired novel structures and determine their stable
in guiding experimental studies, and its concerted interplay regions in a 3D parameter space of an ABC-type BCP with
with experiments has greatly promoted the study of BCPs.8−14 purposely designed architecture. BO that has emerged as a
The chain architecture or topology of BCPs is one key factor powerful solution for adaptive design problems is particularly
controlling their self-assembly to form novel and even useful in scenarios where the nonconvex objective function is
nonclassical ordered structures.15−19 However, complex costly evaluated and the derivatives are hard to be attained due
architectures need more parameters to describe, expanding to its outstanding ability in utilizing the full information
the parameter space and thus making the exploration of new provided by the optimization history. Its efficient search in an
self-assembled structures more challenging. In recent years, a iterative manner relies on a strategy that balances the trade-off
number of guiding principles have been established for the between the best evaluation from inference (exploitation) and
design of the architectures and components of the maximal uncertainty in the predictions (exploration).
BCPs.9,10,13,14,17,20,21 Though the architectures and compo- Therefore, BO has been widely applied in a wide range of areas
nents can be rationally designed, the stable region of desired to accelerate the optimization of expensive functions for design
structures needs to be explored in the high-dimensional space problems.38,41−45
of parameters including the volume fractions of all blocks, the
interaction parameters, etc. For many complex BCPs, there are
more than five independent parameters,17,19 and thus it is time- Received: January 10, 2023
consuming and tedious to manually search for the stable region Accepted: March 3, 2023
of a target structure. How to develop an efficient approach to Published: March 8, 2023
realize rapid search for the target structure and to roughly
determine its stable region is crucial for the inverse design of
BCPs. Recently, there have been some pioneering reports

© 2023 American Chemical Society https://doi.org/10.1021/acsmacrolett.3c00020

401 ACS Macro Lett. 2023, 12, 401−407
ACS Macro Letters pubs.acs.org/macroletters Letter

Figure 1. Schematic workflow of inverse design algorithm incorporating Bayesian optimization (BO) with SCFT for actively searching the phase
region of a target structure. By maximizing the acquisition function a(X), BO probabilistically recommends a prediction for the next explorable
parameter X (herein, composed of τA, fA, and f C), according to the uncertainty estimation of the objective function g(X) (eq S1 in the Supporting
Information) by the Gaussian process regression (GPR)28,36 on our collected SCFT data. FFT-3DCNN which is built on the convolutional neural
network (CNN) in a 3D version is designed for automatically classifying the SCFT-converged structures, which as the input data to our CNN is
preprocessed by the low-pass filter (LPF) based on fast Fourier transform (FFT). The whole iterative workflow proceeds until the termination
criterion is reached.

We use the standard BO algorithm to realize high-efficient converges to an initially given candidate structure presents a
search for the target structure in the overall space of adjustable great challenge. When the calculated parameters are far from
parameters X.40,46 Mathematically, one aims to find a global the stable region of the given structure, SCFT calculation
minimizer of an unknown objective function g(X) which searches the minimum in an extremely high-dimensional
free-energy landscape may not converge or converge to a
X* = arg min g (X) different structure, outputting an invalid free-energy value.
X
Usually, SCFT calculations were performed under human
= arg min[F *(X) min F *(X)] supervision, so this problem was easy to find. However, human
X (1)
intervention hinders the realization of a fully automated
where Ω represents the parameter space of interest and Fα(X) workflow.
is the free energy of structure α at parameter point X. In the To avoid the manual work for the structure recognition,
present work, Ω is a 3D space constituted by τA, fA, and f C herein, we attempt to develop an effective data-driven scheme
(details will be given later). Herein, g(X) denotes a difference automatically discriminating the spatial structure by combining
between the free energy for the target structure α* and the the convolutional neural network (CNN) in a 3D
minimal free energy for the other candidate structures α. version27,50,51 and the low-pass filter (LPF)52 based on fast
Additional details about the BO algorithm can be found in the Fourier transform (FFT). CNN which facilitates mapping the
Supporting Information. high-dimensional image-like data into a low-dimensional latent
Searching for the target structure by the BO algorithm is space with the concept of weight sharing is applied to
based on the calculation of SCFT for the free-energy data of all distinguish the self-assembled structures by the compositional
candidate structures that can be rationally inferred from density profiles. Considering that the unit cells of different
previous studies.17,19 The pseudospectral method is one of the periodic structures are often divided into different sizes of grid
most efficient approaches for solving SCFT equations.47,48 In lattices for similar calculational accuracy in the pseudospectral
practice, for each candidate structure at specific parameters, method, we use the LPF technique as the data preprocessing
one utilizes a special initialization scheme49 to generate the approach to remove the high-frequency modes of density
initial fields and then obtains the converged density and field profiles in reciprocal space and retain only the same number of
distributions through SCFT iterations. The converged fields low-frequency modes that adequately capture the spatial
can be used as input for the SCFT calculations of other features of self-assembled structures. A total of 2028 density
parameters. How to judge whether the SCFT calculation files, including 27 classes of ordered structures, are collected
402 https://doi.org/10.1021/acsmacrolett.3c00020
ACS Macro Lett. 2023, 12, 401−407
ACS Macro Letters pubs.acs.org/macroletters Letter

