Free-flight measurement for a small-scale model via synthetic image-based template matching technique

The paper investigates the interaction of saturated flows propagating along a slope and finally impacting a barrier (a) fixed to the base ground or (b) movable, i.e., free to slide along the base ground. Two volumes are considered, two positions of the barrier (near or far), and two soils (a monogranular medium sand and a volcanic silty sand). The tests are conducted at 40 g in a small-scale centrifuge model using the GeoCentrifuge, which is available at Deltares in Delft. Measurements are achieved through high-speed cameras and high-frequency load cells. Some major insights are that: (a) the presence of fine promotes long-lasting flow-barrier interaction, (b) the load transferred to the movable barrier is lower than the fixed barrier, (c) the ratio of lateral and vertical pressure inside the flow front drastically changes over time, (d) a statistic is achieved for both such ratio and the lateral soil pressure. Despite the substantial number of tests considered, the main limitations relate to: the specific geometric configuration of the slope considered; the limited number of load cells and cameras installed. Coupling the experimental results with numerical modeling represents a promising direction for future research.

ABSTRACT Universities face the dual challenge of preparing full-time students for real-world work-life challenges and offering practitioners lifelong learning (LLL) opportunities. The purpose of this article is to develop and assess a small-scale live-case model that focuses on both full-time and LLL students’ learning, reflecting on its implications for students, practitioners, and universities to better understand how to simultaneously promote lifelong learning and work-life integration in higher education. The model addresses two key challenges in previous live-case literature: high resource demands and the tendency to focus on organisational rather than individual learning. By focusing on a shared, small-scale live case involving quality improvement in healthcare, the small-scale live-case model offers a resource-efficient and time-bound format promoting mutual learning and engagement between engineering students and healthcare practitioners.

Active support systems are being increasingly applied in the control of large deformation in soft rock tunnels, and exploring the bearing characteristics of prestressed anchor bolts is of great engineering value for improving the long-term stability of tunnel structures. To address the problems of insufficient quantitative characterization of the bearing performance of prestressed anchor bolt support in soft rock tunnels and the difficulty of small-scale model tests in revealing the synergistic bearing law of support and surrounding rock, this study took a 350 km/h double-line high-speed railway tunnel as the prototype and established a large-scale tunnel structure model test system to conduct comparative tests under three working conditions: unsupported, ordinary bolt support, and prestressed anchor bolt support. By monitoring the tunnel failure process and mechanical response of the support structure throughout the test, the failure modes, bearing capacity, deformation characteristics, and axial force distribution of anchor bolts of tunnels under different support forms were systematically analyzed to quantitatively reveal the active support mechanism and bearing strengthening effect of prestressed anchor bolts. The results show that the design bearing capacity of the tunnel model with prestressed anchor bolt support is increased by 127.3% and 31.6% compared with that of the unsupported and ordinary bolt support models, and the ultimate bearing capacity is increased by 120.0% and 43.5%, respectively. Its secant stiffness in the initial loading stage reaches 80.0 kPa/mm, which is five times that of the ordinary bolt support and can effectively restrain the early plastic deformation of the surrounding rock. When the design bearing capacity is reached, the tensile stress of prestressed anchor bolts accounts for 40.2~69.8% of the ultimate tensile strength, with a more uniform axial force distribution and a much higher utilization rate of material mechanical properties than ordinary anchor bolts, which can fully mobilize the bearing potential of deep rock mass and realize the synergistic bearing of support and surrounding rock. This study accurately quantifies the bearing strengthening law of prestressed anchor bolts on tunnel support systems and clarifies the core mechanism of their active support. The research results provide important experimental basis and theoretical reference for the optimal design and engineering application of prestressed anchor bolts in soft rock tunnel engineering.

Computed tomography (CT) reports for patients with bone metastases include key clinical details, but they are usually written as free text. Different doctors write reports in different ways, which makes it hard to extract information in a consistent way. Large language models (LLMs) may help pull out important details automatically. But using external models can raise data privacy and management concerns in hospitals. This study looked at whether proprietary models and locally deployed open-source LLMs can help extract key diagnostic information and create structured reports in settings where data privacy is important. We analyzed 300 CT reports of metastatic bone tumors and evaluated seven large language models, including proprietary and open-source systems. Tasks included identification of pathological fractures and extraction of fracture and metastatic sites. Expert annotations served as the reference standard. Model performance was assessed using accuracy, recall, precision, F1 score, area under the curve (AUC), and Cohen’s Kappa. Across all three extraction tasks, accuracies ranged from 82% to 98%. Both proprietary and open-source models achieved high performance. Locally deployable models, including Qwen2-72B and WiNGPT2-9B, demonstrated performance levels within similar ranges to GPT-4o. The results show that open-source models can handle structured extraction of bone metastasis CT reports. Medically fine-tuned small-scale open-source models maintain stable performance while being more suitable for on-site deployment within healthcare institutions. This can help reduce data privacy risks. In the future, these models may help doctors make clinical decisions within health systems.

