Machine-Learning-Assisted Selective Synthesis of a Semiconductive Silver Thiolate Coordination Polymer with Segregated Paths for Holes and Electrons

Article abstract of DOI:10.1002/anie.202110629

Coordination polymers (CPs) with infinite metal–sulfur bond networks have unique electrical conductivities and optical properties. However, the development of new (-M-S-)_n-structured CPs is hindered by difficulties with their crystallization. Herein, we describe the use of machine learning to optimize the synthesis of trithiocyanuric acid (H₃ttc)-based semiconductive CPs with infinite Ag?S bond networks, report three CP crystal structures, and reveal that isomer selectivity is mainly determined by proton concentration in the reaction medium. One of the CPs, [Ag₂Httc]_n, features a 3D-extended infinite Ag?S bond network with 1D columns of stacked triazine rings, which, according to first-principle calculations, provide separate paths for holes and electrons. Time-resolved microwave conductivity experiments show that [Ag₂Httc]_n is highly photoconductive (φΣμ_max=1.6×10^?4 cm² V^?1 s^?1). Thus, our method promotes the discovery of novel CPs with selective topologies that are difficult to crystallize.

Full text of DOI:10.1002/anie.202110629

A Journal of the Gesellschaft Deutscher Chemiker
Angewandte
Chemie
International Edition
www.angewandte.org
Accepted Article
Title: Machine-Learning-Assisted Selective Synthesis of
Semiconductive Silver Thiolate Coordination Polymer with
Segregated Paths for Holes and Electrons
Authors: Takuma Wakiya, Yoshinobu Kamakura, Hiroki Shibahara,
Kazuyoshi Ogasawara, Akinori Saeki, Ryosuke Nishikubo,
Akihiro Inokuchi, Hirofumi Yoshikawa, and Daisuke Tanaka
This manuscript has been accepted after peer review and appears as an
Accepted Article online prior to editing, proofing, and formal publication
of the final Version of Record (VoR). This work is currently citable by
using the Digital Object Identifier (DOI) given below. The VoR will be
published online in Early View as soon as possible and may be different
to this Accepted Article as a result of editing. Readers should obtain
the VoR from the journal website shown below when it is published
to ensure accuracy of information. The authors are responsible for the
content of this Accepted Article.
To be cited as: Angew. Chem. Int. Ed. 10.1002/anie.202110629
Link to VoR: https://doi.org/10.1002/anie.202110629

10.1002/anie.202110629
Angewandte Chemie International Edition
RESEARCH ARTICLE
Machine-Learning-Assisted Selective Synthesis of
Semiconductive Silver Thiolate Coordination Polymer with
Segregated Paths for Holes and Electrons
Takuma Wakiya,a Yoshinobu Kamakura,a Hiroki Shibahara,a Kazuyoshi Ogasawara,a Akinori Saeki,b
Ryosuke Nishikubo,b Akihiro Inokuchi,c Hirofumi Yoshikawad and Daisuke Tanaka*,a,e
[a]
T. Wakiya, Dr. Y. Kamakura, H. Shibahara, Prof. K. Ogasawara, Prof. D. Tanaka
Department of Chemistry
School of Science and Technology
Kwansei Gakuin University
2-1 Gakuen, Sanda, Hyogo 669-1337, Japan
*E-mail: dtanaka@kwansei.ac.jp
Prof. A. Saeki, Dr. R. Nishikubo
Department of Applied Chemistry
Graduate School of Engineering
Osaka University
2-1 Yamadaoka, Suita, Osaka 565-0871, Japan
Prof. A. Inokuchi
Department of Informatics
School of Science and Technology
Kwansei Gakuin University
2-1 Gakuen, Sanda, Hyogo 669-1337, Japan
Prof. H. Yoshikawa
Department of Nanotechnology for Sustainable Energy
School of Science and Technology
Kwansei Gakuin University
2-1 Gakuen, Sanda, Hyogo 669-1337, Japan
Prof. D. Tanaka
[b]
[c]
[d]
[e]
JST PRESTO
2-1 Gakuen, Sanda, Hyogo 669-1337, Japan
Supporting information for this article is given via a link at the end of the document.