Figure 2. Search results with SDGN as the target structure. (a) w = 0.2. (b) w = 1. (c) w = 2. (d) w = 5. (a1−d1) Variation of the objective function
g(X) along with the iterative process. The red crosses on the horizontal line indicate that the target structure is unstable and g(X) is not obtained.
(a2−d2) The 3D scattering plot of sample points (the filled circles indicate g(X) > 0, and the filled squares indicate g(X) < 0) and the stable region
of SDGN constructed from these explored points. The stable region is represented by the isosurface of g(X) = 0, within which the target structure
has the lowest free energy among all candidate structures. The horizontal section in (c2) is used to draw Figure 3c.

from the SCFT results of our previous studies as the data set to from the real stable region of α*, we conduct SCFT
train the FFT-3DCNN. For each class of structure, the data are calculations on all candidate structures as the initial guess.
randomly divided into 80% training and 20% testing sets. The After the automatic recognition of SCFT converged
training results show that the predictive accuracy of our FFT- morphologies by our well-trained FFT-3DCNN model, we
3DCNN can reach 99.6% on training set, which ensures the readily sort all morphologies by their free energies to obtain
reliable structural discrimination. With the support of this the objective function g(X) in eq 1. According to g(X), the BO
high-precision FFT-3DCNN, the BO algorithm can be algorithm provides a recommendation for the next X. The
integrated into a fully automated workflow. Details on our iterative procedure continues until the termination criterion is
FFT-3DCNN model are provided in the Supporting reached. It should be noted that the initial parameters and the
Information. corresponding values of g(X) can be empty, which allows the
Once SCFT is equipped with FFT-3DCNN and BO, an researcher to provide as little prior knowledge as possible. It is
iterative search for the thermodynamically stable phase region also worth stressing that our search scheme heavily relies on
of the target structure α* becomes fully automated. The whole SCFT convergence. For the situation that SCFT likely fails to
workflow of our scheme is illustrated in Figure 1. Starting from eventually converge to a specific morphology α (in the case the
one random parameter point X, which is allowed to be far away maximum number of SCFT iteration is beyond a threshold),
403 https://doi.org/10.1021/acsmacrolett.3c00020
ACS Macro Lett. 2023, 12, 401−407
ACS Macro Letters pubs.acs.org/macroletters Letter

Figure 3. (a) Search results with SD as the target structure for w = 2. (b) Search results with NDGN as the target structure for w = 1. (c) 2D phase
regions constructed from BO results (color filled areas) together with those calculated manually (delimited by solid lines) for f C = 0.25.

