Accurate prediction of the hydrodynamic coefficients of non-spherical particles in wall-confined flows is crucial for understanding particle–fluid interactions and reliable modelling of particle motion. Under strong wall confinement, the hydrodynamic coefficients exhibit a highly nonlinear dependence on the Reynolds number, wall distance and particle orientation – posing significant modelling challenges. In this study, we propose a multi-stage physics-informed machine-learning (MSPIML) framework for modelling the drag, lift and pitching torque coefficients of a wall-bounded prolate spheroid over the explored parameter space. In the first stage, a physics-informed mixture-of-experts (PIMoE) model predicts the drag coefficient by intelligently blending empirical correlations with a data-driven statistical expert. The resulting high-fidelity drag coefficient is then injected as an auxiliary input to a second-stage model, either a deep neural network (DNN) or an additional MoE, that predicts lift and pitching torque coefficients, thereby leveraging the strong physical coupling among the three coefficients. Trained on a comprehensive dataset of 720 direct numerical simulations covering wide ranges of Reynolds number, wall distance and particle orientation, the optimal PIMoE–DNN and PIMoE–MoE configurations achieve relative errors below 2.2 % for drag, 11.4 % for lift and 7.0 % for pitching torque while maintaining excellent generalisation across the entire parameter space. Moreover, the Shapley additive explanations analysis confirms that the MSPIML framework correctly captures the physical dependencies: dominant influence of Reynolds number and strong pitching torque dependence on the drag coefficient. The MSPIML framework provides an interpretable and efficient approach to the prediction of hydrodynamic coefficients and offers substantial potential for dynamic modelling of non-spherical particles in multiphase flows.

The FrameNet project is a large-scale frame-semantic database with a seemingly usage-based core: It draws on 200,000 annotated sentences from representative corpora and offers the most comprehensive description of semantic valency patterns in English to date. Nevertheless, its empirical validity is weakened by the lack of statistical information on the distribution of lexical units, frames and frame elements. Similarly, the characterisation of frame elements as core, core-unexpressed, peripheral or extra-thematic – intended to indicate their essentiality to a frame – is primarily motivated on theoretical grounds. This raises the question of whether these labels are consistent with actual language use. After exhaustively extracting frequency data from Python’s NLTK FrameNet Corpus for all attested combinations of verbs, frames and frame elements, hierarchical gradient boosting models were trained on information-theoretic measures and word embeddings to predict the coreness of frame elements. The models provide strong usage-based evidence for a general core versus non-core distinction but cast doubt on further subdivisions such as core versus core-unexpressed or peripheral versus extra-thematic. While further validation is necessary, this contribution offers the first statistical perspective on the current state of FrameNet and its compatibility with usage-based approaches.

Traditional Reynolds-averaged Navier–Stokes (RANS) closures, based on the Boussinesq eddy-viscosity hypothesis and calibrated on canonical flows, often yield inaccurate predictions of both mean flow and turbulence statistics. Here, we consider flow past a circular cylinder over a range of Reynolds numbers ($3900$–$100\,000$) and Mach numbers ($0$–$0.3$), encompassing incompressible and weakly compressible regimes, with the goal of improving predictions of mean velocity and Reynolds forces. To this end, we assemble a cross-validated dataset comprising hydrodynamic particle image velocimetry (PIV) in a towing tank, aerodynamic PIV in a wind tunnel and high-fidelity spectral element direct numerical simulation and large eddy simulation. Analysis of these data reveals a universal distribution of Reynolds stresses across the parameter space, which provides the foundation for a data-driven closure. We employ physics-informed neural networks (PINNs), trained with the unclosed RANS equations, to infer the velocity field and Reynolds-stress forcing from boundary information alone. The resulting closure, embedded in a forward PINN solver and the numerical solver OpenFOAM, significantly improves RANS predictions of both mean flow and turbulence statistics relative to conventional models.

Neural network observers (NNOs) are proposed for online estimation of fluid flows, addressing a key challenge in flow control: obtaining flow states online from a limited set of sparse and noisy sensor data. For this task, we propose a generalisation of the classical Luenberger observer. In the present framework, the estimation loop is composed of subsystems modelled as neural networks (NNs). By combining flow information from selected probes and a neural network surrogate model (NNSM) of the flow system, we train NNOs capable of fusing information to provide the best estimation of the states, that can in turn be fed back to a neural network controller (NNC). The NNO capabilities are demonstrated for three nonlinear dynamical systems. First, a variation of the Kuramoto–Sivashinsky (KS) equation with control inputs is studied, where variables are sparsely probed. We show that the NNO is able to track states even when probes are contaminated with random noise or with sensors at insufficient sample rates to match the control time step. Then, a confined cylinder flow is investigated, where velocity signals along the cylinder wake are estimated by using a small set of wall pressure sensors. In both the KS and cylinder problems, we show that the estimated states can be used to enable closed-loop control, taking advantage of stabilising NNCs. Finally, we present a legacy dataset of a turbulent boundary layer experiment, where convolutional NNs are employed to implement the models required for the estimation loop. We show that, by combining low-resolution noise-corrupted sensor data with an imperfect NNSM, it is possible to produce more accurate and robust estimates. Our approach presents better robustness to noise when compared with direct reconstructions via super-resolution NNs and predictions from graph NNs and Fourier neural operators.

