The paper presents a framework to automatically identify crack patterns and the related features in existing reinforced concrete (RC) bridges. The challenge of this work is to define a tool for detecting the focused defect and highlighting the number and the orientation of cracks, allowing for correct interpretation and driving further evaluations on the residual life of the structure. The study is framed within the increasing interest in monitoring the structural health of existing bridges through automated tools, able to support engineers in the phase of visual assessment and interpretation of structural defects. When dealing with periodic inspection of large bridge portfolios, the support provided by automated tools can be fundamental for planning further strategies aimed at ensuring the structural safety and preventing future disasters. Given a stack of photos of a bridge structural element, an image stitching procedure is proposed to produce a near-complete image of the entire element. On the latter, a pipeline of deep-learning (DL) algorithms is employed to automatically detect and identify cracks (as a combination of object detection and segmentation algorithms). Finally, the proposed tool extracts cracks for counting and defines their orientation (i.e., vertical, horizontal, diagonal), in order to provide near-complete information about the crack pattern for the structural element. A full description of the methodology and the proposed algorithms is reported throughout the manuscript, showing the main pros and cons and assessing the effectiveness of the tool on a real-life case study.

Physics-informed neural networks (PINNs) are a promising alternative for extracting additional time-averaged (mean) flow quantities from experimental data. In the case of particle image velocimetry (PIV), for example, the measured mean flow field is contaminated by noise, has a limited field of view, is restricted to a uniform grid, and does not provide the pressure field. To overcome these limitations, we present a methodology in which PINNs are first trained on a Reynolds-averaged Navier–Stokes (RANS) simulation such that it learns all states at every location in the domain. We then apply transfer learning, which updates the PINN using sub-sampled PIV data. The resulting predictions are in significantly better agreement with the full PIV dataset than PINNs, which are trained on experimental data only. This work builds on the recent literature by integrating a Spalart-Allmaras turbulence model and applying hard constraints to the no-slip wall boundary condition. We apply this methodology to a two-dimensional NACA 0012 airfoil inclined at an angle of attack, $ \alpha $ = 15°, for two Reynolds numbers of Re = 10,000 and 75,000. The proposed methodology is initially validated using large eddy simulation (LES) data and then demonstrated on experimental PIV data. Our transfer learning approach results in improved predictions and a reduction in training time when compared to using a random network initialization.

While Structural Health Monitoring (SHM) has been widely studied, its reliability in real-world applications remains challenged by pronounced operational variability, particularly when system behavior changes discretely between operating regimes. Such methods generally rely on baseline comparisons; however, under multiple operating regimes, the baseline becomes distributed across several distinct regions, each associated with a specific regime. This multi-baseline behavior complicates anomaly detectability, as variations induced by changing operating conditions may mask the subtle changes caused by structural degradation. The challenge is particularly pronounced for systems exhibiting regime-dependent behavior, where transitions between approximately stationary conditions occur frequently and are difficult to isolate. To address this, an efficient probabilistic multi-model framework is proposed, in which each operating regime is represented by a locally Vector Autoregressive (VAR) model. A Bayesian formulation is adopted to account for parameter uncertainty explicitly and to enable sequential updating, allowing the regime models to adapt as additional data become available—an important feature for early-stage SHM. New observations are evaluated against the ensemble of regime-specific VAR models using the marginal likelihood, enabling assessment of statistical consistency with the learned reference behavior. Persistently low consistency is interpreted as indicative of anomalies, which may reflect structural changes or evolving degradation. The proposed method is demonstrated using vibration data from gearboxes onboard a Crew Transfer Vessel operating under multiple regimes. Despite the limitations and uncertainties inherent to early-phase SHM, the framework successfully identifies deviations from the learned reference behavior within consistent operating conditions, demonstrating its potential for SHM under realistic, time-varying operation.