commonly encountered for the selected parameter X far from diagram of A(BC)3 and A(ABC)3 can be reduced to A(BC)3 at
the stable phase region of the morphology initially imposed, we τA = 1, we conjecture that A(ABC)3 would also have a stable
will neglect this structure when computing min F *(X). region of SDGN in its parameter space. Accordingly, we first
Moreover, when the target structure α* is not converged at choose SDGN as the target structure. Sample points suggested
specific X′, we set g(X′) to the maximum g(X) in the historical by BO with different w (from top to bottom) are shown in
searching steps. In practice, we find that this trick efficiently Figures 2 and S9. The search path connecting the sample
prevents the BO algorithm from getting stuck in the vicinity of points in sequence is presented in the left column of Figure 2,
X′. More details about the implementation of BO are provided whereas the 3D surface of the phase region estimated from
in the Supporting Information. these sample points by scattered interpolation is plotted in the
Recently, we have discovered a series of exotic structures in right column. Our results suggest that BO finds the first stable
miktoarm ABC-type BCPs, such as sphere-network hybrid parameter point of SDGN (g(X) < 0) quickly for all values of w,
structures,19 which may have great potential applications. For implying that SDGN should have a considerable stable region in
example, the sphere-diamond structure (SD) can be used to the search space. Apparently, BO with w = 0.2 is trapped in the
generate a single-diamond network for complete photonic region around a minimum due to the low probability of
bandgaps.53,54 Inspired by our previous work that branching A- exploration. As shown by the search path (Figure 2a1), BO
block can greatly amplify the curvature of A-domain,13,18 we rarely suggests points with large g(X) in the last 100 steps.
propose a new miktoarm copolymer, denoted as A1(A2BC)n, Nevertheless, we managed to construct a stable region from
which might stabilize the A-sphere-containing hybrid struc- these sample points. However, the phase region determined in
tures. We choose A1(A2BC)3 with fixed χABN = χACN = χBCN = the BO process with w = 0.2 should be underestimated because
80 as a model BCP to demonstrate the efficiency of our it does not extend to τA = 1.0, at which a stable region of the
inverse-design workflow. To facilitate the visualization, we SDGN phase has been predicted in our previous work.19
choose three search variables, i.e., the total volume fraction of In contrast, the suggested points by BO with w = 5 and w =
A-blocks (fA), the volume-fraction ratio of the linear A-block to 10 are scattered in the parameter space more uniformly. The
search paths show that there is little difference in the
the total A-blocks ( A = fA /fA ), and the volume fraction of all
1 fluctuation of g(X) between the first 100 steps and the last
C-blocks ( f C). According to the known self-assembly 100 steps. Although large w lowers the exploitation of the
mechanisms, the variational ranges of the three control historical sample points, it is not difficult to find the stable
parameters can be reasonably set as τA ∈ [0.5, 1.0], fA ∈ points of SDGN because of its large stable region. The main
[0.1, 0.3], and f C ∈ [0.2, 0.4], where the BCP is most likely to problem of large w lies in that the number of effective samples
form novel hybrid structures. For such a search space, we points around the stable region is reduced, lowering the
judiciously select 27 candidate structures shown in Figure S2.19 accuracy of the interpolated phase boundaries. By comparison,
The acquisition weight (w) is an important hyperparameter a rather good balance between exploration and exploitation is
in our BO algorithm, where the acquisition is more explorative reached when w = 1 and 2. As a result, a reasonably precise
at large w and more exploitative at small w. With smaller w, BO phase region is constructed based on 200 historically suggested
is more likely to be trapped in one local minimum. On the parameter points by BO with w = 1 and 2. It is worthwhile to
contrary, BO at large w prefers to exploring the region with a mention that the first stable point of SDGN (i.e., g(X) < 0) is
high uncertainty, leading to a global search in the parameter found in fewer than 20 steps for all values of w we test, which is
space. A delicate balance between exploitation and exploration at least 5−10 times faster than a brute-force search method
is beneficial for BO to efficiently find the global optimum. To (i.e., grid by grid).
roughly estimate the value of w, we make 200-step search for It should be noted that the BO iteration is started with
the desired structure for several times with different w. It random parameter points, but not specific initial points. As a
should be noted that we are not particularly concerned with consequence, the BO algorithm in the first few steps searches
the exact optimal value of w, which can be determined through parameter points randomly. This is why large g(X) is obtained
trial and error or by hyperparameter tuning techniques.40 or even the SCFT calculation initialized with the target
Therefore, we empirically selected several w values to illustrate structure cannot converge. Nevertheless, the first parameter
its influence on BO. point for g(X) < 0 is reached in a very few steps. This indicates
Because the structure composed of A-spheres within the that the BO algorithm is very robust, only requiring very
double-gyroid C-networks (SDGN) is stable in the phase limited prior knowledge of the studied system.
404 https://doi.org/10.1021/acsmacrolett.3c00020
ACS Macro Lett. 2023, 12, 401−407
ACS Macro Letters pubs.acs.org/macroletters Letter

