Minerva Elements Records
Permanent URI for this collection
Search Results
1 - 10 of 17
-
ItemInformation Theory and Machine Learning: A Coding Approach( 2022-11)This thesis investigates the principles of using information theory to analyze and design machine learning algorithms. Despite recent successes, deep (machine) learning algorithms are still heuristic, vulnerable, and black-box. For example, it is still not clear why and how deep learning works so well, and it is observed that neural networks are very vulnerable to adversarial attacks. On the other hand, information theory is a well-established scientific study with a strong foundation in mathematical tools and theorems. Both machine learning and information theory are data orientated, and their inextricable connections motivate this thesis. Focusing on data compression and representation, we first present a novel, lightweight supervised dictionary learning framework for text classification. Our two-stage algorithm emphasizes the conceptual meaning of dictionary elements in addition to classification performance. A novel metric, information plane area rank (IPAR), is defined to quantify the information-theoretic performance. The classification accuracy of our algorithm is promising following extensive experiments conducted on six benchmark text datasets, where its classification performance is compared to multiple other state-of-the-art algorithms. The resulting dictionary elements (atoms) with conceptual meanings are displayed to provide insights into the decision processes of the learning system. Our algorithm achieves competitive results on certain datasets and with up to ten times fewer parameters. Motivated by the similarity between communication systems and adversarial learning, we secondly investigate a coding-theoretic approach to increase adversarial robustness. Specifically, we develop two novel defense methods (eECOC and NNEC) based on error-correcting code. The first method uses efficient error-correcting output codes (ECOCs), which encode the labels in a structured way to increase adversarial robustness. The second method is an encoding structure that increases the adversarial robustness of neural networks by encoding the latent features. Codes based on Fibonacci lattices and variational autoencoders are used in the encoding process. Both methods are validated on three benchmark datasets, MNIST, FashionMNIST, and CIFAR-10. An ablation study is conducted to compare the effectiveness of different encoding components. Several distance metrics and t-SNE visualization are used to give further insights into how these coding-theoretic methods increase adversarial robustness. Our work indicates the effectiveness of using information theory to analyze and design machine learning algorithms. The strong foundation of information theory provides opportunities for future research in data compression and adversarial robustness areas.
-
ItemDevelopment of efficient flood inundation modelling schemes using deep learning( 2022)Flood inundation models are one of the important tools used to manage flood-related risks in engineering practices such as infrastructure design, flooding disaster mitigation, and reservoir operations. Two-dimensional (2D) hydrodynamic models are commonly used in engineering applications because of their ability to provide robust estimates of flood inundation depth and extent at high temporal and spatial resolutions. However, due to the high computational costs, 2D models are not suited to many applications such as real-time ensemble flood inundation forecasting or uncertainty analysis. Therefore, many models have been developed based on simplified hydraulic rules such as considering only the conservation of water mass. These models are generally faster than 2D models but have reduced accuracy, which is a problem in many studies where a fine simulation timestep is needed or flow dynamics are significant. Recently, emulation models have been developed for fast flood inundation modelling using data-driven techniques including artificial neural networks, machine learning classification models, and deep learning. These computationally efficient emulation models are found to have comparable accuracy with 2D models when used to simulate flood inundation water level or depth provided with rainfall or streamflow discharge inputs. However, most emulation models simulate flood water/depth for each grid cell in the modelling domain separately, which would significantly increase the computational costs when applied for large domains. To add to that, these models have been found to have reduced accuracy in data-scarce regions on the floodplain. To improve the performance of emulation models, the objective of this thesis is to develop computationally efficient flood inundation models using deep learning and new spatial representation methods, that can be used for fast flood inundation simulation on floodplains with various characteristics at high spatial and temporal resolutions. The major contributions of this thesis include: (1) the development of an emulator for rapid flood inundation modelling which incorporates a novel spatial reduction and reconstruction (SRR) method as well as long short-term memory (LSTM) deep learning models to efficiently estimate flood inundation depth and extent; (2) the development of a Python program for the SRR method for flood surface representation; (3) the development of a U-Net-based spatial reduction and reconstruction (USRR) method and one-dimensional convolutional neural network (1D-CNN) models to emulate flood inundation on flat and complex floodplains. In addition, an input selection structure is developed and validated in the architecture of the LSTM models to simplify the model development process and to reduce the effort required for real-world applications. Also, a comparison is carried out for the performance of the combined approaches of the SRR method and LSTM models, as well as the USRR method and 1D-CNN models in an application to a flat and complex floodplain. The comparison demonstrates the advantages of using the USRR-1D-CNN emulator for rapid modelling of flood inundation on flat floodplains with complex flow paths, while the SRR-LSTM emulator is more computationally efficient and suitable for application to steep floodplains. The flood inundation modelling schemes developed in this thesis provide fast estimates of flood inundation surfaces without a material loss of accuracy compared to 2D hydrodynamic models, useful for applications such as ensemble real-time flood forecasting and flood risk analysis. They have the potential to deepen our understanding of the impacts of input uncertainty on temporal and spatial patterns of flood inundation, and to facilitate improved flood risk management.