Abstract: Coordination polymers (CPs) with infinite metal–S bond
networks have unique electrical conductivities and optical properties,
thus drawing attention. However, the development of new (−M−S−)n-
structured CPs is hindered by their crystallization difficulty. Herein,
we use machine learning to optimize the synthesis of trithiocyanuric
acid (H3ttc)-based semiconductive CPs with infinite Ag–S bond
networks, report three CP crystal structures, and reveal that isomer
selectivity is mainly determined by proton concentration in the
reaction medium. One of the CPs, [Ag2Httc]n, features a 3D
extended infinite Ag–S bond network with 1D columns of stacked
triazine rings, which, according to first-principle calculations, provide
separate paths for holes and electrons. Time-resolved microwave
conductivity experiments show that [Ag2Httc]n is highly
photoconductive (φΣμmax = 1.6 × 10−4 cm2 V−1 s−1). Thus, our method
promotes the discovery of novel CPs difficult to crystallize with
selective topologies.
difficult to obtain as single crystals.[2f] Furthermore, the large
ionic radius of S allows this element to adopt different bridging
(μ) modes, which results in the broad structural diversity of the
corresponding CPs and MOFs.[7] Moreover, the flexibility of the S
coordination geometry makes the control of the topological
isomers of M–S CPs and MOFs important for their crystal
structure design.[8] Therefore, facile and versatile methods for
optimizing the syntheses of highly crystalline CPs and MOFs
with metal–thiolate bonds are highly sought after.[2f]
Recently, machine learning applications have rapidly expanded
in the area of materials science[9] and have been used to
optimize syntheses.[10] However, reports on the application of
machine learning to the preparation of novel CPs and MOFs are
few because of the complicated compositions of the solids
obtained in failure experiments. The products obtained during
such optimization procedures are often complex mixtures, and a
facile approach for the assessment of such crude products is yet
to be established.
We have recently reported a method allowing one to explore the
synthetic conditions of novel MOFs based on two machine
learning techniques, namely (i) the cluster analysis of powder X-
ray diffraction (PXRD) patterns and (ii) random forest and
decision tree analysis.[11] Cluster analysis is an unsupervised
learning technique automatically grouping the PXRD patterns of
crystals obtained under various synthetic conditions based on
the degree of similarity (correlation) between these patterns.
Random forest and decision tree analysis are supervised
learning techniques that create a prediction model using the
experimental conditions as explanatory variables and the PXRD
patterns classified by clustering analysis as objective variables.
The above approach allowed us to extract chemical insights
from failure experiments and optimize the conditions for the
synthesis of novel MOFs. Herein, this method is applied to the
Introduction
Coordination polymers (CPs) and metal–organic frameworks
(MOFs) exhibit the features of both inorganic and organic
compounds are therefore of particular interest.[1] Various CPs
and MOFs with coordinated S have been recently reported,[2]
including those featuring infinite metal–S (–M–S–)n bonds
arranged in one-dimensional chains,[3] two-dimensional layers,[4]
and three-dimensionally extended structures.[5] Such materials
exhibit high electrical conductivities and structural stabilities,
thus finding numerous applications in electronic devices.[6]
However, CPs and MOFs with metal–thiolate bonds often exhibit
poor crystallinities, mainly because of the strong M–S bonds and
the rapid precipitation of low-crystalline solids, and are therefore
1
This article is protected by copyright. All rights reserved.

10.1002/anie.202110629
Angewandte Chemie International Edition
RESEARCH ARTICLE
syntheses of novel Ag–S CPs (Figure 1), which have attracted
cases), Ag2S (27 cases), and a minor uncertain impurity phase
(1 case) (Figures 2(a) and S1). At this point, we want to re-state
that the grouping of PXRD patterns into one of the five
categories by humans is inevitably arbitrary, whereas clustering
analysis forces each pattern into one of the categories using
quantitative indicators, which is one of the advantages of this
analysis. As the following supervised learning handles data
statistically, one can expect blurring in the boundary region not
to be a fatal problem.
much attention because of their unique optical properties and
12]
semiconducting nature.[4b,
Specifically, we describe the
syntheses and crystal structures of three novel topological
isomers of CPs comprising Ag+ and trithiocyanuric acid (H3ttc).
The dominant reaction parameters for selective isomer
syntheses are determined through the machine-learning-based
analysis of failure experiments, and the optimal isomer is shown
to exhibit favorable photoconductivity.
