Tohoku University Repository (TOUR) / 東北大学機関リポジトリ

福田, 大夢

若島, 孔文

Japanese

ひきこもり者を持つ家族のソーシャル・サポートと受療行動に関する研究

井上, 泰⼈

李, 姿億

⽇本⼈と台湾⼈の夫婦の離婚の実務と理論

Tohoku University博士（情報科学）thesi

細矢, 悠介

English

Rethinking Object Detection from Images: Challenges in Open-Set and Open-Vocabulary Scenarios

李, 云蒙

Chat Translation Evaluation: From Datasets to Metrics

Tohoku University博士（工学）要約のみthesi

ZHAO, ZHIJIAN

論文内容要約 
  Many of today’s applications rely on magnetic field sensors, ranging from navigation systems, current 
sensing, and compasses to military instruments. Among these, magnetic field sensors for biomedical 
applications have gained considerable attention due to their ability to provide reference-free and coherent 
measurements independent of the dielectric properties of biological systems. However, biomagnetic sensing 
is constrained by challenges including signal levels often falling into the sub-nano or picoteslas range, 
particularly for DC and low-frequency magnetic fields. Thus, the application of magnetic field sensors in 
biomedicine is limited by the high cost of device fabrication and operational requirements in specialized 
environments. For example, the superconducting quantum interference device (SQUID), which is the most 
sensitive magnetic field sensor across a wide frequency range (from DC to several GHz), demands 
magnetically shielded rooms and cooling with liquid nitrogen or helium, making it less practical. Similarly, 
commercially available sensors such as Hall effect sensors, Fluxgate sensors, AMR (Anisotropic 
Magnetoresistance), and GMR (Giant Magnetoresistance) sensors each offer specific advantages but are also 
limited in sensitivity, frequency range, or environmental requirements. 
On the other hand, due to the progress in micro- and nano-fabrication technologies, the development of 
highly sensitive micro-fabricated magnetic field sensors becomes a rapidly expanding topic. These magnetic 
field sensors offer advantages including low-cost fabrication methods, flexibility of usage, and the quick 
obtaining of test results. One of the typical micro-fabricated magnetic field sensors is the Loretz force-based 
sensor. These sensors typically include resonators that respond to the Lorentz force generated by the 
interaction between an external magnetic field and an electric current. This force modifies the sensor in 
changes of resonant amplitude or frequency, which can be detected using capacitive, piezoresistive, or optical 
sensing techniques. These sensors provide a wide dynamic range and high resolution. However, the 
           ちょう  ち けん 
氏 名          ZHAO ZHIJIAN 
研究科，専攻の名称 東北大学大学院工学研究科（博士課程）機械機能創成専攻 
学 位 論 文 題 目           Resonant Magnetic Sensor Using Soft Magnetic Material with 
 Piezoresistive Readout 
（軟磁性体を用いたピエゾ抵抗検出型の共振磁気センサ） 
論 文 審 査 委 員          主査 東北大学教授 小野 崇人 東北大学教授 田中 秀治 
             東北大学教授 金森 義明 東北大学副教授 戸田 雅也                               
requirement for current flow through the resonator introduces Joule heating, increasing the 
thermomechanical noise and limiting performance. Another promising category is magnetoelectric (ME) 
composite-based sensors, which combine magnetostrictive and piezoelectric materials. These composites 
generate an electric polarization in response to an external magnetic field, leveraging interactions among 
electric, magnetic, and elastic subsystems. ME-based sensors achieve superior resolution, reaching hundreds 
of picoteslas. Despite their promise, they face constraints such as limited measurement ranges, and the 
magnetostriction of thin-film magnetic material is still poorly understood and requires further research. 
Other methodologies, including magnetic force- or torque-based sensing, are also explored in 
micro-fabricated sensors. However, these approaches continue to face challenges in enhancing sensitivity 
and achieving lower detection limits, underscoring the need for further innovation in this rapidly evolving 
field. 
This research is aimed at developing a highly sensitive magnetic field sensor with a target resolution in 
the sub-nanotesla range and a frequency range covering DC or low-frequency magnetic fields. To achieve 
this, a novel micro-fabricated resonant magnetic sensor based on soft magnetic materials has been proposed 
and systematically investigated. 
In the initial phase, a prototype of the sensor was fabricated to test and verify its feasibility. The sensor 
was designed based on the mechanical reproduction of the Larmor precession, consisted of a silicon 
cantilever structure and a micro-pillared magnet. Micro assembling technique based on micro-manipulator 
and conductive adhesive was utilized to fix the pillar on the end of the silicon cantilever. A piezoresistive 
strain sensing element is integrated into the silicon cantilever for detecting vibration signals. The working 
mechanism of the prototype involved applying an alternating magnetic field via a coil, causing the 
micro-pillared magnet to undergo pendulum-like vibrations due to the magnetic torque. Additionally, the 
synchronized alternating magnetic field was introduced to the system, which synchronized with the 
vibration signal through a Gain adjustable amplifier.  The synchronized field causes the external magnetic 
field to change the effective spring constant of the resonator, leading to shifts in the resonant frequency. This 
configuration enabled external magnetic fields to be translated into detectable frequency variations. The 
sensing performance of the prototype was evaluated using a self-oscillation system based on the 
Phase-locked loop circuit.  The evaluation results demonstrated that the fabricated magnetic sensor has an 
electrical adjustable sensitivity, achieving a maximum sensitivity of 114.02 Hz/mT and a minimum 
detectable magnetic field of 35.7 nT based on the minimum Allan deviation of the resonator. These findings 
suggest the sensor's potential for high-sensitivity magnetic field measurements in various applications. 
Subsequently, a comprehensive redesign of the magnetic field sensor was undertaken with the objective of 
reducing its size and enhancing its integration. The redesigned sensor features a micro-fabricated 
cantilevered silicon resonator with an integrated soft magnet and a copper coil for generating an excitation 
magnetic field. Both the soft magnet and the coil were directly fabricated on the sensor chip using 
photoresist mold electrodeposition, eliminating the need for assembly. Piezoresistive strain sensing 
continues to be employed for detecting the resonator's vibrations. Two distinct operation systems of the 
sensor are proposed, depending on the role of the coil. The first method is the forced vibration method, which 
relies on the magnetic torque generated by the excitation magnetic field from the Cu coil, corresponding to 
the magnetic moment of the soft magnet. When the resonator operates at its resonance frequency, changes 
in the external magnetic field are converted into variations in the resonant amplitude through the change of 
the magnetic torque. The second method utilizes Q-factor enhancement based on active damping control, 
where the coil generates a feedback magnetic field derived from the vibration signal through a phase shifter 
and a gain amplifier. In this configuration, the feedback magnetic field is adjusted to reduce the effective 
damping of the resonator. As a result, changes in the external magnetic field are translated into variations in 
effective damping, enabling higher sensitivity. With the forced vibration method, the sensor was evaluated to 
achieve a sensing range of ±40 mT, and a maximum sensitivity of 6.96 mV/µT with a resolution of 2.87 
nT/Hz0.5. When using the Q-factor enhancement method, the sensitivity was improved to 30 mV/µT with a 
resolution of 1.67 nT/Hz0.5. Compared to the initial prototype, the redesigned sensor demonstrated 
significantly enhanced sensing performance, making it more suitable for applications requiring higher 
precision and integration. 
Finally, to further enhance the sensing performance, the electrodeposition of a high-permeability soft 
magnetic material composed of nickel, iron, and molybdenum (NiFeMo) was investigated, along with an 
evaluation of the magnetic anisotropy of patterned soft magnets. The effects of electrodeposition conditions 
on the composition of NiFeMo films were comprehensively examined by systematically varying parameters 
such as solution pH, electrolyte concentration, and current density. A range of composition ratios for NiFeMo 
films was examined to determine the optimal composition for achieving superior magnetic susceptibility. 
Compared to NiFe films, NiFeMo films with optimized composition exhibited significantly enhanced 
magnetic properties, offering the potential to double the sensitivity of the sensor, thereby reaching the 
resolution of hundreds of picoteslas. Additionally, the influence of geometric patterns on the magnetic 
properties of soft magnetic films was evaluated. Stripe-line patterned soft magnetic films demonstrated 
enhanced magnetic anisotropy, improving the directional sensitivity of the sensor. These findings 
demonstrate the potential to further elevate the performance of the proposed magnetic field sensor, paving 
the way for its application in high-resolution magnetic sensing. 
 


