AI for Engineering

Ultra-low-resolution (ULR) thermal array sensors (e.g., $8\times 8$) offer a privacy-preserving and cost-effective paradigm for intelligent edge perception. However, extracting robust spatial features is severely hindered by extreme spatial aliasing, background thermal interference, and environmental temperature drift. In this paper, we propose VMEMSTNet, an ultra-lightweight framework that tackles the ULR perception bottleneck through a thermodynamics-inspired cross-modal learning strategy. Rather than treating ULR recognition purely as a computer vision task, our approach integrates the underlying sensing physics. To bridge the resolution gap, we leverage a high-resolution visual modality strictly as offline cross-modal guidance. Specifically, a YOLOv8n-seg teacher provides spatial shape priors via Gaussian thermal diffusion—explicitly mimicking physical heat dissipation for spatial knowledge translation—while a MobileNetV2 teacher guides semantic probability distillation. Furthermore, a thermodynamics-inspired preprocessing pipeline fuses dynamic quantile-based normalization with absolute temperature difference ($\Delta T$) to maintain contrast consistency across environmental drifts. Using hand gesture recognition as a representative spatial classification task, experimental results demonstrate that VMEMSTNet achieves 95.7% accuracy. Remarkably, it outperforms heavier generic edge models while utilizing only 20.0 k parameters and 0.135 MFLOPs. By effectively decoupling spatial semantics from thermodynamic noise, this work establishes a robust, hardware-aware deployment pathway for ULR thermal perception on resource-constrained microcontrollers. Full article

Recent advances in Generative AI are creating new opportunities for engineering and geosciences activities. This study examined whether the initial production of Territorial and Environmental Diagnosis (TED) reports could be co-piloted by multi-agent systems. The study proposed a challenge between a classical engineering workflow and a multi-agent framework (MAF) combining specialized agents boosted by LLMs for data retrieval, Data Visualization, writing, statistical description, and publishing within a unified deterministic workflow. To evaluate the effectiveness of the MAF, a blind comparative assessment was conducted using sections of TED reports produced by internal experts for three French municipalities as reference case studies. The framework was tested across several key thematic sections, and MAF outputs were compared with the outputs generated by territorial experts. The evaluation focused on Data Relevance, Writing Quality, Data Visualization, Contextual Relevance, and Overall Rating. The results of our experiments indicate that MAF reduced, within tested conditions, the time required for reporting while maintaining a level of quality comparable to conventional reports. Evaluations confirmed that MAF performed well in themes relying on processing data, while human-produced reports retained an advantage in topics requiring deeper contextual interpretation. These findings support the potential of MAF as co-pilots for TED production in engineering workflows. Full article

High-dimensional engineering datasets often combine class imbalance, noisy structures, and limited ground truth, making unsupervised analysis difficult to evaluate reliably. This study quantifies how three properties—partition stability, minority class recovery, and topological interpretability—are traded off across clustering methods, using a capacity-matched 25-seed comparison on a TCGA-derived RNA expression dataset (10,095 samples, 19 cancer types, 13,634 genes). We compare K-means across cluster counts $k\in \{19,\dots ,400\}$, self-organizing maps (SOMs) across lattice sizes from 25 to 625 nodes, consensus K-means, a granularity-matched SOM-Super20 control, and four modern baselines (HDBSCAN, spectral clustering, Gaussian mixtures, and Leiden). At matched prototype budgets, K-means is both more reproducible and substantially better at recovering minority classes than SOMs: at 400 prototypes, K-means achieves pairwise NMI $0.819$ versus $0.621$ for the $20\times 20$ SOM and recovers the smallest cancers 6–$14\times$ more effectively (pancreas effective coverage $0.760$ vs. $0.054$).Crucially, the SOM does not close this gap even when given more prototypes ($0.07$ at 625 nodes), so, under matched capacity, minority recovery is better explained by representational capacity and centroid allocation freedom than by topology preservation. The recovery is not free: increasing k overfragments the partition and lowers the pairwise ARI stability ($0.643\to 0.419$ from $k=20$ to $k=400$), while the NMI remains robust ($\approx 0.82$). The hardest minority, pancreas, is recovered only by high-capacity K-means and by no other method evaluated, including SOMs at any size, consensus K-means, SOM-Super20, HDBSCAN, Gaussian mixtures, spectral clustering, and Leiden. The SOM’s distinct value is therefore not stability or recovery but the interpretable two-dimensional topological visualization that it uniquely provides, including a gradient-organized structure that is reproducible across seeds for kidney (weaker for uterus). No single method optimizes all three properties; the appropriate choice depends on whether a task prioritizes reproducibility, minority recovery, or visual interpretability. Because these conclusions follow from the shape of the data and the allocation behavior of the algorithms rather than from biological semantics, we expect them to transfer to high-dimensional imbalanced engineering data, such as those from fault clustering, condition monitoring, and anomaly detection. Full article