Random forest analysis was subsequently performed to
evaluate the main reaction parameters affecting product
selectivity (Section S4). The 10 synthetic parameters, namely
reaction temperature, solvent type, reaction container type,
ligand concentration, metal/ligand ratio, metal ion concentration,
isothermal reaction time (h), modulator type, modulator amount
(mol), and modulator concentration (mM), were used as
explanatory variables. Temperature, concentration, reaction time,
and mixing ratio were expressed as numerical data, whereas
solvent type, modulator type, and reaction container type, which
are difficult to handle as numerical data, were expressed as
categorical data. The main phase determined by cluster analysis
of PXRD patterns for different conditions was analyzed as an
objective variable. The results of random forest analysis allowed
the number of explanatory variables with large contributions to
be decreased to eight, namely reaction temperature
(contribution = 30.3%), solvent type (28.6%), reaction container
type (11.7%), ligand concentration (11.0%), metal/ligand ratio
(9.5%), metal ion concentration (7.3%), isothermal reaction time
(0.9%), and modulator amount (0.4%) (Figure S4). For
visualization, decision tree analysis was performed using the
screened explanatory variables. Figure 2(b) shows the decision
tree based on the dataset containing 114 synthetic conditions
and the corresponding PXRD patterns, indicating that Ag2S
mainly forms above 120 °C. This behavior strongly suggests that
high temperatures induce the decomposition of H3ttc. In most
cases, the addition of water as a solvent led to Ag2S formation,
which suggests that H3ttc is hydrolyzed at high temperatures to
produce H2S. Although we did not consider the possibility of this
hydrolysis taking place prior to carrying out decision tree
analysis, H3ttc has been reported to gradually produce H2S in
water at high temperatures by aromatic nucleophilic
substitution–based hydrolysis (Figure S8),[13] which can be
enhanced by Ag+ coordination.[14a, 14b] Therefore, decision tree
analysis indicated the importance of preventing hydrolysis and
suggested that unknown phases 1 and 2 are more likely to
form under the conditions in the lower right branch of the
decision tree. In summary, we inferred that Ag2S is more likely to
form at temperatures above 120 °C, while unknown phases tend
to form in the presence of water or acetonitrile, and the low-
crystalline phase easily forms in the absence of water. On the
other hand, the reaction container type did not appear on the
decision tree node, although random forest analysis suggested
that this parameter affected the reaction outcome (score =
11.7%, see Figure S4). This discrepancy was explained as
follows. All experiments using screw cap glasses were
conducted at low temperatures (≤70 °C) in view of the low
resistance of these glasses to pressure. The p-value of logistic
regression with “temperature,” as the explanatory variable, and
“reaction container type,” as the objective variable, was below
0.0001, indicating a strong correlation between temperature and
reaction container type. Therefore, it is most likely that the
Figure 1. Schematic application of machine learning to the selective
syntheses of novel Ag–S coordination polymers.
Results and Discussion
First exploratory synthesis and machine-learning-based
analysis for 114 synthetic conditions
Initially, we searched for conditions under which Ag+ and H3ttc
form crystalline CPs. Specifically, the CPs were synthesized
using different temperatures (45–130 °C), solvents (acetone,
water, dimethylformamide, dimethyl sulfoxide, acetonitrile, and
their mixtures), metal ion/ligand ratios (0.25–4 mol/mol), reaction
times (48–108 h), metal ion and ligand concentrations (1.6–300
mM), and reaction containers (screw cap glass tube or Teflon
autoclave). Overall, 114 conditions were tested. Although we
intended to fabricate single crystals, only powdered crystals
unsuitable for single-crystal structural analysis were obtained.
Therefore, we attempted to extract chemical insights from failure
experiments. As precipitates were obtained under all conditions,
they were collected by centrifugation and analyzed by PXRD,
which revealed the presence of single-phase Ag2S, a low-
crystalline phase, an unknown phase, or complicated mixtures.
To analyze the experimental data using supervised machine
learning, one must define what phases correspond to successful
results. However, arbitrariness was inevitable when the PXRD
patterns of these mixtures were assessed by humans to
determine reaction success/failure, and thus, an objective
approach was necessary. Therefore, we performed cluster
analysis to automatically classify the PXRD patterns into several
categories based on similarities (Figure S1, Supporting
Information) and further verified placement accuracy on a point-
by-point basis. The patterns were classified into five categories,
namely two unknown phases [unknown phase 1 (23 cases)
and unknown phase 2 (7 cases)], a low-crystalline phase (56
2
This article is protected by copyright. All rights reserved.

10.1002/anie.202110629
Angewandte Chemie International Edition
RESEARCH ARTICLE
container type and the experimental results were indeed
associated but not causally related, i.e., their relationship was
spurious. In this case, temperature would be the originally
dominant factor. Indeed, random forest analysis showed that
temperature had the highest contribution of 30.3% (Figure S4).
concentrations (Figure S9). While no single crystals of unknown
phases were obtained, the 326 conditions were subjected to
cluster analysis, random forest analysis, and decision tree
analysis (Figures 2(c), S2, S5, and S7). The patterns were again
classified into five categories, namely unknown phase 1 (63
cases), unknown phase 2 (19 cases), a low-crystalline phase
(175 cases), Ag2S (61 cases), and a minor uncertain impurity
phase (8 cases) (Figure S2). According to cluster analysis, four
patterns fell into a category different from that assigned by
analysis of the 114 conditions. These four entries were PXRD
patterns of mixtures of two or more phases (Figure S3). The
number of cases where the classification differed from that
observed for 114 conditions was relatively small, which
demonstrated the good reproducibility of cluster analysis.