This study explores how pruning strategies can improve the efficiency of deep neural networks (DNNs), which are widely used for tasks like image processing, medical diagnosis, etc. Although DNNs are powerful, they often contain weaker connections that can lead to increased energy consumption both during training and inference. To address this, we compare two pruning approaches: global pruning, which applies to all layers of the network, and layer-wise pruning, which focuses on the hidden layers. These approaches are tested across two MLP models, small-scale and medium-scale, and are then extended to a VGG-16 model as a representative example of Convolutional Neural Networks (CNNs). We evaluate the impact of pruning on five datasets (MNIST, FashionMNIST, EMNIST, CIFAR-10, and OctMNIST), and considering different sparsity levels (50% and 80%). Our results show that, in comparison to the benchmark dense networks (0% sparsity), layer-wise pruning offers the best trade-offs, by consistently reducing inference time and inference energy usage while maintaining accuracy. For example, training the small-scale model with the MNIST dataset and 50% sparsity led to a 33% reduction in inference energy usage, 33% in inference time, and only a negligible 0.49% decrease in accuracy. Furthermore, we investigate training energy consumption, CO2 emissions estimations, and peak memory usage, which again leads to choosing the layer-wise approach over global pruning. Overall, our findings suggest that layer-wise pruning is a practical approach for designing energy-efficient neural networks, particularly in achieving efficient trade-offs between performance and energy consumption.

Large vision language models (VLMs) show strong open-world generalization but degrade at domain-specific tasks, while traditional small forensic models perform well on in-distribution datasets yet lack cross-distribution generalization and language-based interpretability. To address this gap, the authors propose a vision forgery trace (VFT)-VLM framework, which incorporates forensic features into a VLM without sacrificing its general reasoning ability. Specifically, a lightweight VFT extraction module learns to encode texture anomalies, edge incoherence, pixel artifacts, and frequency-domain deviations. The traces are incorporated into the InternVL2-8B backbone via low rank adaptation fine-tuning, achieving alignment between visual evidence and textual explanations. Across 14 diverse artificial intelligence-generated content benchmark datasets, VFT-VLM outperforms VLM-based large-scale models and achieves comparable or superior performance to relevant traditional small-scale models. Ablation studies confirm both VFT extraction and low rank adaptation fine-tuning are critical to the performance gains.

• AE monitors erosion-corrosion in pipelines with varied contaminants & flow. • Increased particle concentration & velocity boosts AE hits and energy. • AE signal intensity shows inverse correlation with increasing flow velocity. • Sand particles induce higher erosion rates than slurry at similar levels. Pipelines represent a fundamental conduit for the transportation of fluids, and their significance is often likened to the circulatory system of modern society. Thus, having the ability to forecast the likelihood of failure and damage location caused by the erosive-corrosive mechanism is paramount. This study investigates the implementation of the Acoustic Emission (AE) method in monitoring various laminar and turbulent seawater flow regimes with different sand and slurry contamination levels on a relatively small-scale pipeline model. The erosion-corrosion behavior on a carbon-steel pipeline was monitored via chemical analyses and AE techniques. In recent years, AE has emerged as a modern tool for detecting active defects, given its ability to save time and expenses. The study delves into the intricate interplay between AE and the erosive-corrosive mechanism in pipelines. Consequently, this research has experimentally discussed using continuous AE technique to enhance offshore pipeline design, inspection, and maintenance, by enabling the early detection of erosive-corrosive wear in pipeline systems.

This paper investigates the acoustic radiation produced by water entry of objects. While current research predominantly focuses on the fluid dynamic characteristics of object submersion, studies on acoustic radiation are limited. Most existing studies derive conclusions from open-water or laboratory measurements of sound pressure, which are highly susceptible to experimental conditions such as measurement distance, water depth, and medium. This study focuses on solid steel balls and uses total sound energy to characterize the acoustic properties of underwater objects. Different methods are employed in near-field and far-field areas to measure water entry sound, and an optimized segmented calculation method suitable for the total sound energy of water entry sounds is proposed. This method reduces the impact of echo signals from the initial impact sound on bubble pulsation sounds in near-field measurements and avoids the influence of the cutoff frequency of the reverberation tank in far-field measurements. Additionally, the calculation results are integrated with a neural network-based data fitting method to establish a small-scale predictive model for the total sound energy of water entry sound. The measurement-calculation-modeling approach proposed in this study provides a theoretical foundation for engineering applications such as extracting acoustic characteristics of underwater targets in shallow marine environments and acoustically locating water entry events.