The SD phase is not stable in the phase diagram of A(BC)3 complex ABC-type BCP. The assumption about all possible
but may be stable in A(ABC)3 with branched A-blocks.19 structures is a limitation of our framework. This limitation can
Therefore, searching for SD is also valuable but more be overcame by coupling our framework with a recently
challenging. A stable region of SD is found for all tested developed method of discovering globally stable morpholo-
values of w (Figures 3a and S10), indicating that the SD phase gies.34 Nevertheless, our work advances the inverse design of
can be stabilized by the branching architecture of A-blocks. block copolymers significantly. Furthermore, our inverse
Encouraged by the search results of SD, we turn to another design framework can be readily extended to build up the
previously unreported structure, i.e., the network−network “architecture-ordered structure-properties” relationships, which
phase (NDGN) consisting of double-gyroid A-struts while C- are helpful for the design of nanostructured polymeric
blocks form a perforated layer (or network) sitting on the materials.
minimal surface between the two A-struts. Surprisingly, a
noticeable stable region of NDGN is successfully identified by
BO in Figures 3b and S11. Although the stable region is very
■
*
ASSOCIATED CONTENT
sı Supporting Information
small, the first parameter point for g(X) < 0 can be found
within about 50 steps. To search for the structure with such a The Supporting Information is available free of charge at
small stable region, the brute-force search method need at least https://pubs.acs.org/doi/10.1021/acsmacrolett.3c00020.
10 sample points along each axis, that is, 1000 grid points in Additional BO details; FFT-3DCNN details; Tables
the considered 3D space. In this case our BO algorithm is at S1−S2 and Figures S1−S12; efficacy of the BO
least 20 times faster. algorithm (PDF)
To verify the validity of BO searching results, we manually
calculate a 2D cross section of the phase diagram at fixed f C = FFT-3DCNN code (ZIP)
0.25 under the guide of the BO results in Figure 3c, which
contains the stable regions of all three target structures. The
2D stable regions for SDGN (w = 2), SD (w = 2), and NDGN (w
= 1) at f C = 0.25 are superimposed on the phase diagram. The
■ AUTHOR INFORMATION
Corresponding Authors
BO results are in well consistent with our manually calculated Ying Jiang − School of Chemistry, Center of Soft Matter
phase regions, especially for the SDGN and SD phases with Physics and its Applications, Beihang University, Beijing
relatively large stable regions. It should be noted that the 100191, China; orcid.org/0000-0002-6041-2272;
inverse design framework mainly aims to search for the target Email: yjiang@buaa.edu.cn
morphology, not for accurate phase diagrams. Therefore, the Liangshun Zhang − Shanghai Key Laboratory of Advanced
sample points suggested by BO to minimize g(X) are not Polymeric Materials, School of Materials Science and
perfect for constructing accurate phase boundaries, but those Engineering, East China University of Science and
that are evenly distributed near the boundaries. Nevertheless, Technology, Shanghai 200237, China; orcid.org/0000-
our results are very useful to guide both the design of 0002-0182-7486; Email: zhangls@ecust.edu.cn
experimental systems and the calculations of precise phase Weihua Li − State Key Laboratory of Molecular Engineering of
diagrams. On one hand, the stable region of the target Polymers, Key Laboratory of Computational Physical
structure can be found quickly (if it exists), which is enough to Sciences, Department of Macromolecular Science, Fudan
guide the experimental research because there are inevitably University, Shanghai 200433, China; orcid.org/0000-
differences in the parameters between theory and experiment. 0002-5133-0267; Email: weihuali@fudan.edu.cn
On the other hand, the search results can be used as a prior
knowledge to guide the further calculation of precise phase Authors
diagrams. Qingshu Dong − State Key Laboratory of Molecular
In order to quantitatively evaluate the efficacy of BO, we Engineering of Polymers, Key Laboratory of Computational
manually calculated the phase region of SDGN in the 3D Physical Sciences, Department of Macromolecular Science,
parameter space. The overlap ratio between the BO predicted Fudan University, Shanghai 200433, China
phase region and the manually calculated phase region is Xiangrui Gong − School of Chemistry, Center of Soft Matter
plotted to show the model quality as a function of steps Physics and its Applications, Beihang University, Beijing
(Figure S12).55 The results clearly show that BO with w = 1 100191, China
and 2 is superior to w = 0.2, 5, and 10. The accuracy of BO Kangrui Yuan − State Key Laboratory of Molecular
with w = 2 converges up to 0.6 after 150 steps. Details about Engineering of Polymers, Key Laboratory of Computational
the efficacy of BO are discussed in the Supporting Information. Physical Sciences, Department of Macromolecular Science,
In conclusion, we propose an inverse design framework that Fudan University, Shanghai 200433, China
combines BO, FFT-3DCNN, and SCFT to efficiently search Complete contact information is available at:
the stable phase regions of novel target structures in high https://pubs.acs.org/10.1021/acsmacrolett.3c00020
dimensions. We first develop a single 3DCNN model based on
FFT low-pass filtering scheme to recognize the converged
Author Contributions
structures of SCFT, thus making SCFT calculations output
valid values of free energy for the BO algorithm. As a result, the Q.D. and X.G. contributed equally to this work. CRediT:
BO iteration becomes fully automated. Taking three exotic Xiangrui Gong investigation (supporting), methodology
ordered phases as target, we demonstrate that the BO (equal); Kangrui Yuan investigation (supporting), method-
algorithm with a tailored balance between exploitation and ology (supporting).
exploration can find their global optima and determine their Notes
stable regions efficiently in the 3D parameter space of a The authors declare no competing financial interest.
405 https://doi.org/10.1021/acsmacrolett.3c00020
ACS Macro Lett. 2023, 12, 401−407
ACS Macro Letters pubs.acs.org/macroletters Letter