The results of decision tree analysis (Figures 2(c) and S6)
indicated that Ag2S formed in the presence of water above 90 °C.
Below 90 °C, unknown phase 2 was obtained in the presence
of AcOH and HBA as modulators. In contrast, the low-crystalline
phase was mainly obtained in non-aqueous solvents or at high
concentrations of Pyr, EDA, DEA, or TEA. Unknown phase 1
was predominantly obtained when water was used as the
solvent in the presence of a moderate modulator amount, which
suggested that proton concentration had a significant effect on
product selectivity. The information extracted from decision trees
revealed the conditions most likely to yield each phase.
Specifically, Ag2S formed in the presence of water at high
temperature; unknown phase 1 formed in the presence of
water under weakly basic to acidic conditions; unknown phase
2 formed in the presence of water under acidic conditions; and
the low-crystalline phase formed under basic conditions or in the
absence of water. These results suggested that proton
concentration is an important factor for the selective syntheses
of the unknown phases.
Second exploratory synthesis and machine learning
analysis for 326 synthetic conditions
We performed 212 additional experiments to synthesize single
crystals of unknown phases 1 and 2. Based on the results of
decision tree analysis, the reaction temperature was set below
100 °C to avoid ttc hydrolysis, and water- or acetonitrile-
containing solvents were used. As container type did not appear
to be a dominant factor in the nodes of the decision tree in
Figure 2(b), all additional experiments were conducted using
Teflon containers. We also focused on certain parameters with
low (according to random forest analysis) contributions. The low
contribution of a given parameter not only suggests that it
cannot influence the experimental results, but also implies that
the range of the explored experimental conditions may have
been too narrow to contribute. For this reason, parameters with
contributions evaluated as low by random forest analysis should
not be ignored but rather be expanded to the currently
unexplored conditions. For example, in 114 experiments, we
added a monodentate ligand (acetic acid) as a modulator to
improve crystallinity.[15] The results of random forest analysis
performed for the 114 conditions indicated that the contribution
of the modulator was small. However, as the addition of acetic
acid was only examined under four conditions, it was impossible
to determine whether modulators actually play an important role
in determining product selectivity. Therefore, we performed
additional experiments using various modulator types [acetic
acid
(AcOH),
benzoic
acid
(HBA),
sodium
N,N-
diethyldithiocarbamate (DDC), pyridine (Pyr), diethylamine
(DEA), triethylamine (TEA), ethylenediamine (EDA)] and
3
This article is protected by copyright. All rights reserved.

10.1002/anie.202110629
Angewandte Chemie International Edition
RESEARCH ARTICLE
Figure 2. (a) Representative powder X-ray diffraction patterns for various categories. Decision trees based on these patterns and (b) 114 or (c) 326 synthetic
conditions. The entireties of the corresponding decision trees are shown in Figures S6 and S7, respectively.
Preparation of single crystals and structural analysis
parameters with small (according to random forest analysis)
contributions proved to be useful. Although the products were
obtained as mixtures with silver halides, their crystal structures
could still be determined (Table 1), which allowed us to validate
the results of decision tree analysis as follows.
The composition of unknown phase 1 (KGF-6) was determined
as [Ag2Httc]n, i.e., one proton was present per ttc unit, in line
with the moderate proton concentration used for the synthesis of
this phase, as suggested by decision tree analysis. The Ag
atoms of KGF-6 were shown to feature an unusual five-fold
coordination in a trigonal bipyramidal structure (the N atoms of
the triazine ring were not coordinated), with the Ag ions
According to the random forest analysis of 326 conditions, the
contribution of metal ion type was small (2.0%, Figure S5),
possibly because only two Ag+ sources were explored, i.e., the
influence of the Ag+ source could not be sufficiently evaluated.
Therefore, we used insoluble silver halides and successfully
obtained single crystals of unknown phases mixed with the silver
halide precursors. The formation of large single crystals was
most likely due to the deceleration of Ag–S bond formation
caused by the sustained release of Ag+ from the insoluble silver
halides.[16] As machine learning analysis was not performed with
crystal size as the objective variable, the formation of single
crystals in the presence of silver halides was an accidental result. disordered along the axial direction of this structure (Figures 3(a)
However, to get a result different from that of previous
experiments, the guideline of exploring a wider range of
and S10). The Ag–S bond lengths in the triplanar structure were
obtained as 2.50–2.66 Å, and those between the Ag and axial S
4
This article is protected by copyright. All rights reserved.