Precise station-keeping of construction barges during offshore operations remains a demanding control problem because the underlying dynamics are highly nonlinear and the disturbance environment is seldom known a priori. This work investigates how a learning-based controller can be embedded into the coordinated winch-control architecture of a specialized engineering vessel to deliver accurate positioning in shallow water. Vessels such as rock-dumping platforms and pipe-laying barges routinely rely on a spread of mooring lines to hold station, and the tensions on these lines are, in current industrial practice, still adjusted manually by the winch operator. The scheme proposed here replaces that manual loop with an adaptive neural feedback law synthesized through backstepping, allowing the unknown portions of the ship model and the exogenous environmental loads to be compensated online without requiring prior identification. The 3DOF control wrench produced by the feedback law is then mapped to the physical line tensions through a constrained allocation that respects the unilateral and breaking-load constraints of the spread. The closed-loop system is shown to be semi-globally uniformly ultimately bounded (SGUUB) in the Lyapunov sense, and its performance is benchmarked against a conventional PD regulator and a nominal model-based design through simulation of a full-scale rock installation barge. When the model-based baseline is given the nominal plant, it attains the cleanest tracking; the proposed neural law achieves comparable steady-state accuracy without requiring prior identification of the hydrodynamic coefficients. A model-free deep reinforcement learning (PPO) controller is additionally benchmarked under irregular (JONSWAP) seas; it attains bounded sub-metre station-keeping without any model knowledge, on par with the PD baseline but less precise than the model-based and adaptive-neural laws. Full article

Recent advances in artificial intelligence (AI) and computer vision technologies have enabled practical applications in industrial environments where safety, quality, and productivity are critical. To support both researchers and practitioners, this survey categorizes AI-based vision systems by their functional objectives rather than algorithmic classification. Specifically, representative implementations are organized across key application areas including safety monitoring, product quality inspection, assembly line support, and worker productivity enhancement. Most of the surveyed studies are in the manufacturing and construction sectors, where real-world deployments have demonstrated measurable improvements. Unlike many previous reviews, this survey focuses on image-centric applications, using visually interpretable outputs such as photographs, video frames, and real-world examples to illustrate the on-site usability of AI vision systems. It also organizes prior work by functional roles and practical deployment considerations, rather than algorithm-centric evaluations, to provide practitioners with actionable insights for industrial adoption. Full article

This pilot study presents a surrogate modeling framework for estimating carbon emissions for 35 data centers in the New York City metropolitan area. Using publicly available facility data (square footage, operator type, location), we calculated the annual CO₂e emissions based on standard industry assumptions. These calculated values, which represent modeled emissions rather than measured data, served as the target variable for surrogate model development. A Random Forest regression model was implemented. The model achieved strong performance in producing the calculated emissions with the test set with cross-validated performance (CV R² = 0.960 ± 0.022 and CV MAE = 2431 ± 739 MT CO₂e). Analysis indicated that data center size was the major predictor, accounting for 79.7% of the total feature importance, while location and operator type contributed 13.6% and 6.6%, respectively. As a localized, preliminary feasibility study, this case study demonstrates that surrogate modeling using only publicly available facility data can provide modeled carbon footprint estimates for infrastructure planning and grid decarbonization efforts. The reproducible methodology can be applied to other metropolitan regions, though generalizability requires further validation with larger datasets. Full article

Fast and accurate resistance prediction is critical in early-stage ship design. While Michell’s thin-ship theory provides rapid evaluations, its linear assumptions limit accuracy, particularly as hull forms deviate from ideal slenderness. This paper introduces a physics-preserving neural calibration method that improves Michell’s theory without replacing the underlying solver. We train a two-dimensional convolutional encoder–decoder, conditioned on Froude numbers via global FiLM modulation, to predict a bounded correction to the geometric effective-slope field. Because the solver remains unchanged, the learned correction acts as an interpretable spatial perturbation rather than a black-box resistance map. Evaluated under a strict leave-one-family-out (LOFO) protocol on a fleet of five slender hull families (DTMB, NPL-4A, Wide-Light Canoe, Wigley, and Delft 372), the neural calibration achieves a mean absolute percentage error (MAPE) of 0.0741. This represents a 24% improvement over a reproduced 2020 baseline and a 7.9% improvement over the uncorrected Michell solver. The 2020 baseline is the rigid boundary-layer and phase-deflection correction of an earlier study by the present group, re-evaluated here on the present hulls at their measured attitudes. Ablation studies show that much of this aggregate gain is captured by a bounded global slope offset, indicating that a spatially uniform displacement correction accounts for most of the improvement on slender hulls, while the spatially varying field mainly adds per-family headroom. Finally, we map the physical boundaries of this approach. Dedicated recovery campaigns on fuller forms (KCS and Series 60) show that the model regresses compared to baselines. This confirms that while the correction successfully refines the linear source distribution for slender hulls, it cannot synthesize missing physics, such as stagnation pressure, separated flow, or wave interference, for fuller or unrelated geometries. Full article