Recent advances in learning dynamical systems from data have shown significant promise. However, many existing methods assume access to the full state of the system—an assumption that is rarely satisfied in practice, where systems are typically monitored through a limited number of sensors, leading to partial observability. To address this challenge, we draw inspiration from the Mori–Zwanzig formalism, which provides a theoretical connection between hidden variables and memory terms. Motivated by this perspective, we introduce a constant-lag neural delay differential equations (NDDEs) framework, providing a continuous-time approach for learning non-Markovian dynamics directly from data. These memory effects are captured using a finite set of time delays, which are identified via the adjoint method. We validate the proposed approach on a range of datasets, including synthetic systems, chaotic dynamics, and experimental measurements, such as the Kuramoto–Sivashinsky equation and cavity-flow experiments. Results demonstrate that NDDEs compare favorably with existing approaches for partially observed systems, including long short-term memory (LSTM) networks and augmented neural ordinary differential equations (ANODEs). Overall, NDDEs offer a principled and data-efficient framework for modeling non-Markovian dynamics under partial observability. An open-source implementation accompanies this article.

Digital Twin Construction (DTC) is a data-centric mode of construction that leverages digital twin technologies to maintain a continuously updated representation of a project and to enable short cycle Plan Do Check Act (PDCA) planning and control with continuous feedback and improvement. Industrial implementations reported to date have been too narrow in scope and too few in number to provide comprehensive proof of feasibility and empirical evidence of impacts. We address these gaps using a laboratory setup that implements a complete DTC PDCA workflow for a precast residential project. The experimental DTC system stores project intent, monitors and captures current status, and supports human in the loop replanning or automated optimization across factory production, logistics, and erection on site for a 1:25 scale model building. Validation through numerous full construction project runs demonstrates consistent end-to-end operation and practical usability. The testbed itself provides a reference architecture for DTC systems and a platform for controlled and replicable experiments that provide comparable quantitative evidence on DTC impacts under varied levels of automation.

Space utilization is a key driver for yacht design. Larger hulls increase the cost of the yacht substantially, and so space is always at a premium. The designs should feel spacious and luxurious, making the best use of the available volume while also conforming to standard engineering constraints such as safety and efficiency. To maximize the useful space, genetic algorithms have been utilized to design the layout with the aim of supporting the concept design process. Although a number of approaches have been developed, integrating these tools within the design process requires a substantial effort in human–machine teaming. This is at least in part due to the layout optimization problem being trivial in a general sense of providing effective space utilization, but the general layouts produced do not resonate with the designers as they do not conform to a brand often developed over decades. Essentially, does the yacht feel like other designs by the same company to a human designer? This article explores these two aspects of integrating an AI tool into a traditional design studio, generating concepts that resonate with the industrial designers while conforming to engineering constraints. It is shown that this only requires two objectives to achieve, in contrast to other approaches that use a large number of constraints or objectives to push the layout into a specific part of the design space. The resulting AI layout engineer is integrated into the team as the designs are competitive with those developed by the design team.

The Bayesian approach offers a systematic framework for updating finite element (FE) models and quantifying the remaining uncertainty given measured data. However, an inappropriate formulation of the probabilistic model can compromise accuracy. This paper presents an improved hierarchical Bayesian method for FE model updating by formulating the likelihood function in a fully probabilistic manner and incorporating time-varying stiffness parameters. A key methodological novelty lies in latent variable treatment of unmeasured mode shapes within the Bayesian hierarchy, yielding the joint inference of structural parameters and modal quantities in a fully generative manner without explicit eigenvalue decomposition. Furthermore, the geometric nature of mode shapes is rigorously respected by constraining them to the unit hypersphere using a Bingham distribution. A Metropolis-within-Gibbs sampling algorithm is developed to approximate the posterior distribution, with QR and Cholesky decompositions ensuring computational efficiency and accuracy. Three case studies, including synthetic, lab, and field test data, validate the effectiveness of the proposed approach. The updated model can be used as a reference model for structural damage detection and condition assessment in structural health monitoring.