10.1002/anie.202110629
Angewandte Chemie International Edition
RESEARCH ARTICLE
atoms were determined as 2.77–3.35 Å (this Ag atom was
disordered over three positions). In addition, the Httc ligand was
revealed to have two crystallographically non-equivalent S
atoms, S1 (μ2-S atom) and S2 (μ4-S atom), which could
coordinate Ag ions to afford a AgS layer in the ac-plane (Figure
3(b)). Furthermore, the layers were stacked with S atoms as
shared vertices to form a three-dimensional Ag–S network
(Figure 3(d)). Moreover, one-dimensional channels were present
between the AgS layers, where the Httc triazine rings aligned
The composition of unknown phase 2 (KGF-7) was determined
as [AgH2ttc]n, i.e., two protons were present per ttc unit (Figures
3(e) and S11). The two Ag atoms formed Ag–S bonds (2.52–
2.88 Å) with the four H2ttc S atoms to afford a dinuclear complex
connected by H2ttc units, i.e., a one-dimensional CP. The
presence of one proton in KGF-6 and two protons in KGF-7
agreed with the tendency of KGF-6 to form under near-neutral
conditions and that of KGF-7 to form under acidic conditions, as
suggested by the decision trees. Furthermore, a single crystal of
proton-free [Ag3ttc]n (KGF-8) was accidentally prepared in the
presence of 4,4’-bipyridine (bpy). Although the temperature
factor could not be optimized, and the electron density around
the Ag atoms was high because of the poor quality of the KGF-8
single crystal, the simulated PXRD pattern of this CP agreed
well with the broad PXRD patterns of some samples classified
as the low-crystalline phase (Figure S13). The formation of
proton-free single-crystalline KGF-8 was ascribed to the action
of bpy as a good proton acceptor. The fully deprotonated
structure of ttc in KGF-8 well agreed with the decision tree
results, showing that the low-crystalline phase was more likely to
form under basic or water-free conditions, i.e., at low proton
concentrations. KGF-8 also featured Ag–Ag and Ag–N bonds,
unlike KGF-6 and -7 (Figures 3(f) and S12). KGF-6 and -7 were
stable in water and 0.1 M HCl for 24 h, but were deprotonated to
afford KGF-8 when immersed into 0.1 M NaOH (Figure S14).
Thermogravimetric analysis revealed that KGF-6, -7, and -8 are
stable up to 300 °C (Figure S15), confirming the thermal and
chemical stabilities of these thiolate-coordination-based CPs.
perpendicularly to the channel direction to form
a one-
dimensional columnar structure (Figure 3(c)). The distance
between the neighboring C and N atoms in the triazine rings was
determined as ~3.09 Å, indicating strong aromatic interactions.
Such a columnar structure is also expected to act as an electron
carrier path.[17]
Table 1. Crystallographic data for KGF-6, -7, and -8.
Compound
CCDC number
Formula
KGF-6
2052285
C3HAg2N3S3
390.99
150
KGF-7
2052284
C3H2AgN3S3
284.13
150
KGF-8
2071617
C3Ag3N3S3
497.85
150
Formula weight/Da
Temperature/K
Crystal system
Space group
a/Å
orthorhombic
Pnma
monoclinic
C2/m
orthorhombic
Pna21
6.9154(2)
13.0731(4)
7.4840(2)
90
7.080(4)
12.924(7)
7.431(4)
90
7.5715(10)
7.9093(7)
12.5668(13)
90
b/Å
c/Å
α/o
β/o
90
103.196(6)
90
90
γ/o
90
90
Volume/Å3
Z
676.60(3)
4
661.9(7)
4
752.57(14)
4
ρcalc/g cm−3
μ/mm−1
3.838
2.851
4.394
6.627
3.898
8.488
F(000)
728.0
544.0
912.0
Reflections collected
Unique reflections
Goodness-of-Fit on F2
21245
807
3541
6764
732
2158
1.253
0.994
1.008
Final R1 index
[I ≥ 2σ (I)]
0.0263
0.0276
0.0815
0.0246
0.0255
0.0637
0.0710
0.1179
0.1916
Figure 3. (a) Ligand coordination environment of [Ag2Httc]n (KGF-6). (b)
Portion of KGF-6 structure along the b-axis. (c) One-dimensional column of
Httc molecules in KGF-6. (d) Perspective view of the KGF-6 crystal structure
along the a-axis. (e) Perspective view of the [AgH2ttc]n (KGF-7) crystal
structure along the a-axis. (f) Perspective view of the [Ag3ttc]n (KGF-8) crystal
structure along the a-axis. Ag: green, S: yellow, N: blue, C: gray, H: pink. The
displayed Ag atoms represent the major occupancy.