Resonant Magnetic Sensor Using Soft Magnetic Material with Piezoresistive Readout

ZHANG, YUKUAN

論文内容要約 
Garment manipulation is a fundamental aspect of both daily life and industrial production, yet it presents 
significant challenges for robotic systems due to the deformable and unpredictable nature of fabrics. Tasks 
such as folding, dressing, or industrial processes like ironing and sewing require precision and adaptability, 
characteristics that traditional robotic systems struggle to achieve. In response to these challenges, this 
dissertation proposes an innovative framework for teaching garment manipulation to dual-arm robotic 
systems by integrating simulation, reinforcement learning, and human guidance. The research not only 
addresses the technical obstacles in automating these tasks but also seeks to contribute to broader 
advancements in robotics, simulation technologies, and human-robot collaboration. 
The foundation of this work lies in the creation of a highly customized simulation environment capable of 
replicating complex garment manipulation tasks. Simulation is crucial in this context, as it provides a 
risk-free, scalable, and efficient platform for training and testing robotic control models. The developed 
simulation environment incorporates virtual sensors to measure key parameters such as contact, 
deformation, wrinkles, and positional data. By enabling realistic task replication, this environment 
accelerates the development of control algorithms while reducing the reliance on expensive and 
time-consuming real-world experiments. Tasks such as dragging garments onto a printing board or dressing 
them onto an ironing board were successfully modeled in this environment, forming the basis for training 
reinforcement learning models. 
Building upon this simulation framework, the research introduces a novel reinforcement learning 
  チュウ ユクアン 
氏 名   ZHANG YUKUAN 
研究科，専攻の名称 東北大学大学院工学研究科（博士課程）ロボティクス専攻 
学 位 論 文 題 目          Teaching Garment Manipulation to Dual-Arm Robotic System by 
                                 Combining a Simulator and Human Guidance
                    (シミュレータと人間指示を組み合わせた双腕ロボットシステムへの衣服操作教示） 
論 文 審 査 委 員          主査 東北大学教授 平田 泰久 東北大学教授 林部  充宏 
  東北大学教授 鏡  慎吾 東北大学助教 ホセ サラザル 
architecture tailored to the unique challenges of garment manipulation. Traditional reinforcement learning 
methods often fall short in handling the temporal and multi-phase nature of garment tasks. To address this, 
a multi-critic Deep Deterministic Policy Gradient (DDPG) algorithm was developed. This architecture 
divides the task into distinct phases, with each phase evaluated by a dedicated critic network. By 
aggregating the evaluations across all phases, the system learns to optimize actions that consider both 
immediate and future consequences. The multi-critic approach significantly enhances the system's ability to 
manage complex tasks, as demonstrated through extensive experiments in both simulated and real-world 
settings. Comparative studies revealed that this method outperforms conventional reinforcement learning 
algorithms in adaptability, precision, and overall task success. 
 
The transition from simulation to real-world applications is another critical challenge addressed in this 
dissertation. To bridge the gap between the controlled conditions of simulation and the variability of 
real-world environments, an adaptive incremental hybrid impedance control strategy was proposed. This 
method integrates force-position impedance control, allowing the robot to safely manipulate garments while 
maintaining the desired deformation states. By adapting to external forces and fabric elasticity, the control 
system ensures that the trajectories generated in simulation are faithfully executed in reality. This seamless 
transfer of learned models is essential for practical deployment in industrial and domestic contexts. 
 
Recognizing the limitations of purely autonomous systems, the research incorporates human-robot 
interaction (HRI) as a pivotal component. Human guidance allows for real-time intervention, enabling the 
robot to adapt to unforeseen environmental changes and task requirements. For instance, a human operator 
can correct the robot's actions during task execution, and the system uses these interventions to refine its 
control model and update the simulation environment. This feedback loop not only enhances the robot's 
adaptability but also ensures continuous improvement of the control algorithms. Experiments demonstrated 
that this hybrid approach, combining autonomous learning with human guidance, significantly improves 
task performance in dynamic and unpredictable conditions. 
 
Extensive experimental validation was conducted to evaluate the effectiveness of the proposed framework. 
Tasks such as dragging garments onto a printing board and dressing them onto a vertical ironing board 
were used as benchmarks. The results highlighted the framework's ability to handle a wide range of 
garment types and manipulation tasks, demonstrating its generalizability and robustness. Key metrics such 
as task completion time, error rates, and fabric condition were used to assess performance, with the 
proposed methods consistently achieving superior results compared to baseline approaches. The integration 
of simulation, reinforcement learning, and HRI proved to be particularly effective in addressing the core 
challenges of garment manipulation, including deformability, variability, and safety. 
 
Despite its successes, the research acknowledges several limitations and areas for improvement. One major 
challenge lies in the discrepancies between simulation and reality, such as variations in fabric material 
properties and external forces. These differences can lead to inconsistencies in performance when 
transitioning from the simulation environment to real-world applications. Additionally, the absence of 
real-to-simulation feedback limits the system's ability to dynamically adapt to complex and evolving 
real-world conditions. Addressing these gaps will require further advancements in simulation fidelity and 
the development of adaptive real-time control systems. Leveraging data-driven models for force prediction 
and generation is another promising avenue for future research. 
 