This work presents a nonlinear system identification framework for modeling the power extraction dynamics of wind turbines, including both freestream and waked conditions. The approach models turbine dynamics using data-driven power coefficient maps expressed as combinations of compact radial basis functions and polynomial bases, parameterized in terms of tip-speed ratio and upstream conditions. These surrogate models are embedded in a first-order dynamic system suitable for model-based control. Experimental validation is carried out in two wind tunnel configurations: a low-turbulence tandem setup and a high-turbulence wind farm scenario. In the tandem case, the identified model is integrated into an adapted $ K{\omega}^2 $ controller, resulting in improved tip-speed ratio tracking and power stability compared to steady-state models, while also outperforming the BEM-based model in the present low Reynolds number setting, where aerodynamic predictions are more uncertain. In the wind farm scenario, the model captures the statistical behavior of the turbines despite unresolved turbulence. The proposed method enables interpretable, adaptive control across a range of operating conditions without relying on black-box learning strategies.

Underwater three-dimensional (3D) scanning systems play a crucial role in marine archaeology, offshore engineering, and underwater robotics by capturing detailed representations of areas of interest. However, conventional underwater scanning methods are often time-consuming and inefficient, frequently collecting redundant data points that add little value to the overall representation. This study introduces an artificial intelligence (AI)-driven approach to underwater surface scanning that leverages machine learning techniques such as Bayesian optimisation and Gaussian Process regression to address these inefficiencies. A prototype 3D scanner, controlled by machine learning algorithms, was developed and tested in laboratory conditions that replicated the conditions of offshore deployment. Surfaces with different geometries, including flat, conical, and wavy shapes, were scanned to evaluate the performance of the proposed method against traditional approaches. The new scanning method autonomously selects the most informative measurement locations, reducing the number of scans required while exceeding the accuracy of conventional techniques. The results demonstrate that the proposed approach provides more precise surface representations for most geometries while significantly reducing scanning time. The approach not only reduces computational and storage requirements but also enables efficient data transmission in low-bandwidth scenarios, such as with underwater wireless communication systems. This research highlights the potential of machine learning-enhanced scanning systems to outperform traditional methods, offering a faster, more adaptable, and more accurate solution for underwater visualization in diverse scientific and engineering applications.

Accurate prediction of bridge crack evolution is essential for infrastructure safety assurance and maintenance optimization. This study develops an interpretable machine learning framework to predict the expansion of cracks on the main beam in small- and medium-span highway beam bridges and identify the underlying mechanisms of structural deterioration. A comprehensive database was constructed from inspection and monitoring records of over 100 bridges, featuring critical degradation indicators, including crack density (CD) and maximum crack width (MCW). Following data preprocessing and feature selection through correlation analysis, three machine learning algorithms, that is, support vector regression (SVR), random forest (RF), and extreme gradient boosting (XGBoost), were implemented and evaluated using statistical metrics (R2, RMSE, and MAE). The XGBoost model demonstrated superior predictive performance with R2 values of 0.9433 and 0.9413 for MCW and CD, respectively, reducing RMSE by up to 66.8% and MAE by up to 72% compared to alternative models. SHAP (SHapley Additive exPlanations) analysis revealed that four factors, namely, vehicle load (VL), annual average daily truck traffic (ADTT), bridge age (BA), and annual average daily traffic (ADT), collectively contributed 61.45 ± 2.35% to crack development, with VL (19.7%) being the most influential factor. These findings identify excessive traffic loading and aging as the dominant drivers of crack propagation in beam bridges, providing valuable insights for targeted maintenance strategies and bridge management.

We present an efficient neural-based approach to estimate the instantaneous flow field around an airfoil from limited surface pressure measurements. The model, denoted SNN-POD, relies on two independent shallow neural networks to predict the instantaneous flow over a wide range of angles of attack $ \left[10{}^{\circ},20{}^{\circ}\right] $. At all angles the global model correctly recovers the average characteristics of the flow from single-time sensor data, thus allowing combination with local, angle-dependent models. The method is applied to 2D URANS simulations of a thick airfoil at a Reynolds number of $ \mathit{\operatorname{Re}}=4.5\times {10}^6 $. The training set consists of snapshots obtained from a coarse sampling $ \left(1-2{}^{\circ}\right) $ of the angle of attack range. A variance-based criterion is used to determine the number and positions of sensors. Tests are carried out for unseen snapshots at angles of attack within the set (sampled angles) as well as outside the set (interpolated angles). The maximum MSE error of attack for sampled and interpolated angles is respectively $ 2.9\% $ and $ 6.6\% $. This makes it possible to develop adaptive strategies to improve the estimation if necessary.

