[pt] O ensemble smoother with multiple data assimilation (ES-MDA) se tornou um poderoso estimador de parâmetros. A principal ideia do ES-MDA é assimilar os mesmos dados com a matriz de covariância dos erros dos dados inflada. Na implementação original do ES-MDA, os fatores de inflação e o número de assimilações são escolhidos a priori. O único requisito é que a soma dos inversos de tais fatores seja igual a um. Naturalmente, escolhendo-os iguais ao número de assimilações cumpre este requerimento. Contudo, estudos recentes mostraram uma relação entre a equação de atualização do ES-MDA com a solução para o problema inverso regularizado. Consequentemente, tais elementos agem como os parâmetros de regularização em cada assimilação. Assim, estudos propuseram técnicas para gerar tais fatores baseadas no princípio da discrepância. Embora estes estudos tenham propostos técnicas, um procedimento ótimo para gerar os fatores de inflação continua um problema em aberto. Mais ainda, tais estudos divergem em qual método de regularização é sufiente para produzir os melhores resultados para o ES-MDA. Portanto, nesta tese é abordado o problema de gerar os fatores de inflação para o ESMDA e suas influências na performance do método. Apresentamos uma análise numérica do impacto de tais fatores nos parâmetros principais do ES-MDA: o tamanho do conjunto, o número de assimilações e o vetor de atualização dos parâmetros. Com a conclusão desta análise, nós propomos uma nova técnica para gerar os fatores de inflação para o ES-MDA baseada em um método de regularização para algorítmos do tipo Levenberg-Marquardt. Investigando os resultados de um problema de inundação de um reservatório 2D, o novo método obtém melhor estimativa tanto para os parâmetros do modelo tanto quanto para os dados observados. / [en] The ensemble smoother with multiple data assimilation (ES-MDA) gained much attention as a powerful parameter estimation method. The main idea of the ES-MDA is to assimilate the same data multiple times with an inflated data error covariance matrix. In the original ES-MDA implementation, these inflation factors, such as the number of assimilations, are selected a priori. The only requirement is that the sum of the inflation factors inverses must be equal to one. Therefore, selecting them equal to the number of assimilations is a straightforward choice. Nevertheless, recent studies have shown a relationship between the ES-MDA update equation and the solution to a regularized inverse problem. Hence, the inflation factors play the role of the regularization parameter at each ES-MDA assimilation step. As a result, they have also suggested new procedures to generate these elements based on the discrepancy principle. Although several studies proposed efficient techniques to generate the ES-MDA inflation factors, an optimal procedure to generate them remains an open problem. Moreover, the studies diverge on which regularization scheme is sufficient to provide the best ES-MDA outcomes. Therefore, in this work, we address the problem of generating the ES-MDA inflation factors and their influence on the method s performance. We present a numerical analysis of the influence of such factors on the main parameters of the ES-MDA, such as the ensemble size, the number of assimilations, and the ES-MDA vector of model parameters update. With the conclusions presented in the aforementioned analysis, we propose a new procedure to generate ES-MDA inflation factors based on a regularizing scheme for Levenberg-Marquardt algorithms. It is shown through a synthetic two-dimensional waterflooding problem that the new method achieves better model parameters and data match compared to the other ES-MDA implementations available in the literature.

We build on the method introduced by Ringer, et al., applying it to an 1820 event that happened near South Sulawesi, Indonesia. We utilize other statistical models to aid our Metropolis-Hastings sampler, including a Gaussian process which informs the prior. We apply the method to multiple possible fault zones to determine which fault is the most likely source of the earthquake and tsunami. After collecting nearly 80,000 samples, we find that between the two most likely fault zones, the Walanae fault zone matches the anecdotal accounts much better than Flores. However, to support the anecdotal data, both samplers tend toward powerful earthquakes that may not be supported by the faults in question. This indicates that even further research is warranted. It may indicate that some other type of event took place, such as a multiple-fault rupture or landslide tsunami.

No description available.

No description available.

No description available.