In conclusion, this dissertation makes significant contributions to the field of robotics by presenting a 
comprehensive framework for garment manipulation. The integration of simulation, reinforcement learning, 
and human guidance offers a scalable, adaptable, and efficient solution for automating complex garment 
handling tasks. This work not only advances the state of the art in robotic control but also lays the 
groundwork for practical applications in industries such as textile manufacturing, healthcare, and domestic 
assistance. By addressing both technical and practical challenges, this research represents a step forward in 
realizing the potential of robotics to improve efficiency, accessibility, and quality of life in diverse settings. 
 
 
 


Teaching Garment Manipulation to Dual-Arm Robotic System by Combining a Simulator and Human Guidance

OKAWARA, TAKU

論文内容要約 
 In recent years, robots have been gradually deployed to our daily lives thanks to the progress of robot 
technologies. Particularly, we often see mobile robots such as clean robots, servicing robots, and autonomous 
driving cars. These mobile robots accomplished their tasks automatically based on navigation systems in the 
case of structured environments (e.g., rooms and restaurants). Basically, autonomous navigation systems 
heavily depend on perception systems because their output is the input of control and planning systems (e.g., 
humans also easily cause a car accident in the case of a whiteout). Therefore, the perception 
system—particularly the ability to localize the robot’s position within its environment—is crucial for reliable 
navigation because the robot can easily collide with objects. While mobile robots can stably accomplish their 
tasks in structured environments, it is difficult for them to conduct autonomous navigation in extreme 
environments (e.g., featureless areas such as tunnels and lunar surfaces, noises produced by fogs and rains) 
with exteroceptive sensor failure or degradation. The reason is that perception systems heavily depend on 
exteroceptive sensors (e.g., LiDAR, camera). Therefore, we propose robust odometry estimation algorithms 
in extreme environments where exteroceptive sensors degrade and fail. In this thesis, we solve the 
aforementioned problems by fusing exteroceptive sensors and proprioceptive sensors to realize a robust 
odometry estimation. Particularly, we apply various kinematic models of mobile robots to sensor 
fusion-based odometry estimation algorithms. However, kinematic models of mobile robots are fragile to 
model identification errors such as kinematic parameters (e.g., wheel radius, link length of a leg) and 
terrain-dependent parameters (e.g., slip ratio, friction coefficients). Furthermore, terrain-dependent 
parameters are basically unknown; thus, this identification must be conducted online such that all 
  おお かわら   たく 
氏 名   大 河 原 拓 
研究科，専攻の名称 東北大学大学院工学研究科（博士課程）航空宇宙工学専攻 
学 位 論 文 題 目          Sensor Fusion-based Robust Odometry in Extreme Environments with 
  Exteroceptive Sensor Failure and Degradation（外界センサの故障や 
  劣化に頑強なセンサ融合によるオドメトリ推定） 
論 文 審 査 委 員          主査 東北大学教授 吉田 和哉 東北大学教授  林部 充宏 
東北大学教授 橋本 浩一  主任研究員 小出 健司 
（産総研） 
parameters fit the current conditions related to the robot and environment. Therefore, we calibrate the 
kinematic models online to accurately conduct odometry estimation even in extreme environments with 
exteroceptive sensor failure or degradation. We proposed many types of online calibration problems for 
achieving autonomous navigation by many types of robots (e.g., wheeled robots, and legged robots). 
 
  First, we propose an online calibration algorithm for an explicit kinematic model of wheel robots. 
Particularly, we tried to apply our algorithm to a skid-steering robot whose kinematic model is complex due 
to producing wheel slippage for rotational motions. We present a tightly coupled LiDAR-IMU-wheel 
odometry algorithm incorporating online calibration of the model-based kinematic model named the full 
linear model. The full linear model describes robot motion without explicit and rigorous modeling; thus, this 
model is robust to modeling errors related to complex terms (e.g., wheel slippage). We proposed a full linear 
wheel odometry factor, which not only serves as a motion constraint but also performs the online calibration 
of kinematic models for factor graph optimization. We also estimate the uncertainty (i.e., covariance matrix) 
of the wheel odometry-based motion prediction online to create a constraint with a reasonable statistical 
model even in rough terrains. To create the proposed factor graph, matching cost factor and IMU 
preintegration factor are used as LiDAR and IMU constraints, respectively. In addition, the kinematic 
parameter fixation factor is created to help optimize stability in the case of exteroceptive sensor failure and 
degradation. The constant kinematic parameter factor is created to describe the time transition of the 
kinematics parameter to enable the kinematics parameter to adapt to terrain condition changes. 
  According to the experimental results, we demonstrated that the proposed method is robust in two severe 
situations: 1) exteroceptive sensor failure and degradation, and 2) large model errors (i.e., all-wheel radii 
were enlarged by 25%). We also demonstrated that our online uncertainty estimation for kinematic models 
estimates reasonable uncertainty of the full linear wheel odometry factor. Specifically, while its uncertainty 
is low on an indoor flat plane (i.e., ideal condition for wheel odometry model), its uncertainty is large on grass 
(i.e., rough terrain).  Furthermore, the proposed method enables robust odometry estimation in a condition 
that terrains change (bricks, outdoor stone tiles, and indoor stone tiles), thanks to terrain adaptability based 
on online calibration. 
 
