DR-NTU (Digital Repository of NTU)
Not a member yet
116018 research outputs found
Sort by
Designing perovskite quantum dots for single photon emission in near-infrared range
No full textHalide perovskites are positioned at the forefront of photonics, optoelectronics and photovoltaics, owing to their excellent optical properties, with emission wavelengths ranging from blue to near infrared, and their ease in manufacturing. However, their vast composition space and the corresponding emission energies are still not fully mapped, and a guided high-throughput screening that allows for targeted material synthesis would be desirable. To this end, we use experimental data from the literature to build a Machine Learning model, predicting the band gap of 10920 possible compositions. Focusing on one of the most promising candidates, Cs2PbSnI6, we validate the model by synthesizing and characterizing nanocrystals of ordered 2-2 elpasolite (double perovskite) structure. The measured photoluminescence spectra agree with both ab initio GW band structure calculations and the machine learned predicted band gap. Therefore, our study not only provides a machine learned model for the composition space of the halide perovskites, but also introduces elpasolite Cs2PbSnI6 as a new candidate material for Near-IR emitting quantum photonic applications.Doctor of Philosoph
A conversation on space
No full textAs children, we found joy in the most unexpected places —under beds, within shelves, in the narrow gap between furniture and wall —not only to hide, but to imagine. These spaces which existed as a result of functionality for adults; became intimate worlds for children where imagination unfolded, solitude was embraced, and where wonder took root. A Conversation on Space is a return to these forgotten corners—not to dwell in nostalgia, but to question: could joy actually just be anywhere, waiting to be found?
This project explores the hidden spaces of childhood through sculpture and photography. Working closely with seven individuals, I gathered memories of the places they once played, hid, dreamed, and waited to be found. Some returned with fondness to spaces where they spent most of their time; others recalled entire imaginary worlds. Through revisiting, this work becomes a gentle act of restoration—an invitation to remember how it felt to dream, and a hope that we might learn to do so again.
In this paper, I will take you through the conceptualisation of the project—beginning with the personal and theoretical reflections that informed its development. I will explore themes of childhood perception, memory, and space, drawing from both artistic and academic research. I will also outline my process of working with participants, the choice of materials, and the making of the sculptures. Finally, I will reflect on the insights gained through this journey, and how this project has shaped my understanding of space, imagination, and the act of remembering.Bachelor's degre
Fabrication of defects in honeycomb specimens for ultrasonic C testing
No full textHoneycomb composite specimens are widely used in various industries due to it being lightweight and have desirable strength to weight ratio properties. During fabrication or throughout the structure’s life cycle, defects that are not visible to the naked eye may appear. This requires certain Non-Destructive Techniques such as Ultrasonic Testing to detect and characterise defects found within the honeycomb composite specimens.
The main objective was to explore the ability of single probe pulse echo method using C-scan (Auto-XY scanner) and Phased Array Ultrasonic Testing (Dolphicam2) equipment to detect defects from the defect and opposite side. Two different specimens were made of Carbon Fibre Reinforced Polymer face sheet with an aluminium or aramid honeycomb core. The fabricated defects were fibre delamination, lack of adhesion and cuts in honeycomb core.
Results showed that both experiments were unable to detect the cuts in the honeycomb core as well as any defects from the opposite side. However, the author was able to effectively detect and map out the fibre delamination and lack of adhesion defect from the defect side. The results show the ability of pulse echo technique, particularly C-scan and PAUT to detect and map out the size and depth of defects from defect side.Bachelor's degre
Out-of-distribution detection as a safety monitor in cyber-physical systems
No full textCyber-physical systems (CPS) are a class of systems that tightly integrate computation with the physical world. Advances in machine learning (ML) have enabled CPS to process complex data, be deployed in difficult and challenging environments, and perform more essential functions in an increasingly digital world. However, because CPS have tight integration with the physical world, safety is a key concern. Traditional verification and validation techniques are unable to guarantee safety properties of machine learning models due to the complexity of the input space and the black-box nature of these models. There are two aspects to safety in CPS: verification actions performed during the design phase and runtime monitoring and fallback during deployment. This work focuses on the latter.
When ML models are trained, they use data drawn from a particular distribution and their predictive power on other samples drawn from the same distribution is good. However, in the field, there is no guarantee that real world data will respect this training distribution. Out of distribution samples could result from a class of object unseen in the training data or from a covariate shift where the concept represented by the data remains unchanged, but some kind of noise is present to distort the data, e.g., different levels of precipitation or brightness. To detect such samples, out-of-distribution (OOD) detectors have been proposed in previous works. These works have focused on detecting OOD samples caused by novel classes, covariate shifts, or temporal disturbances. They have also considered the problem of explainability, i.e., providing a reason why a particular sample has been classified as OOD. However, previous works have not focused on the deployment of OOD detectors to resource-constrained CPS. In this work we focus on three aspects of the deployment of OOD detectors in CPS: model size, system provisioning, and exploiting explainable OOD results.