■ ACKNOWLEDGMENTS
This work was supported by the National Natural Science
(21) Liu, M. J.; Qiang, Y. C.; Li, W. H.; Qiu, F.; Shi, A. C. Stabilizing
the Frank-Kasper Phases via Binary Blends of AB Diblock
Copolymers. ACS Macro Lett. 2016, 5, 1167−1171.
Foundation of China (Grants 21925301, 22203018, 22073004, (22) Tsai, C. L.; Delaney, K. T.; Fredrickson, G. H. Genetic
and 22073028) and the Fundamental Research Funds for the Algorithm for Discovery of Globally Stable Phases in Block
Central Universities. Copolymers. Macromolecules 2016, 49, 6558−6567.
(23) Terayama, K.; Tamura, R.; Nose, Y.; Hiramatsu, H.; Hosono,

■ REFERENCES
(1) Polymeropoulos, G.; Zapsas, G.; Ntetsikas, K.; Bilalis, P.;
H.; Okuno, Y.; Tsuda, K. Efficient Construction Method for Phase
Diagrams Using Uncertainty Sampling. Phys. Rev. Mater. 2019, 3,
033802.
Gnanou, Y.; Hadjichristidis, N. 50th Anniversary Perspective: (24) Nakamura, I. Phase Diagrams of Polymer-Containing Liquid
Polymers with Complex Architectures. Macromolecules 2017, 50, Mixtures with a Theory-Embedded Neural Network. New J. Phys.
1253−1290. 2020, 22, 015001.
(2) Bates, C. M.; Bates, F. S. 50th Anniversary Perspective: Block (25) Dai, C.; Glotzer, S. C. Efficient Phase Diagram Sampling by
Polymers-Pure Potential. Macromolecules 2017, 50, 3−22. Active Learning. J. Phys. Chem. B 2020, 124, 1275−1284.
(3) Huang, C. H.; Zhu, Y. Y.; Man, X. K. Block Copolymer Thin (26) Case, L. J.; Delaney, K. T.; Fredrickson, G. H.; Bates, F. S.;
Films. Phys. Rep.-Rev. Sec. Phys. Lett. 2021, 932, 1−36. Dorfman, K. D. Open-Source Platform for Block Polymer
(4) Xu, Z. W.; Dong, Q. S.; Zhang, L. S.; Li, W. H. Enhanced Formulation Design Using Particle Swarm Optimization. Eur. Phys.
Dielectric Permittivity of Hierarchically Double-Gyroid Nanocompo- J. E 2021, 44, 115.
sites via Macromolecular Engineering of Block Copolymers. Nanoscale (27) Aoyagi, T. Deep Learning Model for Predicting Phase
2022, 14, 15275−15280. Diagrams of Block Copolymers. Comput. Mater. Sci. 2021, 188,
(5) Drolet, F.; Fredrickson, G. H. Combinatorial Screening of 110224.
Complex Block Copolymer Assembly with Self-Consistent Field (28) Ethier, J. G.; Casukhela, R. K.; Latimer, J. J.; Jacobsen, M. D.;
Theory. Phys. Rev. Lett. 1999, 83, 4317−4320. Shantz, A. B.; Vaia, R. A. Deep Learning of Binary Solution Phase
(6) Matsen, M. W. Effect of Architecture on the Phase Behavior of Behavior of Polystyrene. ACS Macro Lett. 2021, 10, 749−754.