 In contrast to the full linear model (explicit kinematic model), we introduce a learning-based kinematic 
model of wheel robots because a neural network can express nonlinearly complex terms (e.g., large wheel 
slippage such as drifting) of a wheel odometry model. The proposed network is trained online to consider 
terrain-dependent parameters (e.g., friction coefficients, slip ratio, and slip angle) to demonstrate the 
versatility of terrain changes. We tightly fuse the LiDAR and IMU constraints, and online learned neural 
kinematic model for enhancing robustness toward nonlinear motions (e.g., drifting) in addition to 
exteroceptive sensor degradation and failure. According to the proposed constraint named neural adaptive 
wheel odometry factor, we jointly solve odometry estimation and online training of the learning-based wheel 
odometry model on a unified factor graph. To realize online learning calculation in real-time, we designed 
the network divided into an offline learning model and an online learning model. While the online learning 
model describes terrain-dependent features, the offline learning model describes features invariant to 
terrain (e.g., fundamental kinematics such as the relationship between wheel speeds and robot velocity).  
  According to the experimental results, we demonstrated that the proposed network adapted to changing 
terrain conditions such as concrete (solid terrain) to grass (soft terrain), resulting in an accurate odometry 
estimation across different environments. We also validated that the proposed odometry estimation 
algorithm was robust against two extreme situations: 1) exteroceptive sensor failure and degradation, and 2) 
nonlinearity (e.g., large wheel slippage by drifting) of the kinematic model. Furthermore, we confirmed that 
our online trained model is trained such that this model consists of unique parameters describing each 
environment (e.g., indoor flat floor, bricks, outdoor stone tiles, indoor stone tiles, concrete, and grass).   
 
  Finally, we conduct robust odometry estimation for legged robots dependent on the above works that 
incorporate kinematic models of wheeled robots into LiDAR-IMU constraints. We propose an online 
learning-based leg odometry model incorporating tactile information (foot reaction force) to address 
challenging conditions where a foot velocity relative to the ground is non-zero (e.g., terrain deformation, and 
foot slippage). Conventional leg odometry models cannot describe accurate motion prediction because the 
assumption (no foot slippage) is invalid in such challenging conditions. Online training of this model 
enhances its versatility for weight loads of a robot (e.g., assuming delivery or transportation tasks) in 
addition to terrain conditions. This versatility is needed because foot reaction forces vary depending on the 
robot's mass. According to the neural adaptive leg odometry factor and online uncertainty estimation of the 
leg odometry-based motion prediction, we also jointly solve online training of this leg kinematics model and 
odometry estimation.  
  Experimental results showed that the proposed method outperforms state-of-the-art approaches (e.g., 
Unitree proprietary odometry incorporating LiDAR-IMU constraints, FAST-LIO2) in challenging terrains 
(e.g., deformable terrains such as sandy beaches, and gravel) and featureless environments. Furthermore, 
we validated that the proposed method is robust to extreme situations where exteroceptive sensor 
degradation and terrain condition changes occurred jointly, thanks to the proposed online learning-based leg 
odometry model with tactile information. We assume that adding tactile information to wheel robots’ cases 
such as reaction forces or torque applied to wheels or their axes can deal with more challenging conditions 
(e.g., frozen roads covered by tunnels). Furthermore, tactile information can provide information regarding 
each wheel's contact with the ground; tactile information helps kinematic model accuracy in the case of 
rough terrains, where all wheels are not insured to contact with the ground. 
 