Thin-walled truncated conical shells subjected to axial compression are extremely susceptible to buckling, with experimentally observed buckling loads often falling well below classical theoretical predictions. The ratio of the experimentally measured critical load to its theoretical counterpart is defined as the Knockdown Factor (KDF). Although design guidelines proposed by agencies such as NASA provide conservative estimates of KDFs to ensure safety, recent research has highlighted the need to revisit and refine these provisions due to their excessive conservatism. In this context, the present study compares robust machine learning (ML) models for predicting buckling loads, or equivalently KDFs, of truncated conical shells using Artificial Neural Network (ANN), Support Vector Regression (SVR), Random Forest Regression (RFR) and Histogram Gradient Boosting (HGB). These models are able to capture strong nonlinear and complex feature interactions which are inherent in buckling phenomena. A comprehensive database compiled from existing literature and complemented with a set of simulated data is employed for model training and testing. To lead a new direction in the line of data-driven KDF prediction, a novel hybrid ML framework integrating Gaussian Process Regression (GPR) with Extreme Gradient Boosting (XGB), referred to as (GPR + XGB), is proposed. Additionally, a sensitivity analysis is performed to identify the most influential features governing the KDF predictions of truncated conical shells. The proposed hybrid framework that leverages experimental data as well as simulated data to accurately predict buckling KDFs of truncated conical shells, achieve significantly improved accuracy over existing ML models and conservative design guidelines.

Ferry quays experience rapid deterioration due to their exposure to harsh maritime environments and operational ferry impacts. Vibration-based structural health monitoring offers a valuable approach to assessing structural integrity and enhancing the understanding of the structural implications of these impacts. However, practical limitations often restrict sensor placement at critical locations. Consequently, virtual sensing techniques become essential for establishing a Digital Twin and estimating the structural response. This study investigates the application of the Gaussian Process Latent Force Model (GPLFM) for virtual sensing through Bayesian inference on the Magerholm ferry quay, combining in-operation experimental data collected during a ferry impact with a detailed physics-based model. The proposed Physics-Encoded Machine Learning model integrates a reduced-order structural model with a data-driven GPLFM representing the unknown impact forces via their modal contributions. This formulation enables joint probabilistic inference of system states and unmeasured responses while accounting for modeling and measurement uncertainties. Results show that the GPLFM provides accurate posterior mean acceleration response estimates at most locations, even under simplifying modeling assumptions such as linear time-invariant behavior during the impact phase. Lower accuracy was observed at locations in the impact zone. A numerical study was conducted to explore an optimal real-world sensor placement strategy using a Backward Sequential Sensor Placement approach. Sensitivity analyses examined the influence of measurement noise, sensor types, incorrectly assumed damping ratios, and sampling frequencies. The results suggest that the GP latent forces can help accommodate modeling and measurement uncertainties, maintaining acceptable estimation accuracy across scenarios.

Simulating turbulent fluid flows is a computationally prohibitive task, as it requires the resolution of fine-scale structures and the capture of complex nonlinear interactions across multiple scales. Consequently, extensive research has focused on analysing turbulent flows from a data-driven perspective. However, due to the complex and chaotic nature of these systems, traditional models often become unstable. To overcome these limitations, we propose a purely stochastic approach that separately addresses the evolution of large-scale coherent structures and the closure of high-fidelity statistical data. To this end, the dynamics of the filtered data are learnt using an autoregressive model. This combines a variational-autoencoder (VAE) and Transformer architecture. The VAE projection is probabilistic, ensuring consistency between the model’s stochasticity and the flow’s statistical properties. The mean realisation of stochastically sampled trajectories from our model shows relative $ {L}_1 $ and $ {L}_2 $ distances of 6% and 10%, respectively. Moreover, our framework enables the construction of meaningful confidence intervals, achieving a prediction interval coverage probability of 80% with minimal interval width. To recover high-fidelity velocity fields from the filtered space, Gaussian Process (GP) regression is employed. This strategy has been tested in the context of a Kolmogorov flow exhibiting chaotic behavior. We compare the performance of our model with state-of-the-art probabilistic baselines, including a VAE and a diffusion model. We demonstrate that our Gaussian process-based closure outperforms these baselines in capturing first and second moment statistics in this particular test bed, providing robust and adaptive confidence intervals.