A major challenge is that OOD detectors are often implemented using deep neural networks (DNNs), which can take significant computational resources for execution. While existing techniques like neural architecture search exist to compress DNNs without sacrificing accuracy, OOD detectors pose some unique challenges. First, they are composed of additional preprocessing and postprocessing steps that are not present in other DNNs. The interplay between the performance of these steps and the effect of any changes to the DNN needs to be considered when compressing any architecture. Also, unlike a standalone DNN, the accuracy of OOD detection is not determined by the loss alone, so it is possible to prune and compress more aggressively without significantly influencing performance.
The next major challenge lies in system provisioning, i.e., when deploying an OOD detector, it is necessary to ensure that the processing time used for OOD detection does not detract from other safety critical tasks in a CPS. In general, larger DNNs can obtain more accurate results at the expense of more resources. On a system that is already maximally provisioned, introducing an OOD detector necessarily requires the use of smaller companion DNNs. We show that despite this effect, the OOD detector is still useful at ensuring system safety, provided the performance of all ML components is taken into account.
The last major challenge is the control action that should be taken when OOD samples are detected. Simple examples are handing control back to a human or bringing the system into a safe state (e.g., stopping along the side of a road), however, these options are not always available. Furthermore, not all OOD samples degrade ML performance to the point of uselessness. Building on previous works that cover explainable out-of-distribution detection, we show that OOD detectors can provide valuable information that can be used online to increase system performance and reduce risk.
Together, these three advancements -- compressing OOD detectors, co-design with other ML components, and development of runtime policies for OOD scenarios -- show that OOD detectors can be deployed and increase the safety of CPS. While OOD detection is not the only runtime safety monitor for CPS, it is a critical part of ensuring the safety of ML-enabled CPS. However, from a safety perspective, more research on runtime monitoring is needed. While OOD detectors help deal with unknown inputs where high uncertainty impacts a system, other types of monitors can observe environmental state or system state and provide valuable information. Taking advantage of this information to make a system safer and better performing, while still meeting real-time deadlines, remains uninvestigated. This work has contributed towards the safety of CPS, but it is only one piece of the safety puzzle.Doctor of Philosoph
Intelligent interface and data analytics of the model predictive controller of smart building in the cloudbased platform
No full textThe integration of predictive analytics and control in smart buildings has become
essential to reduce energy consumption and maintain occupant comfort. This study
presents the development of a cloud-based intelligent interface that utilizes stored
sensor data—rather than real-time integration—to simulate a Model Predictive Control
(MPC) system for optimizing building operations, specifically air conditioning and
mechanical ventilation (ACMV), lighting, and shading systems.
The proposed system utilizes the Light Gradient Boosting Machine (LightGBM)
model to forecast environmental parameters and energy needs. This model is chosen
for its exceptional ability to manage large-scale tabular data, offering enhanced
accuracy and faster training times compared to deep learning techniques like Long
Short-Term Memory (LSTM). The LightGBM model predicts essential variables such
as cooling power, room temperature, humidity. These forecasts are incorporated into a
Model Predictive Control (MPC) system that dynamically modifies the temperature
and humidity setpoints to reduce overall energy use, while still meeting comfort
requirements and indoor air quality standards. The entire control process is displayed
on a cloud-based dashboard, allowing users to interactively observe predicted
behaviors, energy patterns, and system reactions.
This study demonstrates the effectiveness of a data-driven, simulation-based control
approach using stored data, making it scalable for future deployment in buildings with
limited access to real-time sensor networks. The results validate the feasibility of
LightGBM-assisted MPC in supporting energy-efficient and adaptive smart building
management.Master's degre
Embodied carbon optimisation for building layout with hollow core slab
No full textCarbon emissions stand as one of the biggest issues today, with the construction industry being a key factor. Emissions are classified into two types: operational and embodied. Significant strides in the efficiency of buildings and renewable energy have mitigated operational carbon. Attention has increasingly moved towards lowering embodied carbon, which is associated with emissions from material production, transportation, and construction activities.
Despite growing awareness, systematic optimization studies focusing on reducing structural design embodied carbon are scant, especially considering the interactions of various components within a system at the design level. In changing this, this study presents different configurations of beam spans and hollow core slab (HCS) arrangements within a complete structure to see how they influence carbon emissions.
The research uses ETABS for structural analysis and building design. While the software’s application programming interface (API) is used for automating the iterative process, and the output data is managed and converted to embodied carbon emission in Microsoft Excel. By applying specific beam design criteria, curtailment rules, and optimal HCS thickness and column estimations, the credibility and applicability of the results can be conserved.