  In summary, we proposed the model-based and learning-based online calibration of the kinematic model 
for wheeled robots. We consider that while the model-based approach is suitable for low-speed operation in 
which wheel slippage can be approximated to linear, the learning-based approach is suitable for high-speed 
operation in that nonlinear terms (e.g., large wheel slippage) are included in kinematic models. Furthermore, 
we demonstrated that the learning-based approach can be applied to legged robots’ cases. Particularly, we 
proposed incorporating tactile information (foot reaction forces) to take into account the dynamics between 
robot feet and the ground. Because tactile information directly includes terrain-dependent features, we also 
demonstrated our kinematic model conducted accurate odometry estimation even in challenging conditions, 
where featureless environment and terrain condition changes jointly occurred. According to the all 
experimental results, we successfully demonstrated that our odometry estimation algorithms can apply to 
many application cases such as 1) terrain conditions can be changed in the middle of robot applications, 2) 
various robot types such as skid-steering robots (two-wheeled, four-wheeled, six-wheeled, robots, tracked 
robots, quadruped robots), and 3) high-speed operation inducing nonlinear motions.  
  Our learning-based kinematic models are trained in a specific robot (e.g., Rover mini made by Rover 
robotics, Go2 made by Unitree) that is used in offline learning procedures. Hence, our learning method can 
be applied to only a specific robot. To address this versatility problem, we will train many types of data 
regarding robot sizes (e.g., large wheel radius, small wheelbase), robot types (e.g., two-wheeled, four-wheeled 
robots, and quadruped robots), and materials (e.g., wheels covered by rubber, wheels made by plastic only) 
as well as terrain types. However, training data will become a huge amount because the number of each 
type affects the number of entire data types by multiplication. For example, in the case of robot types: 4 (e.g., 
two-wheeled, four-wheeled robots and, quadruped, hexapod robot), terrain types: 10 (e.g., same as our 
experiments), robot size: 3 (e.g., small, medium, and large size), the number of entire data types are 120 (= 4 
× 10 × 3). Therefore, we plan to use simulation environments to collect datasets without real data to reduce 
time costs because the recent progress of simulation environments can realize realistic phenomena (e.g., 
wheel rut). 
 
 
 


Sensor Fusion-based Robust Odometry in Extreme Environments with Exteroceptive Sensor Failure and Degradation

董, 庭瑞

Study of Generalized Quantum Measurement of Qubit Systems in a Small-Scale Quantum Computer