Final R1 index [all data]
Final wR2 index
[I ≥ 2σ (I)]
Final wR2 index
0.0807
0.0629
0.1625
[all data]
5
This article is protected by copyright. All rights reserved.

10.1002/anie.202110629
Angewandte Chemie International Edition
RESEARCH ARTICLE
Evaluation of electronic properties
Based on the diffuse reflectance ultraviolet-visible-near-infrared
spectra of KGF-6, -7, and -8, their optical bandgaps were
determined as 2.76, 2.79, and 2.95 eV, respectively (Figure
S19). Photoelectron yield spectra revealed that the valence
band maximum (VBM) levels of all three isomers were nearly
identical, equaling −5.92, −5.98, and −5.93 for KGF-6, -7, and -8,
respectively (Figures S20 and S21). The band structures and
densities of states of each isomer were evaluated using
CASTEP first-principle calculations,[18] and the VBM and
conduction band minimum (CBM) of KGF-7 and -8 were found to
be relatively flat. Contrastingly, KGF-6 exhibited a steeply
dispersed band structure with VBM and CBM levels of ~0.4 eV
(Figures 4(a), (c), and (e)), which indicated semiconductor-like
behavior. The effective masses of holes and electrons were
obtained by density functional theory calculations as mh* =
0.96m0 and me* = 0.81m0 (m0 = electron rest mass), respectively,
and were therefore comparable to those of a previously reported
highly-conductive MOF, Fe3(THT)2(NH4)3 (THT = 2,3,6,7,10,11-
triphenylenehexathiol), which was claimed to have a mobility as
high as ~220 cm2 V−1 s−1.[19] Notably, the VBM and CBM of KGF-
6 mainly originated from the Ag and S atoms and the C and N
atoms, respectively (Figure 4(b)). Furthermore, the highest
occupied molecular orbital and lowest unoccupied molecular
orbital distributions indicated that these orbitals are distributed
across the AgS network and the triazine ring, respectively
(Figure S22), to form an alternating p–n heterostructure at the
molecular level for charge separation. The segregated structure
and the relatively low effective masses of holes and electrons
could provide ambipolar pathways, which may be a useful
attribute in the context of photocatalysis and solar cells.
Moreover, time-resolved microwave conductivity (TRMC)
measurements were performed to evaluate the CP
photoconductivities. The maximum values of φΣμmax (φ = charge
carrier generation yield, Σμ = sum of hole and electron
mobilities) equaled 1.6 × 10−4, 2.7 × 10−5, and 2.8 × 10−5 cm2 V−1
s−1 for KGF-6, -7, and -8, respectively (Figure 5). The value
obtained for KGF-6 is one order of magnitude higher than those
of KGF-7 and -8 and is comparable to those of previously
reported conductive CPs, MOFs, and covalent organic
frameworks.[20] The TRMC results were consistent with the band
dispersions (flat for KGF-7 and -8; steeper for KGF-6) obtained
by first-principle calculations. The effective lifetime (i.e., the
weight-averaged lifetimes of double-exponential fits) of KGF-6
was determined as 10.2 μs; thus, this compound exhibited both
a long-lived charge separation state and high mobility. This
finding was consistent with the crystal structure having
segregated paths for hole and electron transport.[21] With
decreasing excitation intensity (~1016–1014 photon cm−2 pulse−1),
φΣμmax slightly increased, but the half-life times decreased,
which suggests that charge carrier decay was mainly due to
charge trapping (Figure S26).
Figure 4. (a, c, e) Band structures and (b, d, f) density of states of (a, b) KGF-
6, (c, d) KGF-7, and (e, f) KGF-8. In the band structures, the energy level of
the valence band maximum is represented as zero.
Figure 5. Results of time-resolved microwave conductivity measurements (λex
= 355 nm) for KGF-6 (red), KGF-7 (orange), and KGF-8 (blue).
Conclusion
Semiconductive CPs containing Ag–S bonds were synthesized
using machine learning to extract chemical insights from failure
experiments. Single-crystal structure analysis demonstrated that
the number of protons in the isomer crystal structures was
consistent with the proton concentration dependency in the
synthetic conditions suggested by decision tree analysis. One
topological isomer, KGF-6, possessed an infinite three-
dimensional Ag–S network and a one-dimensional columnar
triazine ring structure, therefore exhibiting high photoconductivity,
as evidenced by first-principle calculations and TRMC
measurements. The proposed machine-learning-assisted
6
This article is protected by copyright. All rights reserved.

10.1002/anie.202110629
Angewandte Chemie International Edition
RESEARCH ARTICLE
J. Yan, X. S. Du, D. H. Wei, S. Q. Zang, T. C. W. Mak, Angew. Chem.
Int. Ed. 2018, 57, 8560; c) Z. Y. Wang, M. Q. Wang, Y. L. Li, P. Luo, T.