<p>Deep learning has been increasingly adopted in a multitude of settings. Yet, its strong performance relies on processing data during inference that is in-distribution with its training data. Deep learning input data during deployment, however, is not guaranteed to be in-distribution with the model's training data and can often times be distorted, either intentionally (e.g., by an adversary) or unintentionally (e.g., by a sensor defect), leading to significant performance degradations. In this dissertation, we develop algorithms for a variety of applications to improve the performance of deep learning models in the presence of distorted data. We begin by first designing feature engineering methodologies to increase classification performance in noisy environments. Here, we demonstrate the efficacy of our proposed algorithms on two target detection tasks and show that our framework outperforms a variety of state-of-the-art baselines. Next, we develop mitigation algorithms to improve the performance of deep learning in the presence of adversarial attacks and nonlinear signal distortions. In this context, we demonstrate the effectiveness of our methods on a variety of wireless communications tasks including automatic modulation classification, power allocation in massive MIMO networks, and signal detection. Finally, we develop an uncertainty quantification framework, which produces distributive estimates, as opposed to point predictions, from deep learning models in order to characterize samples with uncertain predictions as well as samples that are out-of-distribution from the model's training data. Our uncertainty quantification framework is carried out on a hyperspectral image target detection task as well as on counter unmanned aircraft systems (cUAS) model. Ultimately, our proposed algorithms improve the performance of deep learning in several environments in which the data during inference has been distorted to be out-of-distribution from the training data. </p>

<p dir="ltr">Nuclear energy, renowned for its clean, carbon-free attributes and cost-effectiveness, stands as a pivotal pillar in the global quest for sustainable energy sources. Additionally, nuclear power, being a spatially high-concentrated industry, offers an unparalleled energy density compared to other sources of energy. Despite its numerous advantages, if a nuclear power plant (NPP) is not operated safely, it can lead to long-term shutdowns, radiation exposure to workers, radiation contamination of surrounding areas, or even a national-scale disaster, as witnessed in the Chernobyl incident of 1986. Therefore, ensuring the safe operation of nuclear reactors is considered the most important factor in their operation. Recognizing the intricate tradeoff between safety and economy, economic considerations are often sacrificed in favor of safety.</p><p dir="ltr">Given this context, it becomes crucial to develop technologies that ensure NPPs’ safety while optimizing their operational efficiency, thereby minimizing the sacrifice of economic benefits. In response to this critical need, scientists introduced the term “digital twin (DT)”, derived from the concept of product lifecycle management. As the first instance of the term, the DT model comprises the physical product, its digital representation, data flowing from the physical to the DT, and information flowing from the digital to the physical twin. In this regard, various nuclear stakeholders such as reactor designers, researchers, operators, and regulators in the nuclear sector, are pursuing the DT technologies which are expected to enable NPPs to be monitored and operated/controlled in an automated and reliable manner. DT is now being actively sought given its wide potential, including increased operational effectiveness, enhanced safety and reliability, uncertainty reduction, etc.</p><p dir="ltr">While a number of technical challenges must be overcome to successfully implement DT technology, this Ph.D. work limits its focus on one of the DT’s top challenges, i.e., model validation, which ensures that model predictions can be trusted for a given application, e.g., the domain envisaged for code usage. Model validation is also a key regulatory requirement in support of the various developmental stages starting from conceptual design to deployment, licensing, operation, and safety. To ensure a given model to be validated, the regulatory process requires the consolidation of two independent sources of knowledge, one from measurements collected from experimental conditions, and the other from code predictions that model the same experimental conditions.</p><p dir="ltr">and computational domains in an optimal manner, considering the characteristics of predictor and target responses. Successful model validation necessitates a complete data analytics pipeline, generally including data preprocessing, data analysis (model training), and result interpretation. Therefore, this Ph.D. work begins by revisiting fundamental concepts such as uncertainty classification, sensitivity analysis (SA), similarity/representativity metrics, and outlier rejection techniques, which serve as robust cornerstones of validation analysis.</p><p dir="ltr">The ultimate goal of this Ph.D. work is to develop an intelligent inference framework that infers/predicts given responses, adaptively handling various levels of data complexities, i.e., residual shape, nonlinearity, heteroscedasticity, etc. These Ph.D. studies are expected to significantly advance DT technology, enabling support for various levels of operational autonomy in both existing and first-of-a-kind reactor designs. This extends to critical aspects such as nuclear criticality safety, nuclear fuel depletion dynamics, spent nuclear fuel (SNF) analysis, and the introduction of new fuel designs, such as high burnup fuel and high-assay low-enriched uranium fuel (HALEU). These advancements are crucial in scenarios where constructing new experiments is costly, time-consuming, or infeasible with new reactor systems or high-consequence events like criticality accidents.</p>