-
ItemSHEAR BEHAVIOUR OF REINFORCED CONCRETE ELEMENTS: AN INSIGHT INTO SHEAR TRANSFER MECHANISMS IN CRACKED CONCRETE( 2022)Despite 70 years of investigations in understanding the shear behaviour of reinforced concrete members, it is again gaining attention among structural engineers as the recently issued Australian concrete design standard, AS 3600-2018, updated its shear provisions and ACI 318-19 unveiled its new one-way shear design equation. The shear behaviour of reinforced concrete elements is governed by several shear transfer mechanisms. Among them, the aggregate interlock is responsible for 50-70% of the ultimate shear transfer of cracked concrete elements. Despite its importance, a finite element model for shear transfer due to aggregate interlock considering realistic crack surfaces was still not developed. The complexity of developing a FE model lies due in the mesoscopic nature of the problem. In this study, a novel finite element approach is presented for evaluating shear transfer in crack concrete using realistic concrete crack surfaces. Concrete mesoscale models and zero-thickness cohesive elements were employed to develop the proposed method. Validation of the proposed FE models were conducted on two different experimental setups namely, small scale test and push-off test. The study comprises the evaluation of the surface roughness index of the cracked concrete surfaces. The proposed FE modeling approach demonstrated excellent performance against the most widely used analytical models and empirical equations for predicting the shear transfer in cracked concrete. Stress transfer in cracked concrete has been investigated since the 1970s, yet the existing code-based expressions for predicting shear transfer in cracked concrete were based on limited experimental data leading to insufficient prediction capabilities. Thus, this study further developed a machine learning-based framework for shear transfer in cracked concrete. The research outcomes present a novel finite element approach that is capable of evaluating stress transfer in cracked concrete and a machine learning-based framework to predict the maximum shear transfer in cracked reinforced concrete. The significance of the outcomes is that it enables to the evaluation of the stress transfer in crack concrete numerically while providing a clear pathway to solve the riddle of shear failures.
-
ItemData-driven large eddy simulation modelling in natural convection( 2022)Natural convection is a commonly occurring heat-transfer problem in many industrial flows and its prediction with conventional large eddy simulations (LES) at higher Rayleigh numbers using progressively coarser grids leads to increasingly inaccurate estimates of important performance indicators, such as Nusselt number (Nu). Thus, to improve the heat transfer predictions, we utilize Gene Expression Programming (GEP) to develop sub-grid scale (SGS) stress and SGS heat-flux models simultaneously for LES. With that as the focus, in the present study, two geometrically distinct natural convection cases are considered to develop and generalize turbulence models. The Rayleigh-Benard Convection (RBC) is used to develop models, while the Concentric Horizontal Annulus (CHA) is used to test the model generalization. An in-house compressible solver, HiPSTAR, for simulating natural convection flows for low Mach number problems is benchmarked against the experiments and Direct Numerical Simulations (DNS) results. Subsequently, HiPSTAR is used to run simulations for the RBC and CHA configurations and the generated DNS database is then used to train and assess LES models. The models’ development starts with RBC, where the fluid is in a cubic box with the bottom wall as the hot wall and the top wall as the cold wall. The alignment between different basis functions and the Gaussian-filtered SGS stress and SGS heat flux is used to determine the most suitable training framework. The trained models in isotropic form, by utilizing the norm of the grid cell as the length scales demonstrate good performance in the bulk region, but less improved performance in the near wall region. It is shown, that for LES of wall-bounded flow, the GEP models in anisotropic form, i.e. using different grid length scales for the different spatial directions, are required to obtain generalized models suitable for different regions. Consequently, the a-priori results demonstrate a significant improvement in the prediction of both instantaneous and mean quantities for a wide range of filter widths. However, developing accurate LES models that generalize well to complex geometries poses a challenge, particularly for data-driven methods. Thus, in the next stage, machine-learned closure models with embedded geometry independence are proposed, where the subgrid-scale (SGS) stress and heat-flux models developed by using Gene Expression Programming (GEP) are built in the computational space. The CHA case is chosen to develop and generalize the models. Subsequently, the formulation between the SGS closures, the total, and the resolved large-scale turbulent stress and heat flux is derived in the compressible LES context. The a-priori results show that the GEP models developed in computational space significantly improve both the SGS stress and SGS heat-flux prediction while being robust to complex flows. Similarly, the a-posteriori results demonstrate that the GEP models perform better than the wall-adapting local eddy-viscosity (WALE) model in the prediction of the mean SGS stress and the SGS heat-flux. The data-driven approach for turbulence model development presented clearly offers promising geometry independence for LES in the prediction of SGS stress and heat flux.