笠原, 遥太郎

論文内容要約 
【① General introduction】 In developing organic functional materials, the design and construction of molecular 
assembly structures based on substituents, molecular shape, and symmetry are important from the perspective of physical 
property manifestation, in addition to molecular synthesis. High-symmetry cyclic hydrocarbons were widely used in the study 
of molecular assembly control and external stimulus-responsive materials, 
as their molecular assembly structures could be altered through guest 
additions or side-chain design. Compared to typical C2 and C4 symmetric 
molecules such as perylene and porphyrin, C3-symmetric molecules not 
only formed characteristic crystal structures, such as honeycomb structures 
but also attracted attention due to their ability to exhibit physical phenomena 
such as spin and charge frustration. Among them, dehydrobenzoannulene 
(DBA),1 which was the C3-symmetric polyaromatic hydrocarbon with a 
void surrounded by ethynyl linker, had unique properties such as redox 
properties, aromaticity, absorption–emission properties,2 and the ability to 
form complexes with transition metals. DBA formed a herringbone 
structure in single crystals due to C–H•••π interactions. On the other hand, 
introducing substituents into DBA led to the formation of various crystal 
structures, such as one-dimensional (1D) columnar structures,3 
two-dimensional (2D) layered structures,4 and three-dimensional tetragonal 
structures.5 These studies demonstrate that the interactions of substituents dominate the molecular aggregate structure of DBA, 
and the dynamic molecular assemblies based on DBA are expected to be applied as materials responsive to external stimuli.  
In this doctoral thesis, I focused on the C3-symmetrical DBA derivatives and studied the control of the molecular 
assemblies and the exploration of physical properties. In Chapters 2 and 3, I designed and prepared the DBA derivatives with 
          かさはら ようたろう 
氏 名          笠 原 遥 太 郎 
研究科，専攻の名称 東北大学大学院工学研究科（博士課程）応用化学専攻 
学 位 論 文 題 目           Studies on Synthesis, Control of Molecular Assembly Structures, and 
Physical Properties of C3-Symmetrical Dehydrobenzoannulene Derivatives 
(C3対称性デヒドロベンゾアヌレン誘導体の合成、分子集合体構造制御、および物性に関する研究) 
指 導 教 員 東北大学教授 芥川 智行 
論 文 審 査 委 員          主査 東北大学教授 芥川 智行 東北大学教授 蟹江 澄志 
             東北大学教授 三ツ石 方也 准教授 武田 貴志 
                              （信州大学） 
Figure1. The molecular structures of 
the DBA derivatives, which were the 
research subjects of the doctoral thesis. 
octylbenzoate group (C8[X]DBA, Figure 1, n = 0 ~ 2) to focus on the molecular assembly formed by isotropic van der Waals 
interaction and studied the formation of its molecular glass. In Chapter 4, I discussed the study of the polymorphism and 
physical properties exhibited by the alkylamide derivative 4-1a (Figure 1), which introduced anisotropic and strong 
intermolecular amide hydrogen bonds.  
【② Fluorescent molecular glass of the [12]DBA derivative6】 I prepared the C8[12]DBA (Figure 1, n = 1), which 
coexisted with dynamic alkyl chains and phenylene groups that suppress π–π interactions, and developed the luminescent 
molecular glass. Furthermore, from the single-crystal XRD of the methyl benzoate derivative C1[12]DBA, the mechanism of 
glass phase formation was discussed.  
To investigate the phase transition behavior and molecular assembly 
structure, I carried out differential scanning calorimetry (DSC), observation of 
polarized optical microscopy (POM), and variable-temperature X-ray 
diffraction (VT-PXRD). These results revealed that C8[12]DBA formed a 
glass (G) phase and exhibited a G– isotropic liquid (IL) phase transition 
around 320 K. This is the first report of a molecular glass material based on a 
DBA derivative. Variable frequency (f) and temperature (T) dependent 
dielectric constants indicated that the G phase of C8[12]DBA (250 – 320 K) 
exhibited the dielectric relaxation caused by the rotational fluctuations of 
alkylester chains. To investigate the mechanism of G formation in 
C8[12]DBA, I prepared the methyl ester derivative C1[12]DBA, and 
performed the single-crystal X-ray structural analysis. As a result, C1[12]DBA 
in the single-crystal formed dimers shifted by 180° relative to each other 
(Figure 2). When considering the extension of alkyl chains to this dimer 
structure, it was thought that the glass formation of C8[12]DBA was due to 
weak π-stacking interactions and bulky phenylene groups inhibiting regular and long-period molecular arrangements. 
Generally, solids formed by π-conjugated molecules exhibit strong intermolecular interactions, leading to excimer formation in 
the excited state, and their absorption-fluorescence spectra are red-shifted and quenched (ACQ) compared to solution states. 
However, the absorption-fluorescence spectra of C8[12]DBA in the glass state (Figure 3) closely matched those in the solution 
state. This result indicated that the intermolecular interactions between π-units in the glass state of C8[12]DBA were weak, and 
C8[12]DBA exhibited the emission like an isolated molecule without quenching.  
【③ Ring size effect on molecular assembly of C8[X]DBA】 To gain insights into the intermolecular interactions of 
