We present a critical survey on the consistency of uncertainty quantification used in deep learning and highlight partial uncertainty coverage and many inconsistencies. We then provide a comprehensive and statistically consistent framework for uncertainty quantification in deep learning that accounts for all major sources of uncertainty: input data, training and testing data, neural network weights, and machine-learning model imperfections, targeting regression problems. We systematically quantify each source by applying Bayes’ theorem and conditional probability densities and introduce a fast, practical implementation method. We demonstrate its effectiveness on a simple regression problem and a real-world application: predicting cloud autoconversion rates using a neural network trained on aircraft measurements from the Azores and guided by a two-moment bin model of the stochastic collection equation. In this application, uncertainty from the training and testing data dominates, followed by input data, neural network model, and weight variability. Finally, we highlight the practical advantages of this methodology, showing that explicitly modeling training data uncertainty improves robustness to new inputs that fall outside the training data, and enhances model reliability in real-world scenarios.

Sentiment analysis and stance detection are key tasks in text analysis, with applications ranging from understanding political opinions to tracking policy positions. Recent advances in large language models (LLMs) offer significant potential to enhance sentiment analysis techniques and to evolve them into the more nuanced task of detecting stances expressed toward specific subjects. In this study, we evaluate lexicon-based models, supervised models, and LLMs for stance detection using two corpuses of social media data—a large corpus of tweets posted by members of the U.S. Congress on Twitter and a smaller sample of tweets from general users—which both focus on opinions concerning presidential candidates during the 2020 election. We consider several fine-tuning strategies to improve performance—including cross-target tuning using an assumption of congressmembers’ stance based on party affiliation—and strategies for fine-tuning LLMs, including few shot and chain-of-thought prompting. Our findings demonstrate that: 1) LLMs can distinguish stance on a specific target even when multiple subjects are mentioned, 2) tuning leads to notable improvements over pretrained models, 3) cross-target tuning can provide a viable alternative to in-target tuning in some settings, and 4) complex prompting strategies lead to improvements over pretrained models but underperform tuning approaches.

Social scientists have quickly adopted large language models (LLMs) for their ability to annotate documents without supervised training, an ability known as zero-shot classification. However, due to their computational demands, cost, and often proprietary nature, these models are frequently at odds with open science standards. This article introduces the Political Domain Enhanced BERT-based Algorithm for Textual Entailment (DEBATE) language models: Foundation models for zero-shot, few-shot, and supervised classification of political documents. As zero-shot classifiers, the models are designed to be used for common, well-defined tasks, such as topic and opinion classification. When used in this context, the DEBATE models are not only as good as state-of-the-art LLMs at zero-shot classification, but are orders of magnitude more efficient and completely open source. We further demonstrate that the models are effective few-shot learners. With a simple random sample of 10–25 documents, they can outperform supervised classifiers trained on hundreds or thousands of documents and state-of-the-art generative models. Additionally, we release the PolNLI dataset used to train these models—a corpus of over 200,000 political documents with highly accurate labels across over 800 classification tasks.

This article examines the governance challenges of human genomic data sharing. The analysis builds upon the unique characteristics that distinguish genomic data from other forms of personal data, particularly its dual nature as both uniquely identifiable to individuals and inherently collective, reflecting familial and ethnic group characteristics. This duality informs a tripartite risk taxonomy: individual privacy violations, group-level harms, and bioterrorism threats. Examining regulatory frameworks in the European Union (EU) and China, the article demonstrates how current data protection mechanisms—primarily anonymisation and informed consent—prove inadequate for genomic data governance due to the impossibility of true anonymisation and the limitations of consent-based models in addressing the risks of such sharing. Drawing on the concept of “genomic contextualism,” the article proposes a nuanced framework that incorporates interest balancing, comprehensive data lifecycle management, and tailored technical safeguards. The objective is to protect individuals and underrepresented groups while maximising the scientific and clinical value of genomic data.

Recently, data-driven methods have shown great promise for discovering governing equations from simulation or experimental data. However, most existing approaches are limited to scalar equations, with few capable of identifying tensor relationships. In this work, we propose a general data-driven framework for identifying tensor equations, referred to as symbolic identification of tensor equations (SITE). The core idea of SITE – representing tensor equations using a host–plasmid structure – is inspired by the multidimensional gene expression programming approach. To improve the robustness of the evolutionary process, SITE adopts a genetic information retention strategy. Moreover, SITE introduces two key innovations beyond conventional evolutionary algorithms. First, it incorporates a dimensional homogeneity check to restrict the search space and eliminate physically invalid expressions. Second, it replaces traditional linear scaling with a tensor linear regression technique, greatly enhancing the efficiency of numerical coefficient optimization. We validate SITE using two benchmark scenarios, where it accurately recovers target equations from synthetic data, showing robustness to noise and flexible expressive capability. Furthermore, SITE is applied to identify constitutive relations directly from molecular simulation data, which are generated without reliance on macroscopic constitutive models. It adapts to both compressible and incompressible flow conditions and successfully identifies the corresponding macroscopic forms, highlighting its potential for data-driven discovery of tensor equation.