The result of this study demonstrates that strategic layout design can lead to significant embodied carbon savings. In general, building layout utilizing shorter HCS spans is more efficient in reducing embodied carbon emissions. Moreover, the optimal Y-span of each HCS slab is also provided in the outcome of this study. Another finding in this study is although higher live loads capacity and stronger concrete classes typically result in greater embodied carbon emissions, certain combinations of HCS span and load can offer similar emissions level. Hence, the flexibility of choosing different layout configurations
is available without compromising sustainability. The results offer practical guidelines for engineer to
make decision on building layout configurations, with minimizing embodied carbon emissions as a key consideration.Bachelor's degre
Unveiling observable, hybrid phases and dual phases in topological condensed matters
No full textThe topological electrical transport of condensed matter has been investigated fordozens of years. The microscopic mechanisms of the topological phases, such as quan-tum anomalous Hall insulators (QAHIs), quantum spin Hall insulators (QSHIs) andhigher-order topological insulators (HOTIs), have been established from the perspec-tive of band theory and fied theory. At the same time, the ”topology” is related toelectronic system’s other properties, such as the spin, the orbital, the correlation andthe symmetry, giving rise to more novel responses to the external fieds. Herein, fromQAHIs to the system with multiple topological phases, we study the inner mechanismsstep by step. In the firs section, through investigating the chiral transport of QAHIs, we fin that the topological invariant is coupled with the angular-momentum difer-ence between the electron and hole around the Fermi level, giving rise to the chiralmagnetic circular dichroism efect. In the second section, we uncover the hybrid topo-logical phases, the surface state and the step-edge state, in one elemental solid, which
is a novel phenomenon determined by the hybrid bulk topology of α-As’s electronic structure. In the third section, we investigate the interplay between the topology and electron’s correlation in a QSHI. By this work, we clarify the origin of the topology ofthe correlated gap by experiment and theory simultaneously. In summary, our findings uncover the microscopic mechanisms for the topological properties of condensed matters and the related efects of the crystals.Doctor of Philosoph
Very short-term chiller energy consumption prediction based on simplified heterogeneous graph convolutional network
No full textImplementing predictive control within heating, ventilation, and air conditioning (HVAC) systems is imperative for maintaining operational efficacy and concurrently attaining energy efficiency. In the context of HVAC systems, accurate prediction of chiller energy consumption at very short-term, minute-level intervals is crucial for the timely implementation of optimal predictive control strategies. Current research predominantly focuses on shortterm prediction at hourly or daily intervals, relying heavily on historical data for predictive insights. However, this approach imposes significant data acquisition burdens and may deviate from practical feasibility for very short-term prediction needs. This study introduces a novel approach employing the graph convolutional network (GCN) model for very short-term energy prediction based on an integrated simple graph (ISG-GCN) leveraging solely antecedent temporal information. The ISG addresses the challenges of determining connection weights between interdependent parameters, thereby eliminating the need for trial-and-error methods or reliance on potentially inaccurate historical correlation coefficients. Conceptualizing the chiller system as a black box, the model necessitates only seven discrete data points, significantly mitigating the data acquisition workload and facilitating seamless integration into buildings devoid of sophisticated sensor infrastructure. Applied to two simulation datasets encompassing large office buildings and one real operational dataset, the proposed model attains a mean absolute percentage error as minimal as 5.2%, demonstrating its effectiveness in real operational environments.National Research Foundation (NRF)AI SingaporeSubmitted/Accepted versionThis work was supported by the National Research Foundation Singapore under the AI Singapore Programme (AISG Award No. AISG2-TC-2023-008-SGKR)
Deep-ultraviolet laser generation for high harmonic generation
No full textThis Final Year Project (FYP) report discusses what I have learnt and done throughout the duration of my FYP. The concepts involved in nonlinear optics, including the basic processes like SFG and 2HG as well as the properties such as phase mismatch that can affect the efficiency of the process, and the ways in which it is mitigated.
The report also discusses the main objective of this FYP, to design a system for DUV generation through high harmonics generation, and the process of optimizing this system, some of the challenges faced and the conclusions made as well as a discussion on possible future work.Bachelor's degre
Towards practical automatic document understanding
No full textDocuments have been essential for information preserving and exchanging, with their significance continually increasing in the information age. Since the excessive documents make manual processing impractical, automatic document understanding has become a long-standing task with urgent and practical needs. This thesis focuses on practical research in automatic document understanding, where remaining four realworld challenges: few-shot, large-scale, multi-modal and long-context. We aim to address these challenges in the two sub-tasks of document understanding, i.e., Information Extraction and Document Reading Comprehension. Regarding Information Extraction, we introduce three IE algorithms which are effective under few-shot settings and efficient under large-scale scenarios. And in terms of Document Reading Comprehension, we focus on native multi-modal documents and explore (1) the end-to-end reading comprehension for long-context documents and (2) the efficiency of visualized document retriever towards large-scale applications. We believe these contributions in this thesis offer valuable insights and lay a solid foundation for future research towards more robust and scalable document understanding systems.Doctor of Philosoph