AB-Type Block Copolymer Melts. Macromolecules 2012, 45, 2161− (29) Zhao, C.; Cai, T. Y.; Zhang, L. S.; Li, W. H.; Lin, J. P.
2165. Autonomous Construction of Phase Diagrams of Block Copolymers
(7) Matsen, M. W. Field Theoretic Approach for Block Polymer by Theory-Assisted Active Machine Learning. ACS Macro Lett. 2021,
Melts: SCFT and FTS. J. Chem. Phys. 2020, 152, 110901. 10, 598−602.
(8) Lee, S.; Bluemle, M. J.; Bates, F. S. Discovery of a Frank-Kasper (30) Xuan, Y.; Delaney, K. T.; Ceniceros, H. D.; Fredrickson, G. H.
Sigma Phase in Sphere-Forming Block Copolymer Melts. Science Deep Learning and Self-Consistent Field Theory: A Path Towards
2010, 330, 349−353. Accelerating Polymer Phase Discovery. J. Comput. Phys. 2021, 443,
(9) Xie, N.; Li, W.; Qiu, F.; Shi, A.-C. σ Phase Formed in 110519.
Conformationally Asymmetric AB-Type Block Copolymers. ACS (31) Arora, A.; Lin, T.-S.; Rebello, N. J.; Av-Ron, S. H. M.;
Macro Lett. 2014, 3, 906−910. Mochigase, H.; Olsen, B. D. Random Forest Predictor for Diblock
(10) Schulze, M. W.; Lewis, R. M.; Lettow, J. H.; Hickey, R. J.; Copolymer Phase Behavior. ACS Macro Lett. 2021, 10, 1339−1345.
Gillard, T. M.; Hillmyer, M. A.; Bates, F. S. Conformational (32) Shen, Z.; Sun, Y.; Lodge, T. P.; Siepmann, J. I. Development of
Asymmetry and Quasicrystal Approximants in Linear Diblock a Pointnet for Detecting Morphologies of Self-Assembled Block
Copolymers. Phys. Rev. Lett. 2017, 118, 207801. Oligomers in Atomistic Simulations. J. Phys. Chem. B 2021, 125,
(11) Duan, C.; Zhao, M.; Qiang, Y.; Chen, L.; Li, W.; Qiu, F.; Shi, 5275−5284.
A.-C. Stability of Two-Dimensional Dodecagonal Quasicrystalline (33) Sattari, K.; Xie, Y. C.; Lin, J. Data-Driven Algorithms for
Phase of Block Copolymers. Macromolecules 2018, 51, 7713−7721. Inverse Design of Polymers. Soft Matter 2021, 17, 7607−7622.
(12) Miyamori, Y.; Suzuki, J.; Takano, A.; Matsushita, Y. Periodic (34) Tsai, C. R. L.; Fredrickson, G. H. Using Particle Swarm
and Aperiodic Tiling Patterns from a Tetrablock Terpolymer System Optimization and Self-Consistent Field Theory to Discover Globally
of the A1BA2C Type. ACS Macro Lett. 2020, 9, 32−37. Stable Morphologies of Block Copolymers. Macromolecules 2022, 55,
(13) Li, C.; Dong, Q.; Li, W. Largely Tunable Asymmetry of Phase 5249−5262.
Diagrams of A(AB)n Miktoarm Star Copolymer. Macromolecules (35) Wei, D.; Zhou, T.; Huang, Y.; Jiang, K. A Multi-Category