-
ItemMachine learning approaches to identify early keratoconus and classify keratoconus progression( 2021)Background: Keratoconus (KC) is a common corneal condition affecting children and young adults that has remained a major cause of corneal transplant surgery during the last decade. A newer management method named corneal collagen crosslinking (CXL) may now be used to slow the progression of KC in many individuals. Several studies have shown that the introduction of CXL resulted in a decrease in corneal transplantation rates in many countries. The CXL strategy is most beneficial to patients in early stage of KC to preserve best visual outcomes. The progression of KC must also be documented before executing a CXL procedure on KC patients. Identifying individuals with early KC and progressive KC is thus critical for planning early intervention care. Despite these requirements, existing approaches to early KC diagnosis and progression detection are inefficient. This includes a subjective assessment of early KC, as well as the absence of standardised criteria for KC progression that has resulted in the absence of a globally accepted KC management plan. The advancement of corneal topography and tomography imaging equipment has dramatically increased the quality and quantity of data acquired in KC clinics. There are currently no tools available in clinics that allow for the analysis of all corneal topography data to identify early KC and progression of KC, further signalling that their existing decision-making process may be inadequate. Purpose: This thesis sought to combine all 1692 parameters (containing demographic information and examination characteristics that are not clinical measurements.) available from an advanced corneal tomography system known as Pentacam into the decision-making process for improving early KC identification and progression classification. To cope with large amounts of data, this thesis used a variety of machine learning algorithms to accomplish the following aims: 1. To evaluate the performance of a number of various machine learning algorithms for discrimination of early keratoconus from eyes without KC (control eyes) using 11 commonly derived KC parameters; 2. To explore the impact of using all available Pentacam parameters with the highest-performing machine learning model for detection of early keratoconus; 3. To identify a key set of parameters supporting detection of early keratoconus as an optimal subset of all Pentacam parameters; and 4. To improve the classification of keratoconus progression by stratifying longitudinal clinical changes in KC using all Pentacam parameters with unsupervised machine learning algorithms. Research Design and Methods: This thesis is part of the Australian Study of Keratoconus (ASK), and was undertaken on a retrospective cohort of 3042 KC eyes and 700 eyes without KC (control eyes) collected at the Royal Victorian Eye and Ear Hospital. The diagnosis of each subject was retrieved from their electronic medical records. Early KC eyes were further labelled from the KC group by an expert optometrist. Early KC was defined as having a normal appearance on slit-lamp biomicroscopy and retinoscopy examination and abnormal corneal topography, such as inferior-superior localised steepening or an asymmetric bowtie pattern. The fellow eye may or may not be affected by KC. All 1692 accessible parameters on each individual eye were acquired on their first (baseline) visit and subsequent visits using the Pentacam v1.20r127 (Oculus, Wetzlar, Germany). Two clinical measurements, vision and refraction, were obtained from the patient's electronic medical record (where available) and included in the analysis for aim 3. From aims 1 to 3, supervised machine learning methods were used to develop models for distinguishing early KC and control eyes while in aim 4, an unsupervised machine learning method was used. Results: In Aim 1, the random forest algorithm was found to be the optimal machine learning algorithm to detect early KC from control eyes in the current dataset. This result was obtained by comparing the performance of eight different machine learning algorithms across 11 commonly derived KC