Seg-LLaVA: A small-scale large vision-language model with external visual prompts

This paper proposes an integrated reinforcement learning (RL) framework for optimizing energy management in smart and micro grids that addresses both real-time operations and day-ahead market trading. By designing reward functions that incorporate real-time market prices, grid demand, peak penalties, and forecasted load values, the framework directs optimal charging, discharging, or holding actions of a Battery Energy Storage System (BESS). A comprehensive battery model captures state-of-charge (SoC) dynamics with round-trip efficiency losses, cycle-based degradation using rainflow counting algorithms, and operational constraints including ramp rate limits. This physics-based degradation modeling, which accounts for nonlinear depth-of-discharge effects and electrochemical aging mechanisms (SEI growth, lithium plating, electrode stress), enables the RL agent to balance immediate energy arbitrage profits against long-term asset preservation through optimized shallow cycling strategies. The framework employs Proximal Policy Optimization (PPO) for stable multi-objective policy learning and integrates day-ahead load forecasting using Transformer models. A novel contribution is the application of vertical agents powered by small-scale large language models (sLLM) to translate RL decisions into executable schedules through an intuitive human-machine interface, bridging the gap between optimal policies and practical implementation.

Large language models (LLMs) have begun to function as assistants or teammates in language learning, teaching, and research. However, what prerequisites are required for LLMs to reliably play these roles, and how such prerequisites should be measured, remains under-discussed. This study focuses on measuring Pedagogical Grammar Pattern Recognition (P-GPR) and establishes the Chinese Pedagogical Grammar Evaluation (CPG-EVAL), a multi-tiered benchmark designed to evaluate P-GPR within International Chinese Language Education. CPG-EVAL operationalizes grammar–instance correspondence through five task types that progressively increase contextual load and interference. We evaluate multiple proprietary and open-source LLMs as well as human participants. Results show a monotonic ordering across groups (humans > larger-scale models > semi-larger-scale models > smaller-scale models). In comparison with human participants, LLM performance is more sensitive to task-format complexity. In addition, we identify a set of completely failed items that consistently mislead all evaluated LLMs, exposing shared and systematic weaknesses in current models’ pedagogical grammar recognition. Overall, this study provides an operational framework for diagnosing the capabilities and risks of LLMs when they are deployed as assistants or teammates in grammar-related language-education tasks and offers empirical reference for safer and more syllabus-aligned use of LLMs in educational settings.

A static load test of small-scale models of driven and bored piles with a change in length in conditions of homogeneous soil was carried out. The bearing capacity of the piles was calculated using the method recommended by the current standards of Ukraine [1]. It was established that the real bearing capacity of bored and driven piles differs from that calculated using the methods recommended by the current regulatory documents [1]. The purpose of this work is to compare the operation of driven and bored piles under load in small-scale physical modeling. Based on the results of small-scale physical modeling, the difference in the operation of bored and driven piles was established, and the inaccuracy of analytical calculations of the bearing capacity of piles was confirmed and indicates the complex processes of pile-soil interaction.

Compositional generalization—the systematic combination of known components into novel structures—is fundamental to flexible human cognition, yet the mechanisms that enable it in neural networks remain poorly understood in both machine learning and cognitive science. [1] showed that a compact encoder-decoder transformer can achieve simple forms of compositional generalization in a sequence arithmetic task. In this work, we identify and mechanistically interpret the circuit responsible for this behavior in such a model. Using causal ablations, we isolate the circuit and show that this understanding enables precise activation edits to steer the model’s outputs predictably. We find that the circuit performs function composition without encoding the specific semantics of any given function—instead, it leverages a disentangled representation of token position and identity to apply a general token remapping rule across an entire family of functions. Although the circuit mechanism was identified in a limited number of small scale models with a synthetic task, it sheds light to how symbolic compositionality can emerge in transformers and offer testable hypotheses for similar mechanisms in large-scale models. Code for model and analysis is publicly available.