2020, 53, 10907−10917. Inverse Design Neural Network and Its Application to Diblock
(14) Seo, Y.; Woo, D.; Li, L. Y.; Li, W. H.; Kim, J. K. Phase Behavior Copolymers. Mathematics 2022, 10, 4451.
of PS-(PS-b-P2VP)3 Miktoarm Star Copolymer. Macromolecules 2021, (36) Audus, D. J.; McDannald, A.; DeCost, B. Leveraging Theory for
54, 7822−7829. Enhanced Machine Learning. ACS Macro Lett. 2022, 11, 1117−1122.
(15) Grason, G. M.; Kamien, R. D. Interfaces in Diblocks: A Study (37) Khadilkar, M. R.; Paradiso, S.; Delaney, K. T.; Fredrickson, G.
of Miktoarm Star Copolymers. Macromolecules 2004, 37, 7371−7380. H. Inverse Design of Bulk Morphologies in Multiblock Polymers
(16) Qin, J.; Bates, F. S.; Morse, D. C. Phase Behavior of Using Particle Swarm Optimization. Macromolecules 2017, 50, 6702−
Nonfrustrated ABC Triblock Copolymers: Weak and Intermediate 6709.
Segregation. Macromolecules 2010, 43, 5128−5136. (38) Schneider, L.; Schwarting, M.; Mysona, J.; Liang, H. Y.; Han,
(17) Xie, N.; Liu, M. J.; Deng, H. L.; Li, W. H.; Qiu, F.; Shi, A.-C. M.; Rauscher, P. M.; Ting, J. F. M.; Venkatram, S.; Ross, R. B.;
Macromolecular Metallurgy of Binary Mesocrystals via Designed Schmidt, K. J.; Blaiszik, B.; Foster, I.; de Pablo, J. J. In Silico Active
Multiblock Terpolymers. J. Am. Chem. Soc. 2014, 136, 2974−2977. Learning for Small Molecule Properties. Mol. Syst. Des. Eng. 2022, 7,
(18) Qiang, Y.; Li, W.; Shi, A.-C. Stabilizing Phases of Block 1611−1621.
Copolymers with Gigantic Spheres via Designed Chain Architectures. (39) He, J.; Su, X.; Wang, C.; Li, J.; Hou, Y.; Li, Z.; Liu, C.; Xue, D.;
ACS Macro Lett. 2020, 9, 668−673. Cao, J.; Su, Y.; Qiao, L.; Lookman, T.; Bai, Y. Machine Learning
(19) Dong, Q.; Li, W. Effect of Molecular Asymmetry on the Assisted Predictions of Multi-Component Phase Diagrams and Fine
Formation of Asymmetric Nanostructures in ABC-Type Block Boundary Information. Acta Mater. 2022, 240, 118341.
Copolymers. Macromolecules 2021, 54, 203−213. (40) Garnett, R. Bayesian Optimization; Cambridge University Press:
(20) Gao, Y.; Deng, H. L.; Li, W. H.; Qiu, F.; Shi, A. C. Formation Cambridge, 2023, DOI: 10.1017/9781108348973.
of Nonclassical Ordered Phases of AB-Type Multiarm Block (41) Herbol, H. C.; Hu, W. C.; Frazier, P.; Clancy, P.; Poloczek, M.
Copolymers. Phys. Rev. Lett. 2016, 116, 068304. Efficient Search of Compositional Space for Hybrid Organic-