parameters using 49 early KC eyes (49 patients) and 39 control eyes (39 patients). Notably, Aim 1 proved the usefulness of performing a machine learning algorithm comparison and feature selection to identify an optimal model that provided the maximal model performance. Aim 2 established the effectiveness of integrating all Pentacam parameters in identifying early KC. Using a reduced dimensionality space of all Pentacam parameters, the random forest model achieved a 98% accuracy (97% sensitivity and 98% specificity) in detecting early KC with 145 early KC (141 patients) and 122 controls eyes (85 patients). This machine learning model outperformed the majority of established machine learning modes in the literature. Further, in Aim 3, the same dataset as in Aim 2 was analysed, and a key combination of Pentacam parameters was identified for recognising early KC. This key set of parameters included the eccentricity value at a 30-degree angle of the front cornea, the eccentricity value in the 9 mm diameter zone of the cornea, and the inferior versus superior corneal asymmetry. The random forest model developed using these parameters correctly identified 94% of eyes and had a sensitivity of 97% and a specificity of 91% for distinguishing early KC from controls in the internal test dataset. Additionally, Aim 3 revealed the beneficial impact of combining vision and refraction measurements towards early KC detection. Finally, for the progression study of KC in Aim 4, three clusters/subgroups were identified with KC clinical change, defined as rapid-change, moderate-change and limited-change groups. The clusters were derived using hierarchical clustering based on half-year longitudinal clinical changes in all Pentacam parameters in 903 KC (588 patients) and 119 control eyes (92 patients). In Aim 4, 39 corneal curvature-related parameters were also found to be significantly different across the three subgroups. Discussion: By incorporating all Pentacam parameters in the machine learning model, the machine learning methods created in this work considerably improve the completeness of early KC detection. This enables assessment of the whole cornea during the early KC evaluation, possibly boosting the accuracy of current approaches for identifying early KC. These findings demonstrate the value of incorporating more pertinent information in order to improve early KC identification. Additionally, a key set of parameters for diagnosing early KC was identified, and this finding established the vital importance of corneal eccentricity in the early identification of KC, and the pertinent parameters could be used in clinics for early detection. These findings increase the generalizability of early KC diagnosis. There are numerous corneal topography systems available for usage in various clinics, but no machine learning model created in the literature so far can be applied across multiple imaging systems. The machine learning models developed in this thesis have a greater chance of being applicable to a variety of imaging systems due to the small number of parameters required and their widespread availability. For the first time, a three-classification scheme for KC clinical changes associated with progression was established using data-driven clustering. The finding may change our view of disease progression, since there were more categories than previously thought, in addition to progressed and non-progressed KC. This classification may alter the therapy protocol for individuals with KC since the care of patients with varying degrees of progression may differ. Finally, the research emphasises the critical need of monitoring the complete corneal curvature while assessing KC progression, rather than concentrating just on the central corneal curvature, as is presently done in clinics. Conclusion: The use of machine learning indicates its ability to better define early KC as well as begin to dissect out key aspects of KC progression. The machine learning-based potential decision support systems developed in this thesis addressed several gaps in the field and have the potential to improve clinical examination of KC patients and lead to better patient management.