C8[12]DBA that forms a glass, I investigated the influence of ring size at the molecular center on the molecular assembly 
Figure 3. Absorption and emission spectra 
of C8[12]DBA in CHCl3 soln. and glass. 
Figure 2. Dimer structure of  
C1[12]DBA in single crystal 
structure. Unlike typical cyclized hydrocarbons, DBA can control its ring size while maintaining molecular symmetry, 
depending on the number of ethynyl linkages bridging the benzene rings. In this chapter, I performed a comparative study on 
three DBA derivatives (C8[X]DBA, Figure 1) with the same molecular shape and substituents but different ring sizes. The 
larger-ring C8[18]DBA (Figure 1, n = 2) and the triphenylene derivative C8Trip (Figure 1, n = 0), which lacks bridging 
ethynyl linkages, were compared with C8[12]DBA. The effects of ring size on the molecular assembly structures and optical 
properties were examined from the molecular assembly structure and physical properties of C8[X]DBA, as well as from the 
single-crystal X-ray crystallographic analysis of the methyl benzoate derivatives C1[X]DBA.  
DSC, POM observation, and PXRD measurements of C8Trip revealed that it forms a crystalline phase with low 
crystallinity at room temperature and exhibited a crystal-isotropic liquid phase 
transition around 450 K (Table 1). Additionally, C8[18]DBA formed a stable 
crystalline phase with a hexagonal arrangement at room temperature (Table 1). 
These molecular assembly structures differed from those of C8[12]DBA, 
despite similar terminal substituents and molecular shape. The X-ray 
crystallographic analysis of C1[X]DBA, which can form single crystals, was 
used to discuss the intermolecular interactions of C8[X]DBA. C1[X]DBA 
formed dimers with the DBA core arranged 180° offset from each other, 
which were aligned along the b-axis direction (Figure 4). Moreover, the three 
derivatives had different steric hindrances at the terminal phenylene units, which altered the π-stacking interactions of intra or 
inter-dimers. As a result, the ring size effect on the molecular assembly structure was observed. Finally, the correlation between 
the molecular assembly structure and optical properties was examined. The absorption-emission spectra and quantum yield of 
C8[X]DBA in chloroform solution were different, which was due to the different electronic states of the π-framework. 
Compared to the quantum yield in the aggregated state, the glass state of C8[12]DBA did not exhibit quenching, while 
C8[18]DBA and C8Trip, which form crystalline phases, showed significant ACQ. These results indicated that the quenching 
behavior was due to the promotion of non-radiative deactivation through π-π interactions, which was facilitated by the 
increased regularity of the molecular assembly structure. 
Table1. Molecular assemblies, phase transitions, and phase transition enthalpy changes of C8[X]DBA 
Compounds Molecular assemblies T, K ΔH, kJ mol−1 Quantum yields (CHCl3 soln. → 
solid state) 
C8Trip Crystal (low crystallinity) 450, Cr → IL −23.8 
(+16.7) 
11 % → 5.7 % 
C8[12]DBA Glass 320, G → IL − [a] 13 % → 10 % 
C8[18]DBA Crystal (hexagonal 
arrangement) 
440, 
Polymerization 
− [b] 29 % → 1.9 % 
[a] Glass transition. [b] Polymerization. 
Figure 4.  Molecular arrangements of 
C8[X]DBA. 
【④ Polymorphism caused by dynamic hydrogen-bonds】 In Chapters 2 and 3, it was difficult to explore the 
function of DBA based on the C3-symmetry due to the random molecular arrangement. In this chapter, to address this 
challenge, the DBA derivative (4– 1a) with three tetradecylamide groups was designed to introduce anisotropic and strong 
intermolecular hydrogen bonding (Figure 1).  
4-1a formed lamellar (LM) structure from the chloroform (CHCl3) / acetonitrile (MeCN) mixed solvent and metastable 
one-dimensional nanofiber (NF) structure from toluene caused by different association behaviors in solution due to 
intermolecular hydrogen-bonds between the amide groups (Figure 5). In the LM structure, two hydrogen bonds were formed, 
whereas three hydrogen bonds were formed in the NF structure. These results revealed that polymorphism was due to 
differences in the hydrogen bonding patterns in the C3-symmetric molecule. DSC, VT-PXRD, and dielectric constant 
measurements showed the NF exhibited the phase transition to the LM phase around 440 K with −2.2 kJ mol−1 of exothermic 
phase transition enthalpy change due to the thermal activation of alkylamide units. These results revealed that LM was a 
thermodynamically stable phase compared to the NF phase. Molecular dynamics (MD) calculations revealed two equilibrium 
structures with different numbers of hydrogen bonds, and the polymorphism was successfully reproduced. The LM 
equilibrium structure had higher crystal density than NF, which was considered a polymorphic phenomenon where the number 
of hydrogen bonds and packing density competed. The correlation between the molecular assembly structure of 4-1a and the 
absorption spectra showed that the absorption spectrum of NF showed a blue shift due to the formation of H-aggregates, 
compared to the chloroform solution.  
 