This paper aims to elucidate the physical mechanisms underlying airfoil–vortex gust interaction and mitigation. The vortex gust mitigation problem consists in finding the pitch rate sequence that minimises the gust-induced lift disturbance of an NACA0012 airfoil at Reynolds number 1000. The instantaneous flow fields and resulting lift are obtained from numerical resolution of the Navier–Stokes equations. The controller is modelled as an artificial neural network and trained to minimise the lift fluctuation using deep reinforcement learning (DRL). The paper shows that DRL-trained controllers are able to mitigate medium- and high-intensity vortex gusts by more than 80 % compared to the uncontrolled scenario. It then presents a comparative analysis of the controlled and uncontrolled lift generation mechanisms using the force partitioning method (FPM). The FPM provides a quantitative assessment of the amount of lift generated by each flow region. For medium-intensity gusts, the main phenomenon is the asymmetry in the airfoil boundary layer induced by the vortex. The control strategy mitigates the gust-induced lift by restoring the flow symmetry around the airfoil. For high-intensity gusts, the boundary layer asymmetry remains, but the gust interaction with the airfoil also triggers flow separation and the formation of a strong leading-edge vortex (LEV). Consequently, the control command balances several aerodynamic phenomena such as boundary layer asymmetry, flow detachment, LEV, and secondary recirculation regions to produce a net quasi-zero lift fluctuation. Thus this work highlights the potential of DRL control, enhanced by advanced post-processing such as FPM, to discover and interpret optimal flow control mechanisms.

In this work we propose a neural operator-based coloured-in-time forcing model to predict space–time characteristics of large-scale turbulent structures in channel flows. The resolvent-based method has emerged as a powerful tool to capture dominant dynamics and associated spatial structures of turbulent flows. However, the method faces the difficulty in modelling the coloured-in-time nonlinear forcing, which often leads to large predictive discrepancies in the frequency spectra of velocity fluctuations. Although the eddy viscosity has been introduced to enhance the resolvent-based method by partially accounting for the forcing colour, it is still not able to accurately capture the decay rate of the time-correlation function. Also, the uncertainty in the modelled eddy viscosity can significantly limit the predictive reliability of the method. In view of these difficulties, we propose using the neural operator based on the DeepONet architecture to model the stochastic forcing as a function of mean velocity and eddy viscosity. Specifically, the DeepONet-based model is constructed to map an arbitrary eddy-viscosity profile and corresponding mean velocity to stochastic forcing spectra based on the direct numerical simulation data at $Re_\tau =180$. Furthermore, the learned forcing model is integrated with the resolvent operator, which enables predicting the space–time flow statistics based on the eddy viscosity and mean velocity from the Reynolds-averaged Navier–Stokes (RANS) method. Our results show that the proposed forcing model can accurately predict the frequency spectra of velocity in channel flows at different characteristic scales. Moreover, the model remains robust across different RANS-provided eddy viscosities and generalises well to $Re_\tau =550$.

This study explores the marriage matching of only-child individuals and the related outcomes. Specifically, we analyze two aspects: First, we investigate the marriage patterns of only children, examining whether people choose mates in a positive or negative assortative manner regarding only-child status. We find that, along with being more likely to remain single, only children are more likely to marry another only child. Second, we measure the matching premium or penalty as the difference in partners’ socioeconomic status between only-child and non-only-child individuals, where socioeconomic status is approximated by years of schooling. Our estimates indicate that among women who marry an only-child husband, only children are penalized, as their partners’ educational attainment is 0.63 years lower. Finally, we discuss the potential sources of this penalty in light of our empirical findings.

The intelligible world of machines and predictive modelling is an omnipresent and almost inescapable phenomenon. It is an evolution where human intelligence is being supported, supplemented or superseded by artificial intelligence (AI). Decisions once made by humans are now made by machines, learning at a faster and more accurate rate through algorithmic calculations. Jurisprudent academia has undertaken to argue the proposition of AI and its role as a decision-making mechanism in Australian criminal jurisdictions. This paper explores this proposition through predictive modelling of 101 bail decisions made in three criminal courts in the State of New South Wales (NSW), Australia. Indicatively, the models’ statistical performance and accuracy, based on nine predictor variables, proved effective. The more accurate logistic regression model achieved 78% accuracy and a performance value of 0.845 (area under the curve; AUC), while the classifier model achieved 72.5% accuracy and a performance value of 0.702 (AUC). These results provide the groundwork for AI-generated bail decisions being piloted in the NSW jurisdiction and possibly others within Australia.

In this article, a 1 × 2 bandwidth (BW) and frequency-reconfigurable dielectric resonator-based multiple input multiple output (MIMO) antenna array is presented for 5G sub-6 GHz (3.3–6.0 GHz)/Wi-Fi 6E (5.925–6.425 GHz)/Wi-Fi 5G (5.15–5.85 GHz) applications. Additional dual-ring-open loop resonator structures with varied dimensions are introduced within antenna’s feeding network to achieve BW and frequency reconfigurability. RF PIN and varactor diodes (VDs) are integrated with proposed structure to enable switching between various modes and continuous tuning of frequency and BW, respectively. Further, Taguchi neural network (TNN) has been incorporated to predict percentage bandwidth of proposed antenna, getting a maximum deviation of only 0.6% from actual value. The proposed structure operated from 4.98 to 6.5 GHz, achieving wide continuous frequency tuning of 20.36% in passband and 6.1% reconfiguration for notch band. It also demonstrates continuous BW tunability from 16.69% to 34.44% with measured BWs of 19.58%, 34.44%, and 16.69% at 0, 3, and 8 V reverse bias voltages of VDs, respectively. MIMO antenna array structure also shows enhanced gain performance with a peak gain of 11.03 dBi and an overall gain above 7 dBi in the whole operating band.