-
ItemSatellite Remote Sensing of forest productivity and water status in an asynchronous climate and its application to the seasonal forecast of soil moisture( 2022)Soil moisture controls the vegetation water status, which is a prominent contributor to defining overall forest ecosystem health. Moreover, it has a crucial connection with forest growth, catchment water yield, bushfire risk, forest fuel moisture, forest carbon sink and streamflow, making real-time mapping and forecasting soil moisture status crucial. In Southeastern (SE) Australia, the plot scale short term feedback of vegetation to soil moisture deficit had been studied in detail; however, long term interaction at a large spatio-temporal scale remains largely unexplored. SE Australia represents an asynchronous climate where the vegetation growth drivers, i.e., rainfall and temperature, are seasonally out-of-phase. This asynchronicity suggests forests survive on deep soil moisture stores during the extensive dry season. However, these soil moisture stores are unknown, complicating the connection between past rainfall events, available moisture, and forest growth response in such climate settings. Consequently, there are limitations in our capability to map vegetation water status and forecast soil moisture. This study aimed to explore the potential of satellite observations in mapping real-time vegetation water stress and soil moisture forecast at a large spatio-temporal scale in SE Australia's forested landscape. Three different methods were developed to address the presented aim. The first method focused on the real-time retrieval of vegetation water stress. It was hypothesized that an improved water stress index could be constructed by the representation of canopy water content information to the Land Surface Temperature (LST) – Normalized Difference Vegetation Index (NDVI) trapezoid model. A new parameterizing method was developed to construct a temporally transferrable vegetation water stress index from the newly augmented spectral space. The performance of the new index was assessed by its capability to predict soil moisture from 60 study sampling windows and soil water fraction from four contrasting FLUXNET sites over the forested landscape in Victoria. Finally, the superiority of the index was presented by comparing its performance against the existing remotely sensed vegetation water stress indices. The second method focused on understanding the long-term relationship between rainfall and Eucalyptus growth. Being an asynchronous climate and forests have access to deep soil moisture stores, the study hypothesized that in SE Australia, there exists a sequential link between some distant (in time) rainfall event, soil moisture stores, and Eucalyptus growth response. It was assumed that NDVI represents yearly accumulated biomass in the Eucalyptus Forest, and it has a link with deep soil moisture dynamics. To address this hypothesis, lagged relationship between antecedent accumulated rainfall and observed NDVI from 2160 study sampling windows was analyzed. The result showed that Eucalyptus forests have a long memory of the previous rainfall, and this memory has a pattern across aridity. This pattern revealed that the middle aridity range forests have a relatively short-term memory of the past rainfall compared to the forests at two extreme ends of aridity, and that this memory pattern correlates with soil depth across the landscape. The third methodology has used this sequential connection between antecedent rainfall, soil moisture stores and forest growth response to forecast soil moisture. The focus centered on using the past rainfall condition, site-specific landscape attributes, and remotely sensed vegetation response in forecasting soil moisture three months in advance. An integrated system was constructed using these as input to the machine learning model trained and tested over 2160 study sampling windows. It was hypothesized that incorporating Moderate Resolution Imaging Spectroradiometer (MODIS) into the machine learning (ML) method will improve the skill for seasonal soil moisture forecast compared to the base model (ML model with all inputs excluding MODIS) and can probably reduce some of the base model's input parameters. A novel methodology was applied where the amount of rainfall driving Eucalyptus growth was back calculated using vegetations response in summer to wintertime rainfall. The result showed that in Mediterranean regions like SE Australia, antecedent rainfall and remotely sensed vegetation response has great potential in forecasting soil moisture three months in advance. Overall, results show that satellite observations have great potential in mapping real-time vegetation water stress and soil moisture forecast over southeastern Australia’s forested landscape.
-
ItemSoftware testing with Monte Carlo tree search( 2022)Achieving high code coverage in modern software is an essential yet challenging task. The size and complexity of programs give rise to the exploitation-exploration dilemma for testers: How to balance the consideration between investigating a) the parts of the program that appears to be interesting based on the existing information (exploitation) vs b) other parts that are less well-understood (exploration)? The Monte Carlo tree search (MCTS) is an artificial intelligence (AI) search algorithm that offers a principled solution to balancing these two concerns and has demonstrated its efficiency in large search spaces. This thesis contributes to tailoring MCTS to automate test generation for coverage-based testing. First, we propose Legion to assist coverage-based whitebox testing with our variation of MCTS. Legion’s APPFuzzing generalises concoilc testing and fuzzing to cover more program states with amortised computation cost: It selectively solves the path constraints of intermediate program states for seed inputs and mutates them to cover the descendants of the selected states. Legion’s variation of MCTS uses the best-first search strategy to learn the most productive program states to apply APPFuzzing, thus mitigating path explosion in symbolic execution. We compare Legion’s performance with KLEE, a famous coverage-based concoilc testing tool developed, maintained, and used by a large community across academia and industry. Legion outperformed KLEE in 7 out of 11 benchmark suites in a worldwide competition and achieved 90% of the best score in 5 suites. Then we extend Legion’s algorithm to seed selection in greybox fuzzing. This algorithm searches for stateful protocol servers’ most progressive seed input sequence. We analyse the impact of multiple server models and selection algorithms with repeated experiments. This study shows that although our algorithm has a limited impact on the overall coverage performance, it vastly increases the probability of covering some representative code blocks in servers. It highlights the bottleneck of overall coverage performance in input generation quality and fuzzing throughput. Finally, we design a contextual MCTS algorithm to reify and transfer the knowledge of past program state evaluations. Most software testing algorithms estimate a state’s productivity only with the state-specific statistics and neglect the knowledge from other states. To use the statistics more efficiently, we identify two contextual features of each program state and quantified their effect state’s productivity as feature coefficients. We then train a regression model to dynamically update the value of the coefficients to evaluate other program states based on the states’ feature values. By analysing our experiment results and reviewing others’ work, we point out the current weaknesses and corresponding future works regarding this topic.