Construction safety inspections typically involve a human inspector identifying safety concerns on-site. With the rise of powerful vision language models (VLMs), researchers are exploring their use for tasks such as detecting safety rule violations from on-site images. However, there is a lack of open datasets to comprehensively evaluate and further fine-tune VLMs in construction safety inspection. Current applications of VLMs use small, supervised datasets, limiting their applicability in tasks they are not directly trained for. In this article, we propose the ConstructionSite 10 k, featuring 10,000 construction site images with annotations for three inter-connected tasks, including image captioning, safety rule violation visual question answering (VQA), and construction element visual grounding. Our subsequent evaluation of current state-of-the-art large pre-trained VLMs shows notable generalization abilities in zero-shot and few-shot settings, while additional training is needed to make them applicable to actual construction sites. This dataset allows researchers to train and evaluate their own VLMs with new architectures and techniques, providing a valuable benchmark for construction safety inspection.

Point clouds derived from UAV photogrammetry are a cost-effective alternative to LiDAR for infrastructure inspections, but they often include both structural and non-structural elements that complicate analysis. Traditional denoising filters remove outliers indiscriminately and frequently erode edges, making it difficult to preserve the curved tunnel lining while distinguishing bolts, access gates, or pipelines. In contrast, segmentation-based approaches leverage geometric context to explicitly separate lining surfaces from ancillary components, thereby enabling more accurate deformation analysis and structural assessment. To that end, this paper presents a novel approach for denoising image point clouds using a synthetic training dataset to address the scarcity of labeled public data for enhancing point cloud quality. Unlike other denoising approaches that rely on projections or assume points lie on a predefined surface shape, this segmentation-based denoising method retains only meaningful points in their original locations, allowing for more accurate analysis of deformation. Enhanced by synthetic training datasets, the application of the proposed denoising method to a road tunnel image point cloud and a subway tunnel terrestrial laser scanning point cloud demonstrates its potential to enhance point cloud quality in tunnels with diverse geometries and point cloud data resources, even when data are limited. The method achieves an 80% mean intersection over union for both the road tunnel and the subway tunnel from manual annotation. This enables an improvement in structural deformation analysis at the mm level.

Building energy management (BEM) tasks require processing and learning from a variety of time-series data. Existing solutions rely on bespoke task- and data-specific models to perform these tasks, limiting their broader applicability. Inspired by the transformative success of Large Language Models (LLMs), Time-Series Foundation Models (TSFMs), trained on diverse datasets, have the potential to change this. Were TSFMs to achieve a level of generalizability across tasks and contexts akin to LLMs, they could fundamentally address the scalability challenges pervasive in BEM. To understand where they stand today, we evaluate TSFMs across four dimensions: (1) generalizability in zero-shot univariate forecasting, (2) forecasting with covariates for thermal behavior modeling, (3) zero-shot representation learning for classification tasks, and (4) robustness to performance metrics and varying operational conditions. Our results reveal that TSFMs exhibit limited generalizability, performing only marginally better than statistical models on unseen datasets and modalities for univariate forecasting. Similarly, inclusion of covariates in TSFMs does not yield performance improvements, and their performance remains inferior to conventional models that utilize covariates. While TSFMs generate effective zero-shot representations for downstream classification tasks, they may remain inferior to statistical models in forecasting when statistical models perform test-time fitting. Moreover, TSFMs’ forecasting performance is sensitive to evaluation metrics, and they struggle in more complex building environments compared to statistical models. These findings underscore the need for targeted advancements in TSFM design, particularly their handling of covariates and incorporating context and temporal dynamics into prediction mechanisms, to develop more adaptable and scalable solutions for BEM.