Online counselling services have seen increased use in recent years, providing critical emergency mental health support. These interactions are typically long, complex, and varied in the dialogue between help seekers and counsellors. The lack of domain-specific models, especially in low-resource languages, poses a significant challenge for the automatic detection of suicide risk in online chat services for mental health support. To address this challenge, our approach adapts a general-purpose large language model (LLM) to the suicide prediction task that employs a two-stage classification architecture to deal with sparse and imbalanced data. It extends the state of the art by: (1) incorporating psychological theory into model training and (2) capturing key aspects of conversation structure in counselling sessions. We evaluate the performance of the proposed LLM against the state-of-the-art LLMs for suicide detection on thousands of conversations in the Hebrew language from a leading national online counselling service in Israel. Results show that the proposed LLM outperformed existing state-of-the-art approaches in detecting suicide risk, as measured by relevant literature metrics. Moreover, the LLM outperforms other approaches even in the early stages of a conversation, which is crucial for real-time detection in practice. We also discuss the ethical implications of combining LLMs in counselling services. The contributions of this work are (1) extending existing LLM architectures to incorporate domain-specific information; (2) evaluating LLM technologies in the context of socially relevant problems; and (3) introducing novel LLM tools for resource-constrained languages.

Humanoid robots, with their capacity for human-like interaction and autonomous decision-making, present novel legal challenges in accident scenarios. This Article argues that existing Chinese accident law cannot fully accommodate accidents involving humanoid robots because hybrid human-algorithm control, adaptive machine learning, and embodied human-like interaction together destabilize traditional assumptions about fault, causation, proof, and remedy. To address these shortcomings, the Article proposes a reconstructed liability system that (1) establishes robust evidentiary processes for determining fault, (2) implements “reasonable person” technical standards for humanoid robot behavior, (3) designates manufacturers as the “least cost avoiders,” and (4) considers behavioral correction and retribution for robots to address victims’ psychological needs. This approach aims to foster an “accidental utopia” where innovation and safety are harmonized.

Velocity measurement techniques, such as particle image velocimetry (PIV), face a trade-off between field of view, spatial resolution and sampling rate, so that small-scale vortices, shear layers and high-frequency turbulent motions are often under-resolved. Most physics-informed reconstructions use a velocity–pressure formulation, even though pressure is not measured in typical PIV experiments, so the Navier–Stokes constraints are only weakly enforced. We address this issue by formulating a vorticity–velocity physics-informed network (VVPINN), in which pressure is eliminated and incompressibility is enforced together with a vorticity transport equation, thereby directly constraining the velocity field and its derivatives. We then compare this formulation with a conventional velocity–pressure PINN (VPPINN) for spatio-temporal super-resolution of planar PIV data in three cases: a laminar multi-cylinder wake, a two-dimensional Taylor–Green vortex and an experimental two-cylinder wake. In the Taylor–Green vortex case, with identical architectures and training strategies, the VVPINN yields smaller velocity errors, reduces the $L_2$ errors in vorticity and shear by approximately $10\,\%$, and the pressure gradient errors by up to approximately $30\,\%$ at moderate super-resolution factors, and produces instantaneous fields with more physically plausible vorticity, shear and fine-scale pressure gradient patterns. Spectral analysis shows that the temporal energy spectrum is recovered accurately, whereas the wavenumber spectra, particularly beyond the Nyquist wavenumber, remain more difficult to match because the training data strongly constrain the time histories at sampled locations, but only indirectly inform the smallest spatial scales. Overall, the results indicate that vorticity-based constraints provide a more effective route to physics-consistent super-resolution of sub-sampled PIV data than the conventional velocity–pressure formulation.

This paper evaluates the performance of baseline and domain-augmented ChatGPT models for literature-based knowledge support in flood susceptibility mapping (FSM) using machine Learning approaches. To assess this, we designed five key questions related to FSM, with benchmark responses derived from our comprehensive review article (Pourzangbar et al., Journal of Flood Risk Management 18, e70042), which analyzed 100 studies on ML applications in FSM. The same questions were posed (i) to standard ChatGPT-4 and ChatGPT-4o models without additional contextual material, and (ii) to a domain-augmented GPT-4 configuration (Chat-FSM) equipped with retrieval access to the 100 reviewed articles. The comparison highlights that GPT-based models can reasonably reproduce frequently reported machine learning models and conditioning factors from the reviewed literature, but show weaker consistency in feature selection methods, often suggesting less relevant techniques. Among the models, ChatGPT-4o demonstrated the weakest alignment with benchmark data, while Chat-FSM demonstrated the highest agreement with the benchmark dataset across most evaluated questions. In terms of application-level efficiency, GPT models required substantially less time and computational effort compared to manual literature synthesis under the defined experimental setup. While ChatGPT-based systems can support literature-informed exploration in FSM, human expertise remains essential for critical reasoning, methodological design, and application to novel or context-specific scenarios.