-
ItemExploiting bacterial adaptation analyses to improve clinical management of Staphylococcus aureus infections( 2022)Despite advances in clinical management, invasive Staphylococcus aureus infections carry a high burden in terms of mortality and morbidity. The outcomes of these infections are likely to be impacted by bacterial adaptation to the host, including antibiotic resistance, persistence, and immune evasion. This project aimed at identifying signatures of bacterial adaptation by applying population genetics approaches to the study of clinical S. aureus infections. I used a combination of convergent evolution analysis (within-host evolution, adaptive laboratory evolution) and large-scale statistical genomics studies (genome-wide association studies, machine learning) to uncover bacterial genetic variation underlying adaptation during infection. Previous research on bacterial adaptation in clinical infections has focussed on methicillin-resistant S. aureus (MRSA) owing to the overall worse prognosis of these infections and the lower efficacy of anti-MRSA regimens. By contrast, less attention has been given to mechanisms of adaptation (and indirectly treatment failure) in methicillin-susceptible S. aureus (MSSA) infections, despite the latter accounting for up to 90% of invasive S. aureus infections. The first part of my thesis aimed at exploring bacterial adaptation during MSSA infections by characterising low-level resistance to flucloxacillin, the key antibiotic for the management of serious MSSA infections in Australia. By combining genomic tools (within-host evolution, population genetics, genome-wide association studies) and adaptive laboratory evolution I was able to challenge traditional paradigms of low-level flucloxacillin resistance. I showed that this type of resistance is adaptive, insofar that it is frequently selected under antibiotic pressure but has fitness cost and it is related to recurrent mutations in a group of key core genes. One of the key findings of the first part of my PhD was that bacterial adaptation is linked to “high-risk” mutations in a small group of core genes. This insight led me to explore S. aureus adaptation at a much larger scale, with the hope to identify robust signatures that could inform prediction of clinical outcomes, in particular treatment failure. To this purpose, I assembled an extensive collection of genomes from within-host evolution studies of S. aureus colonisation and infection (396 episodes; 2,590 genomes). I analysed these sequences using a bespoke within-host evolution genomic pipeline that allows to catalogue and annotate a broad range of genetic variation (including both point mutations and chromosome structural variants) and to infer statistical significance at gene, operon, and functional pathways level. A key finding of this analysis was that there are distinctive evolutionary patterns of S. aureus during colonisation, upon transition from colonisation to infection and during established infection. Crucially, beside the well-known agr-mediated adaptation, I was able to uncover several “non-canonical” signatures that were significantly enriched during within-host evolution. Ultimately if adaptation targets are to inform clinical management, they need to show predictive value in large cohorts of invasive infections. To assess this question, I applied a statistical genomics framework combining GWAS and machine learning to three cohorts of S. aureus bacteraemia (SAB) (1,358 episodes) across a spectrum of outcomes ranging from vancomycin minimum inhibitory concentration (MIC) to duration of bacteraemia to mortality. While low heritability and poor predictive performance indicated that bacterial genomics has a limited role in defining SAB outcomes, this approach was able to uncover previous undescribed genes associated with vancomycin resistance and new, potential genetic signatures of duration of bacteraemia. This work builds a bridge between bacterial genomics and clinical management of bacterial infections by highlighting the crucial role of bacterial (micro)-evolution in S. aureus infections and by providing key insights into adaptive phenotypes and their mutational landscapes. The tools developed for within-host evolution analysis and large statistical genomics are expected to assist future genomics investigations of invasive S. aureus infections.