T. Jia, R. W. Huang, S. Q. Zang, T. C. W. Mak, J. Am. Chem. Soc.
2018, 140, 1069.
method is expected to promote the discovery of novel CPs that
are difficult to crystallize with selective topologies.
[13] E. M. Smolin, L. Rapoport, Chemistry of Heterocyclic Compounds: s-
Triazines and Derivatives, Interscience Publishers, Inc., New York,
1959.
Acknowledgements
[14] a) D. Tanaka, S. Masaoka, S. Horike, S. Furukawa, M. Mizuno, K. Endo,
S. Kitagawa, Angew. Chem. Int. Ed. 2006, 45, 4628; b) R. D. Gillard,
Coord. Chem. Rev. 1975, 16, 67.
This work was supported by JST PRESTO (Grant Nos.
JPMJPR15N6, JPMJPR17NA) and JSPS KAKENHI (Grant Nos.
20H02577, 20H04680, 20H04646, 20H05836, 17K00320, and
20J13900). We would like to thank Editage (www.editage.com)
for English language editing.
[15] T. Tsuruoka, S. Furukawa, Y. Takashima, K. Yoshida, S. Isoda, S.
Kitagawa, Angew. Chem. Int. Ed. 2009, 48, 4739.
[16] K. Nakashima, T. Shimizu, Y. Kamakura, A. Hinokimoto, Y. Kitagawa,
H. Yoshikawa, D. Tanaka, Chem. Sci. 2020, 11, 37.
[17] S. S. Park, E. R. Hontz, L. Sun, C. H. Hendon, A. Walsh, T. Van
Voorhis, M. Dinca, J. Am. Chem. Soc. 2015, 137, 1774.
Keywords: semiconductors • crystal engineering • coordination
polymers • machine learning • silver thiolate network
[18] S. J. Clark, M. D. Segall, C. J. Pickard, P. J. Hasnip, M. J. Probert, K.
Refson, M. C. Payne, Z. Kristallogr. 2005, 220, 567.
[1]
[2]
a) L. S. Xie, G. Skorupskii, M. Dinca, Chem. Rev. 2020, 120, 8536; b)
M. O'Keeffe, O. M. Yaghi, Chem. Rev. 2012, 112, 675; c) S. Kitagawa,
R. Kitaura, S. Noro, Angew. Chem. Int. Ed. 2004, 43, 2334; d) S. L.
Griffin, N. R. Champness, Coord. Chem. Rev. 2020, 414, 213295.
a) X. Li, W. Ma, H. Li, Q. Zhang, H. Liu, Coord. Chem. Rev. 2020, 408,
213191; b) Y. Liu, D. Huang, M. Cheng, Z. Liu, C. Lai, C. Zhang, C.
Zhou, W. Xiong, L. Qin, B. Shao, Q. Liang, Coord. Chem. Rev. 2020,
409, 213220; c) O. Veselska, A. Demessence, Coord. Chem. Rev.
2018, 355, 240; d) G. Givaja, P. Amo-Ochoa, C. J. Gomez-Garcia, F.
Zamora, Chem. Soc. Rev. 2012, 41, 115; e) J. Xie, L. Wang, J. S.
Anderson, Chem. Sci. 2020, 11, 8350; f) Y. Kamakura, D. Tanaka,
Chem. Lett. 2020, 50, 523; g) Z. Ji, C. Trickett, X. Pei, O. M. Yaghi, J.
Am. Chem. Soc. 2018, 140, 13618.
[19] R. Dong, P. Han, H. Arora, M. Ballabio, M. Karakus, Z. Zhang, C.
Shekhar, P. Adler, P. S. Petkov, A. Erbe, S. C. B. Mannsfeld, C. Felser,
T. Heine, M. Bonn, X. Feng, E. Canovas, Nat. Mater. 2018, 17, 1027.
[20] a) A. Saeki, Polym. J. 2020, 52, 1307; b) T. C. Narayan, T. Miyakai, S.
Seki, M. Dinca, J. Am. Chem. Soc. 2012, 134, 12932; c) X. Ding, X.
Feng, A. Saeki, S. Seki, A. Nagai, D. Jiang, Chem. Commun. 2012, 48,
8952.
[21] S. Wang, T. Kitao, N. Guillou, M. Wahiduzzaman, C. Martineau-Corcos,
F. Nouar, A. Tissot, L. Binet, N. Ramsahye, S. Devautour-Vinot, S.
Kitagawa, S. Seki, Y. Tsutsui, V. Briois, N. Steunou, G. Maurin, T.
Uemura, C. Serre, Nat. Commun. 2018, 9, 1660.