As a direct consequence of liquid kerosene injection, aeroengine combustors may be categorized as non-premixed combustion systems, characterized by a swirl-stabilized and highly complex flow field. In addition to the flow of air through the fuel injector, there are a large number of other features through which the oxidizer can enter the heat release region. These can have an impact on local fuel–air mixing, inducing strong spatial and temporal variations in stoichiometry, thereby affecting emissions and combustion system performance. This article discusses a novel statistical methodology, based on principal component analysis (PCA) and K-means clustering, that aims to improve the understanding of fuel–air mixing in realistic aeroengine combustors. The method is applied in a post-processing step to data sampled from a large-eddy simulation, where every chamber inflow has been tagged with a unique passive scalar, which allows it to be traced across space and time. PCA is used to construct a low-dimensional, visually interpretable representation of a spatially localized fuel–air mixing process, while K-means clustering is employed to produce an unsupervised discretization of the flow field into regions of similar fuel–air mixing characteristics. The proposed methodology is computationally inexpensive, and the easily interpretable outputs can help the combustion engineer make better-informed decisions about combustor design.

A key factor in ensuring the accuracy of computer simulations that model physical systems is the proper calibration of their parameters based on real-world observations or experimental data. Bayesian methods provide a robust framework for quantifying and propagating the uncertainties that inevitably arise. Nevertheless, they produce predictions unable to represent the observed datapoints when paired with inexact models. Additionally, the quantified uncertainties of these overconfident models cannot be propagated to other Quantities of Interest (QoIs) reliably. A promising solution involves embedding a model inadequacy term in the inference parameters, allowing the quantified model form uncertainty to influence non-observed QoIs. In this work, we revisit this embedded formulation and analyze how different likelihood constructions affect the inference of model form uncertainty, particularly under the presence of prescribed measurement noise and unavoidable model discrepancies. Two additional likelihood formulations, the global moment-matching and relative global moment-matching likelihoods, are introduced to explore alternative ways of representing the residual distribution. The behavior of these likelihoods is examined alongside existing formulations to show how different treatments of measurement noise and discrepancies shape the inferred parameter posteriors, and thereby affect the uncertainty ultimately propagated to the QoIs. Particular attention is given to how the uncertainty associated with the model inadequacy term propagates to the QoIs for the posteriors obtained from different likelihood formulations, enabling a more comprehensive statistical analysis of the prediction’s reliability. Finally, the proposed approach is applied to estimate the uncertainty in the predicted heat flux from a transient thermal simulation using temperature observations.

Punching shear failure in slab-column connections is a brittle collapse mode that threatens the safety of flat reinforced concrete (RC) slabs. Conventional design provisions are generally conservative but exhibit inconsistencies across geometric and material variations. This study develops an eXtreme Gradient Boosting (XGBoost) model to predict the ultimate punching shear capacity of flat RC slabs, using a database of experimental results categorized by four different geometric domains, including square slab with square column, circular slab with circular column, square slab with circular column, and circular slab with square column, covering the geometric, materials strength, and reinforcement properties of input parameters. The model achieved high predictive accuracy across the domains with coefficient of determination (R2) values > 0.930 in unseen testing datasets with minimal bias (0.994–1.006) and reduced scatter. Model interpretability, addressed through the SHapley Additive exPlanations analysis, confirmed slab thickness and average effective depth as the most critical predictors of shear capacity, followed by concrete strength and reinforcement parameters, while boundary condition parameters showed negligible influence due to the predominance of interior column cases. These findings demonstrate that XGBoost provides accurate, reliable, and interpretable predictions of punching shear capacity, offering a data-driven alternative to code-based methods and supporting safer and more consistent design of flat RC slabs.