-
ItemOptical microspectrometers and chemical classifiers based on silicon nanowires, plasmonic metasurfaces and machine learning( 2021)Spectrometers are a workhorse tool of optics, with applications ranging from scientific research to industrial process monitoring, remote sensing, and medical diagnostics. Although benchtop systems offer high performance and stability, alternative platforms offering reduced size, weight and cost could enable a host of new applications, e.g. in consumer personal electronics and field-deployable diagnostic platforms. To contribute to this trend towards miniaturised optical systems including spectrometers, this thesis presents the realisation of a visible spectrum microspectrometer using structurally coloured silicon nanowires and a reconstruction algorithm. We also experimentally demonstrate a plasmonic mid-infrared filter array-detector array microspectrometer that uses machine learning to determine the chemical compositions of a variety of liquids and solids. In this dissertation, we first present a reconstructive microspectrometer based on vertical silicon nanowire photodetectors. The nanowire photodetectors are designed to have absorption peaks across the visible spectrum. The spectral positions of these peaks are controlled by the radii of the nanowires. The nanowire detectors sit on substrate mesas that also serve as photodetectors for the light transmitted through the nanowires. We demonstrate the fabrication of this device, which has a footprint of a few millimetres. We use it as a spectrometer for the visible spectrum by implementing reconstructive algorithms. The identification of chemicals from their mid-infrared spectra has applications that include the industrial production of chemicals, food production, pharmaceutical manufacturing, and environmental monitoring. This is generally done using laboratory tools such as the Fourier transform infrared spectrometer. To address the need for fast and portable chemical sensing tools, we demonstrate the concept of a chemical classifier based on a filter array-detector array mid-infrared microspectrometer and a machine learning classification algorithm. Our device consists of a thermal camera onto which we have added an array of plasmonic filters. We perform simulations to find design parameters to enable the filters to have spectral features covering the wavelength range of interest. We first investigate this concept via a simulation study. We simulate the data that the device would generate when subjected to different chemicals, including noise. The simulated data is collated to train machine learning classification models. Our model predicts that this approach would be able to classify liquid and gas chemicals with very high accuracy. We later verify this concept by experimentally demonstrating a liquid chemical classifier. We design and fabricate a gold plasmonic filter chip containing 20 filters. The chip is integrated into a thermal camera to realise the mid-infrared microspectrometer platform. We train classifiers using the collected readout data of liquid analytes. The trained liquid classifier can accurately identify each type of analyte.
-
ItemEnzymatic Synthesis of Lactose-Based Bioactive Molecules( 2021)Functional food products and nutraceuticals have attracted considerable attention as they provide potential health benefits in humans. Global bioactive ingredient market data shows that functional food products have turned into a major business, driving a need for efficient and sustainable routes to produce defined compounds with specific health promoting properties. Bioactive lactose-based molecules, including human milk oligosaccharides (HMOs) and galactooligosaccharides (GOS), are examples of functional ingredients that can be supplements to food products. N-Acetyllactosamine (LacNAc) is a structural component of many biologically active compounds such as HMOs, glycoproteins, sialylated carbohydrates and poly-N-acetyllactosamine-type oligosaccharides. Currently, despite excellent research performed to study LacNAc-based building blocks, their commercial production is in an early stage. Commercially available starting substrates can facilitate industrial-scale biochemical production of these value-added ingredients. As an abundant and inexpensive substrate obtained from whey, lactose is an ideal starting substrate that can be used in both in vivo and in vitro biochemical syntheses. The aim of this thesis is to synthesize LacNAc-related molecules from whey-derived substrates, including lactose and casein glycomacropeptide (CGMP), using glycoside hydrolases. As a starting point, a comprehensive analysis of the importance of LacNAc and existing production strategies of both LacNAc and important LacNAc-based structures, including sialylated LacNAcs, as well as poly- and oligo-LacNAcs was compiled. A large-scale process was then designed for the enzymatic synthesis of LacNAc from lactose and N-acetylglucosamine (GlcNAc) based on the use of thermostable B-galactosidases from Bacillus circulans (BgaD-D), Thermus