Configuring urban hydrological models, such as the Storm Water Management Model (SWMM), for operational use remains onerous for many modellers. We propose aiswmm, a SWMM-specialized agentic runtime, together with an Agentic SWMM workflow that embeds (Agent) Skills and Model Context Protocol (MCP) tools to automate QGIS preprocessing, SWMM configuration, execution, and postprocessing. We demonstrate this natural-language triggered workflow on the Tod Creek watershed (located on the Saanich Peninsula, British Columbia). We also validate the proposed Agentic SWMM workflow at three levels: (i) a QGIS-based watershed-pour-point detection that agrees with the commercial PCSWMM^® method to within 0.88% of the watershed perimeter (approximately 7.5 pixels in the digital elevation model); (ii) byte-identical SWMM output files (Secure Hash Algorithm 256-bit identical) between the command-line execution and the MCP paths across 60 paired simulations, and (iii) peak inflow at the watershed outlet matching to three significant digits between the manual SWMM interface and Agentic SWMM workflows. The results confirm that Agentic SWMM workflow can produce the same outputs with the manual SWMM interface, as they are designed to use the same computational engine. We also propose a verification-first contract and byte-level audit chain that record the inputs, parameters, and outputs of each run, thereby supporting the auditability and reproducibility. The aiswmm runtime, Skills, MCP servers, and byte-level audit chain are released as open source and remain compatible with mainstream agentic runtimes (Codex, Claude Code, Hermes, and OpenClaw) to support reproducible SWMM modelling driven by natural language. Full article

This paper presents a Sim2Real policy transfer framework for distributed control in cyber-physical production systems using State-Based Potential Games (SbPGs). While fuzzy inference systems (FISs) or other conventional control policies provide interpretable and stable control policies for manufacturing processes, their direct deployment in real systems is often affected by Sim2Real discrepancies caused by actuator imperfections, sensor uncertainty, and process variability. To address this limitation, we propose a hybrid control architecture in which an optimized rule-based conventional control policy (i.e., FIS used in a non-adaptive, expert-knowledge-driven manner) serves as a baseline controller and SbPG-based policy adaptation refines the control actions online, while keeping the distributed manner, and is proven to converge. To evaluate robustness during Sim2Real deployment, deterministic and stochastic noise injection mechanisms are introduced to emulate systematic actuator biases and random disturbances. The proposed framework is validated on a laboratory-scale distributed production system. Experimental results in both simulation and real-world environments demonstrate that the SbPG-based adaptation compensates for disturbances and maintains production objectives under actuator, sensor, and parameter uncertainties. Compared to standalone FIS control, the proposed approach consistently reduces overflow and power consumption while satisfying production demands under noisy operating conditions. Additional ablation studies further confirm the robustness of the policy transfer strategy and the effectiveness of global and local interpolation mechanisms in the SbPG learning. Full article

Industrial surface defect detection constitutes a fundamental component in automated quality inspection but remains challenging due to complex textures, diverse defect scales, and stringent real-time constraints. To address these issues, we present MWYOLO, an enhanced YOLO11-based detection framework tailored for accurate and efficient industrial inspection. First, a C3k2-Mamba Spatial Fusion Block (C3k2-MSFB) that integrates global contextual information with local structural cues via state-space modeling, enabling more discriminative representations of fine-grained texture variations. Second, a multi-scale wavelet attention (MWA) module is embedded into the backbone, leveraging wavelet-domain feature decomposition and dual attention to capture multi-frequency patterns, thereby improving sensitivity to fine-grained and subtle defect patterns. Third, an Inner-CIoU loss is developed to emphasize interior geometric alignment during bounding-box regression, offering more stable optimization for ambiguous or low-contrast targets. Extensive experiments conducted on three representative industrial datasets—NEU-DET, HRIPCB, and a self-constructed GSD dataset—demonstrate the effectiveness of MWYOLO. The model achieves mAP50 scores of 81.4%, 98.1%, and 67.7%, respectively, while maintaining a lightweight design with only 3.2M parameters and 7.3 GFLOPs. The results validate MWYOLO as a robust and computationally efficient solution, offering a favorable balance between accuracy, interpretability, and deployability for real-world industrial defect detection tasks. Full article

The motion prediction of semi-submersible platforms is of significant importance for improving operational efficiency, ensuring platform safety, and providing early warning information for potential risks. Traditional prediction methods, such as those based on hydrodynamic simulations combined with Kalman filters, often face limitations due to their reliance on precise hydrodynamic parameters, which are difficult to obtain in practice. More recently, data-driven approaches, particularly deep learning models like Long Short-Term Memory (LSTM) networks, have shown promise in predicting complex motions. However, these methods often treat the prediction process as a “black box,” leading to issues such as a lack of generalization ability, overfitting, and an inability to quantify the uncertainty of prediction results. To address these challenges, this paper proposes a novel motion prediction method for semi-submersible platforms based on a Bayesian neural network (BNN). The BNN incorporates Bayesian inference to effectively integrate prior knowledge and measured data, thereby quantifying uncertainties and improving prediction accuracy. The method is validated using field-measured motion data from a semi-submersible platform in the South China Sea. Compared with LSTM and feedforward neural network, the BNN demonstrates superior anti-noise performance and prediction accuracy, achieving an accuracy rate (R²) of up to 91.5%. Moreover, over 92% of the true values are captured within the 95% confidence interval of the prediction results. This study highlights the potential of BNNs for the real-time motion prediction of offshore platforms, providing valuable support for early warning systems and operational decision-making. Full article