Compositional generalization-the systematic combination of known components into novel structures-is fundamental to flexible human cognition, yet the mechanisms that enable it in neural networks remain poorly understood in both machine learning and cognitive science. [1] showed that a compact encoder-decoder transformer can achieve simple forms of compositional generalization in a sequence arithmetic task. In this work, we identify and mechanistically interpret the circuit responsible for this behavior in such a model. Using causal ablations, we isolate the circuit and show that this understanding enables precise activation edits to steer the model's outputs predictably. We find that the circuit performs function composition without encoding the specific semantics of any given function-instead, it leverages a disentangled representation of token position and identity to apply a general token remapping rule across an entire family of functions. Although the circuit mechanism was identified in a limited number of small scale models with a synthetic task, it sheds light to how symbolic compositionality can emerge in transformers and offer testable hypotheses for similar mechanisms in large-scale models. Code for model and analysis is publicly available.

Large and small-scale models’ fusion-driven proactive robotic manipulation control for human-robot collaborative assembly in industry 5.0

Abstract Regarding LCCO 2 , which is the total amount of CO 2 emissions resulting from the procurement of building materials to their disposal, discussions are underway not only on operational carbon, which is generated during the operation of a building, but also on embodied carbon, which is generated during the manufacturing, transportation, construction, renovation, and disposal of materials. Regarding operational carbon, simulation methods and standards for energy use have been established, and their evaluation is progressing with the aim of promoting ZEH and ZEB. However, regarding embodied carbon, it is necessary to understand the various actual situations at the stages of a building’s life cycle, and the establishment of estimation methods and reduction methods are under consideration. Regarding embodied carbon, attention has been focused on the material consumption of building, but it is also important to consider a method of calculation from detailed items assuming actual construction, such as building materials and construction activities associated with construction and demolition. In particular, it is possible to aim to reduce embodied carbon by using different construction methods even though the same materials are used, and this time we focused on the differences in construction methods. In this study, we compared the LCCO 2 of small-scale housing models with different insulation construction methods to examine the construction method that is most effective in reducing CO 2 emissions. In addition, we focused on insulation materials to improve the insulation performance of buildings, thereby reducing energy consumption for heating and cooling, and calculated the LCCO 2 reduction effect of buildings. The construction method using insulation panels was the most effective in reducing LCCO 2 . In addition, construction methods that do not allow separation after the building is demolished were at a disadvantage because they could not be used as recycled resources.

We propose a response-controlled system installing non-compression members to linear structures to avoid the resonance phenomenon that gives a strong influence on the vibration of architectural structures. In this paper, we study the basic seismic response characteristics of the system. Firstly, using a small-scaled rigid-jointed frame model, we study analytically and experimentally. Next, using a cylindrical latticed shell structure, we propose a reinforcement method using cables and study the seismic response characteristics focusing on the initial tensile cable force caused by structural deformation under gravity loading. From this study, we confirmed the high control performance of the system.

工业 AI Agent 是融合工业专业知识与人工智能技术, 能在工业场景中自主完成“感知、决策、执行、学习”闭环的智能体. 工业 AI Agent凭借其自感知、自学习、自决策、自执行、自适应等能力, 在推动工业生产从被动响应向主动预判转变、实现全流程自主化与全局协同优化等方面展现了巨大潜力. 为此, 本文将ChatGPT、Qwen等大模型所具备的全局认知、多模态交互与智能推理等能力, 与工业软件、机器学习算法、数字孪生模型等小模型在专业性、轻量化部署、低成本运行方面的核心优势深度融合, 提出了一种”以大模型为总指挥、以小模型为具体执行”的工业AI Agent全新范式. 据此设计了物理空间、感知空间、决策空间、行为空间、学习空间和应用空间六位一体的工业AI Agent新架构, 阐明了大模型与小模型的协同运行机理. 最后, 通过工业AI Agent赋能的人造卫星关键零部件自主装配案例, 验证了所提架构、模型与方法的合理性与有效性.

Water-quality degradation from nutrient pollution remains a major challenge for resource managers. Developing effective strategies requires tools to characterize nutrient sources and transport. This study used the RSPARROW framework to develop and assess new, smaller-scale models for Total Nitrogen (TN) and Total Phosphorus (TP) transport across Mississippi (MS). These state-level models were built using 15 years (2005–2020) of observation data and considered variables including multiple nutrient sources, land characteristics, and attenuation processes. The MS models demonstrated comparable accuracy to larger regional SPARROW models, validating the use of smaller-scale models for local management. Results showed agricultural sources are the major contributors to TN, dominated by fertilizer in northern MS and livestock manure in the south. Urban land cover also significantly influenced TN and was the second most significant source of TP, following geologic material (background P). Fertilizer and manure were also important TP sources. This study provides valuable, spatially explicit data on nutrient distribution in MS streams, supporting the state’s nutrient reduction planning. It concludes by highlighting the need for future model improvements via updated source data and mean annual flow estimates.