 
【⑤ General conclusion】 
In this doctoral research, molecular glass materials responsive to electric fields and ring sizes, as well as materials exhibiting 
polymorphism dependent on solvents, were created by utilizing the C3 symmetry of DBA and the weakness of π-stacking 
interactions. 
 
Figure 5. Molecular structure of 4-1a and the summary of polymorphism by 
solvent-dependence. 
[1] Staab, H. A.; et al. Tetrahedron Lett. 1966, 7, 751–757; [2] Douhal, A.; et al. Phys. Chem. Chem. Phys. 2018, 20, 
7415–7427; [3] Hisaki, I.; et al. Chem.  Eur. J. 2008, 14, 4178–4187; [4] Hisaki, I.; et al. J. Am. Chem. Soc. 2016, 138, 
6617–6628; [5] Claverie, J. P.; et al. Adv. Funct. Mater. 2024, 34, 2311964; [6] Kasahara, Y.; et al. Chem. Commun.  2021, 57, 
5374–5377. 
  
 


Studies on Synthesis, Control of Molecular Assembly Structures, and Physical Properties of C3-Symmetrical Dehydrobenzoannulene Derivatives

LESAGE, LOUIS

Thermodynamic and Kinetic Control of Liquid Metal Dealloying for the Design of Porous Metallic Powders

Not a member yet