We propose a novel machine-learning-based turbulence closure framework in which a tensor basis neural network (TBNN) is directly embedded into a Reynolds-averaged Navier–Stokes (RANS) formulation, eliminating reliance on traditional baseline turbulence models. The TBNN is trained to predict the Reynolds stress anisotropic tensor from local invariant inputs and geometry-informed features, including stream function and velocity potential. Its output is processed by a regression model that generates an optimised eddy viscosity field, which is then integrated into the RANS equations as a zero-equation turbulence closure. The framework is evaluated on three turbulent flows over complex geometries: a wavy-bottom channel, a smoothed step and a backward-facing step. Incorporating geometry-informed features significantly enhances model robustness, yielding numerically stable and convergent solutions across all cases. The predicted velocity fields and turbulence distributions closely match large eddy simulation (LES) data, confirming the accuracy of the proposed approach and demonstrating its ability to operate independently of conventional turbulence closures.

Variational data assimilation and machine-learning based super-resolution are two alternative approaches to state estimation in turbulent flows. The former is an optimisation problem featuring a time series of coarse observations, the latter usually requires a library of high-resolution ‘ground truth’ data. We show that the classic ‘4DVar’ data assimilation algorithm can be used to train neural networks for super-resolution in three-dimensional isotropic turbulence without the need for high-resolution reference data. To do this, we adapt a pseudo-spectral version of the fully differentiable JAX-CFD solver (Kochkov et al., Proc. Natl Acad. Sci. USA, vol. 118, issue 21, 2021, e2101784118) to three-dimensional flows and combine it with a convolutional neural network for super-resolution. As a result, we are able to include entire trajectories in our loss function, which is minimised with gradient-based optimisation to define the neural network weights. We show that the resulting neural networks outperform 4DVar for state estimation at initial time over a wide variety of metrics, though 4DVar leads to more robust predictions towards the end of its assimilation window. We also present a hybrid approach in which the trained neural network output is used to initialise 4DVar. The resulting performance is more than twice as accurate as other state estimation strategies for all times and performs well even beyond known limiting length scales, all without requiring access to high-resolution measurements at any point.

This study assesses whether a hybrid prediction–optimisation workflow can be used as an exploratory exercise for Brazilian federal budget allocation under severe data constraints. Using executed expenditure by budgetary function (2000–2023; N = 24), a multi-output XGBoost model is estimated to link spending profiles to GDP growth, inflation, and the Gini index; Bayesian optimisation (Tree-structured Parzen Estimator/Optuna) is then applied to search, within explicit bounds and penalties, for allocation vectors that maximise a stated objective function favouring higher growth and lower inflation and inequality. To mitigate data scarcity, the short series is augmented with 1048 synthetic observations generated through controlled noise injection, bootstrapped resampling and variational autoencoder reconstruction. Under randomised K-fold cross-validation on the augmented dataset, the model achieves mean R2 = 0.97 and mean MSE = 0.04, while diagnostics indicate larger errors at extreme values and a persistent training–validation gap. A secondary robustness check uses an anti-leakage design by applying cross-validation to the 24 real observations and generating synthetic data only within each training fold. This yields markedly weaker generalisation for GDP growth and inflation (overall mean MSE = 1.03; overall mean R2 = −0.45), with positive performance remaining only for the Gini index (R2 = 0.60). Under these conditions, the optimisation step identifies a scenario that satisfies the objective function on standardised outputs (GDP growth = 1.15; inflation = −0.04; Gini = −0.17). The results support the use of the workflow to compare scenarios under explicit assumptions, rather than to produce prescriptive budget guidance.

This article examines the use of neural networks in electromechanical sound art and music, where sound is materially enacted through physical means such as motors, solenoids, and physical resonators. It begins with a survey of documented works, outlining a range of current strategies and discussing how technical, material, and performative factors influence their design. Identifying natural language processing as underexplored in this domain, a practice-based case study, Seven Studies for Electric Motors, develops one such language-based approach. The project embeds a small language model for real-time sentence generation, extracts syntax structures, and translates these into patterns of motor-driven sound. Taken together, the survey and case study offer a picture of how machine learning has been integrated into electromechanical practices over the past decade and point to possible directions for further work.