[3]
[4]
a) L. Sun, T. Miyakai, S. Seki, M. Dinca, J. Am. Chem. Soc. 2013, 135,
8185; b) N. E. Horwitz, J. Xie, A. S. Filatov, R. J. Papoular, W. E.
Shepard, D. Z. Zee, M. P. Grahn, C. Gilder, J. S. Anderson, J. Am.
Chem. Soc. 2019, 141, 3940.
a) A. Pathak, J. W. Shen, M. Usman, L. F. Wei, S. Mendiratta, Y. S.
Chang, B. Sainbileg, C. M. Ngue, R. S. Chen, M. Hayashi, T. T. Luo, F.
R. Chen, K. H. Chen, T. W. Tseng, L. C. Chen, K. L. Lu, Nat. Commun.
2019, 10, 1721; b) X. Huang, H. Li, Z. Tu, L. Liu, X. Wu, J. Chen, Y.
Liang, Y. Zou, Y. Yi, J. Sun, W. Xu, D. Zhu, J. Am. Chem. Soc. 2018,
140, 15153.
[5]
a) Y. Kamakura, P. Chinapang, S. Masaoka, A. Saeki, K. Ogasawara,
S. R. Nishitani, H. Yoshikawa, T. Katayama, N. Tamai, K. Sugimoto, D.
Tanaka, J. Am. Chem. Soc. 2020, 142, 27; b) S. Hu, F. Y. Yu, P. Zhang,
D. R. Lin, Dalton Trans. 2013, 42, 7731; c) A. Eichhofer, S. Lebedkin,
Inorg. Chem. 2018, 57, 602.
[6]
[7]
[8]
L. Sun, M. G. Campbell, M. Dinca, Angew. Chem. Int. Ed. 2016, 55,
3566.
a) D. Li, W. J. Shi, L. Hou, Inorg. Chem. 2005, 44, 3907; b) R. D.
Shannon, Acta Crystallogr. A 1976, 32, 751.
X. Huang, Y. Qiu, Y. Wang, L. Liu, X. Wu, Y. Liang, Y. Cui, Y. Sun, Y.
Zou, J. Zhu, W. Fang, J. Sun, W. Xu, D. Zhu, Angew. Chem. Int. Ed.
2020, 59, 22602-22609.
[9]
a) G. Panapitiya, G. Avendano-Franco, P. Ren, X. Wen, Y. Li, J. P.
Lewis, J. Am. Chem. Soc. 2018, 140, 17508; b) Y. Xie, C. Zhang, X. Hu,
C. Zhang, S. P. Kelley, J. L. Atwood, J. Lin, J. Am. Chem. Soc. 2020,
142, 1475.
[10] a) D. W. Davies, K. T. Butler, J. M. Skelton, C. Xie, A. R. Oganov, A.
Walsh, Chem. Sci. 2018, 9, 1022; b) P. Raccuglia, K. C. Elbert, P. D.
Adler, C. Falk, M. B. Wenny, A. Mollo, M. Zeller, S. A. Friedler, J.
Schrier, A. J. Norquist, Nature 2016, 533, 73; c) K. Muraoka, Y. Sada,
D. Miyazaki, W. Chaikittisilp, T. Okubo, Nat. Commun. 2019, 10, 4459;
d) S. M. Moosavi, A. Chidambaram, L. Talirz, M. Haranczyk, K. C.
Stylianou, B. Smit, Nat. Commun. 2019, 10, 539.
[11] Y. Kitamura, E. Terado, Z. Zhang, H. Yoshikawa, T. Inose, H. Uji-i, M.
Tanimizu, A. Inokuchi, Y. Kamakura, D. Tanaka, ChemRxiv Preprint
2020, DOI: 10.26434/chemrxiv.13490925.
[12] a) R. W. Huang, Y. S. Wei, X. Y. Dong, X. H. Wu, C. X. Du, S. Q. Zang,
T. C. W. Mak, Nat. Chem. 2017, 9, 689; b) R. W. Huang, X. Y. Dong, B.
7
This article is protected by copyright. All rights reserved.

10.1002/anie.202110629
Angewandte Chemie International Edition
RESEARCH ARTICLE
Entry for the Table of Contents
Semiconductive trithiocyanuric acid (H3ttc)-based coordination polymers (CPs) featuring Ag–S bonds are synthesized using machine-
learning-assisted techniques. One of the obtained CPs, [Ag2Httc]n (KGF-6), contains a three-dimensionally extended infinite Ag–S
bond network with one-dimensional column structures of stacked triazine rings, which provides segregated paths for holes and
electrons.
Institute and/or researcher Twitter usernames: @DaisukeTanakaKG
8
This article is protected by copyright. All rights reserved.