Deep Neural Networks (DNNs) have shown promising results in classifying known Low-probability-of-intercept (LPI) radar signals in noisy environments. However, regular DNNs produce low-quality confidence and uncertainty estimates, making them unreliable, which inhibit deployment in real-world settings. Hence, the need for robust uncertainty estimation methods has grown, and two categories emerged, Bayesian approximation and ensemble learning. As autonomous LPI radar classification is deployed in safety-critical environments, this study compares Bayesian Neural Networks (BNNs) and Deep Ensembles (DEs) as uncertainty estimation methods. We synthetically generate a training and test data set, as well as a shifted data set where subtle changes are made to the signal parameters. The methods are evaluated on predictive performance, relevant confidence and uncertainty estimation metrics, and method-related metrics such as model size, training, and inference time. Our results show that our DE achieves slightly higher predictive performance than the BNN on both in-distribution and shifted data with an accuracy of 74% and 32%, respectively. Further, we show that both methods exhibit more cautiousness in their predictions compared to a regular DNN for in-distribution data, while the confidence quality significantly degrades on shifted data. Uncertainty in predictions is evaluated as predictive entropy, and we show that both methods exhibit higher uncertainty on shifted data. We also show that the signal-to-noise ratio affects uncertainty compared to a regular DNN. However, none of the methods exhibit uncertainty when making predictions on unseen signal modulation patterns, which is not a desirable behavior. Further, we conclude that the amount of available resources could influence the choice of the method since DEs are resource-heavy, requiring more memory than a regular DNN or BNN. On the other hand, the BNN requires a far longer training time. / Tidigare studier har visat att djupa neurala nätverk (DNN) kan klassificera signalmönster för en speciell typ av radar (LPI) som är skapad för att vara svår att identifiera och avlyssna. Traditionella neurala nätverk saknar dock ett naturligt sätt att skatta osäkerhet, vilket skadar deras pålitlighet och förhindrar att de används i säkerhetskritiska miljöer. Osäkerhetsskattning för djupinlärning har därför vuxit och på senare tid blivit ett stort område med två tydliga kategorier, Bayesiansk approximering och ensemblemetoder. LPI radarklassificering är av stort intresse för försvarsindustrin, och tekniken kommer med största sannolikhet att appliceras i säkerhetskritiska miljöer. I denna studie jämför vi Bayesianska neurala nätverk och djupa ensembler för LPI radarklassificering. Resultaten från studien pekar på att en djup ensemble uppnår högre träffsäkerhet än ett Bayesianskt neuralt nätverk och att båda metoderna uppvisar återhållsamhet i sina förutsägelser jämfört med ett traditionellt djupt neuralt nätverk. Vi skattar osäkerhet som entropi och visar att osäkerheten i metodernas slutledningar ökar både på höga brusnivåer och på data som är något förskjuten från den kända datadistributionen. Resultaten visar dock att metodernas osäkerhet inte ökar jämfört med ett vanligt nätverk när de får se tidigare osedda signal mönster. Vi visar också att val av metod kan influeras av tillgängliga resurser, eftersom djupa ensembler kräver mycket minne jämfört med ett traditionellt eller Bayesianskt neuralt nätverk.

In response to the planned increase in operational speeds and axle loads of passengertrains that may lead to resonance-induced excessive vibrations in railway bridges,recent studies examined the reliability of bridges concerning train running safety andpassenger comfort limit states. In this respect, valuable information regarding theimportance of input variables can be obtained by conducting Sensitivity Analysis (SA).For instance, the determination of unimportant variables (where they can be treated asconstant) reduces the computational time, which is usually very high for probabilisticsimulations. In some of the previous studies, only deterministic SA has beenperformed. This thesis follows a stochastic approach using Global Sensitivity Analysis(GSA) methods. The considered performance functions are vertical acceleration anddeflection of single track ballasted simply supported reinforced concrete bridges.To reduce the computational time, available semi-analytical solution of a planarbeam under the passage of a series of moving loads is employed. To simulatethe bridge behaviour realistically, simplified methods to account for rail irregularityamplification, train-bridge interactions, and axle load redistribution were adopted.The considered random variables are train modal properties, number of train coaches,bogie spacing, axle spacing and loads, bridge mass, flexural stiffness and damping,and rail amplification factor. The analyses were carried out for a selected set of bridgelengths [10-30]m and a range of train speeds [100-400] km/hr. The study findingsshow that, in both acceleration and displacement, the dynamic response of the bridgeis sensitive to randomness in bridge mass, moment of inertia, coach length, and axleloads. Furthermore, the rail amplification factor and Young’s modulus are primarilyimportant for acceleration and displacement, respectively.

No description available.