thermophilus HB27 (TtB-gly) or Pyrococcus furiosus (CelB). Downstream purification of LacNAc was simulated based on anion-exchange chromatography, an activated charcoal-Celite column, GlcNAc crystallization and an activated charcoal-Celite column, as well as selective crystallization. The effect of enzymatic yield, lactose concentration and acceptor to donor ratio on the project costs was discussed. The results showed that the process based on BgaD-D gave the best economics among the enzymes examined. In addition, the minimum LacNAc sales price can be reduced to $2 per gram by the use of selective crystallization as the most economically viable purification step. For most processes, GlcNAc mainly contributed to the raw material costs, while methanol contributed 72% of these costs for the process based on an activated charcoal column. The methanol consumption can, however, be reduced by 73% using a crystallizer for GlcNAc separation before the chromatography column. In the second part of this thesis, the transgalactosylation kinetics of the B-galactosidase from Bacillus circulans in the presence of cations present in dairy whey systems, namely calcium, magnesium, sodium and potassium, was investigated using both molecular modeling and quantitative experimental methods. This study indicated that hydrolysis and transgalactosylation reaction kinetics were not significantly affected at low concentrations of divalent cations (Ca2+ or Mg2+) or up to 100 mM of monovalent cations (Na+ or K+) compared to a control reaction. In contrast, high concentrations of calcium and magnesium (100 mM) triggered enzyme aggregation and progressive formation of an insoluble protein network resulting in the loss of enzyme activity. This decrease in the enzyme activity with time led to significant changes in the enzymatic yield and selectivity. The calculated electrostatic surface potential map of the enzyme was indicative of dominant negatively charged areas, which can interact with calcium and magnesium as strong salt-bridge forming cations. The docking position of these divalent cations were also predicted. This study presents a potential way to regulate the Bacillus circulans B-galactosidase reaction pathways by addition of divalent cations through the formation of protein aggregates. The next part of this thesis deals with the use of layer-by-layer deposition as a versatile technique for immobilizing the B-galactosidase in aqueous solutions under mild conditions. Commercially available silica particles were used as the base support in conjunction with robust pairs of polystyrene sulfonate (PSS) and polyallylamine hydrochloride (PAH) to encapsulate the B-galactosidase into the layers. The effect of multilayer films on the immobilized B-galactosidase stability and catalytic activity was studied in detail. Hydrolytic activity measurements indicated a significant decrease in the enzyme activity after immobilization. In addition, the higher the enzyme dosage applied during the immobilization process, the greater the activity reduction observed. Molecular analysis was further performed to study the possible interactions (electrostatic, covalent and protein-protein interactions) during this encapsulation method, that can contribute to the enzyme hydrolytic activity reduction. In contrast, the immobilized B-galactosidase was able to produce more LacNAc compared to the free counterpart. Moreover, the thermal and operational stability of the enzyme was substantially enhanced after the immobilization, allowing the successful recovery and reuse of the enzyme in consecutive cycles. The analysis of a large-scale process based on the immobilized B-galactosidase demonstrated significantly improved sustainability and economics. In the final part of this thesis, LacNAc directly produced and purified in cheese whey (as the acceptor) and whey-derived casein glycomacropeptide (CGMP) (as the donor) were used to synthesize 3’-sialyl-N-acetyllactosamine (3’-SLN) using a sialidase from the nonpathogenic Trypanosoma rangeli mutated with 15 amino acid substitutions (Tr15). The time-course study of the reaction was performed under different conditions in terms of acceptor to donor ratio, CGMP concentration and enzyme concentration. A high yield of 3’-SLN (75%) based on the available bound alpha-2-3-sialic acid in CGMP was obtained. This demonstrates the potential of this method for industrial applications as compared to sialyltransferase-catalyzed reactions requiring expensive nucleotide sugars. Furthermore, the reaction was successfully modeled using either a mechanistic kinetic analysis or a machine-learning-based approach, using an optimized artificial neural network. This reaction study showed the high trans-sialylation activity of Tr15 in the reaction with CGMP and LacNAc as sialyl donor and acceptor, respectively. In addition, as this enzyme is obtained from nonpathogenic species, it would be of more interest for pharmaceutical and food applications, compared to other sialidases that are virulence factors for pathogenic species.