406 https://doi.org/10.1021/acsmacrolett.3c00020
ACS Macro Lett. 2023, 12, 401−407
ACS Macro Letters pubs.acs.org/macroletters Letter

Inorganic Perovskites via Bayesian Optimization. npj Comput. Mater.

2018, 4, 51.
(42) Minami, T.; Kawata, M.; Fujita, T.; Murofushi, K.; Uchida, H.;
Omori, K.; Okuno, Y. Prediction of Repeat Unit of Optimal Polymer
by Bayesian Optimization. MRS Adv. 2019, 4, 1125−1130.
(43) Wang, Y.; Xie, T.; France-Lanord, A.; Berkley, A.; Johnson, J.
A.; Shao-Horn, Y.; Grossman, J. Toward Designing Highly
Conductive Polymer Electrolytes by Machine Learning Assisted
Coarse-Grained Molecular Dynamics. Chem. Mater. 2020, 32, 4144−
4151.
(44) Graff, D. E.; Shakhnovich, E. I.; Coley, C. W. Accelerating
High-Throughput Virtual Screening through Molecular Pool-Based
Active Learning. Chem. Sci. 2021, 12, 7866−7881.
(45) Sasaki, T.; Nakamura, A.; Flarasawa, J.; Hara, K.; Suzuki, I.;
Takahashi, T. Bayesian Optimization with an Auxiliary Classifier for
the Development of Polymer Materials. IEEE International Conference
on Big Data; Electr Network: 2021; pp 6014−6016.
(46) Jones, D. R.; Schonlau, M.; Welch, W. J. Efficient Global
Optimization of Expensive Black-Box Functions. J. Glob. Optim. 1998,
13, 455−492.
(47) Tzeremes, G.; Rasmussen, K. Ø.; Lookman, T.; Saxena, A.
Efficient Computation of the Structural Phase Behavior of Block
Copolymers. Phys. Rev. E 2002, 65, 041806.
(48) Rasmussen, K. Ø.; Kalosakas, G. Improved Numerical
Algorithm for Exploring Block Copolymer Mesophases. J. Polym.
Sci. Pt. B-Polym. Phys. 2002, 40, 1777−1783.
(49) Li, W. H.; Qiu, F.; Shi, A. C. Emergence and Stability of Helical
Superstructures in ABC Triblock Copolymers. Macromolecules 2012,
45, 503−509.
(50) Aoyagi, T. High-Throughput Prediction of Stress-Strain Curves
of Thermoplastic Elastomer Model Block Copolymers by Combining
Hierarchical Simulation and Deep Learning. MRS Adv. 2021, 6, 32−
36.
(51) Shen, Z. Y.; Luo, K.; Park, S. J.; Li, D. Y.; Mahanthappa, M. K.;
Bates, F. S.; Dorfman, K. D.; Lodge, T. P.; Siepmann, J. I. Stabilizing a
Double Gyroid Network Phase with 2 nm Feature Size by Blending of
Lamellar and Cylindrical Forming Block Oligomers. JACS Au 2022, 2, Recommended by ACS
1405−1416.
(52) Bow, S.-T. Pattern Recognition and Image Preprocessing; CRC
Press: Boca Raton, FL, 2002. Thickness-Controlled Phase Transitions of AB Diblock
(53) Maldovan, M.; Thomas, E. L. Diamond-Structured Photonic Copolymers in Asymmetric Ultrathin Films
Crystals. Nat. Mater. 2004, 3, 593−600. Luyang Li, Weihua Li, et al.
(54) Lequieu, J.; Quah, T.; Delaney, K. T.; Fredrickson, G. H. JULY 19, 2023
Complete Photonic Band Gaps with Nonfrustrated ABC Bottlebrush MACROMOLECULES READ
Block Polymers. ACS Macro Lett. 2020, 9, 1074−1080.
(55) This helpful evaluation method is thoughtfully suggested by the Complex Phase Behavior in Binary Blends of AB Diblock
reviewers. Copolymer and ABC Triblock Terpolymer
So Jung Park, Kevin D. Dorfman, et al.
JANUARY 27, 2023
MACROMOLECULES READ

Alternating Gyroid Stabilized by Surfactant-like Triblock

Terpolymers in IS/SO/ISO Ternary Blends
Pengyu Chen, Kevin D. Dorfman, et al.
MARCH 07, 2023
MACROMOLECULES READ

Emergence and Stability of Exotic “Binary” HCP-Type

Spherical Phase in Binary AB/AB Blends
Fengmei Zhao, Weihua Li, et al.
NOVEMBER 04, 2022
MACROMOLECULES READ

Get More Suggestions >

407 https://doi.org/10.1021/acsmacrolett.3c00020
ACS Macro Lett. 2023, 12, 401−407