Seismic bearing capacity is a critical factor in designing shallow foundations, particularly in earthquake-prone regions, where traditional analytical methods often fall short. This study explores machine learning (ML) techniques to predict the seismic bearing capacity factor of shallow strip footings, explicitly considering the influence of both primary (P) and secondary (S) seismic waves. Multiple evolutionary ML models - Decision Trees (DT), Random Forest (RF), Adaptive Boosting (AdaB), and CatBoost (CB) - were developed and evaluated using robust statistical metrics. Among these, the CatBoost model emerged as the most accurate and reliable, delivering strong performance with R 2 (coefficient of determination) values of 0.957 in training and 0.936 in testing, alongside low mean absolute error (MAE) and root mean square error (RMSE). Notably, CatBoost surpassed the standard analytical formulas, underscoring its superior predictive power. A comprehensive parametric analysis further confirmed the model's reliability, showing strong agreement with trends reported in established literature. These findings demonstrate CatBoost's effectiveness as both a research tool for soil-structure interaction analysis under seismic conditions and a practical solution for geotechnical engineers requiring reliable bearing capacity assessments. • ML captures soil's nonlinear seismic response, replacing conventional analyses • Gradient boosting outperforms classical models in predicting seismic bearing capacity • CatBoost shows high robustness and accuracy in seismic bearing capacity prediction • CatBoost shows stable, generalizable performance for static and seismic loading cases
Segment Anything Models (SAMs) are extensively used in computer vision for universal image segmentation, but deploying them on resource-constrained devices is challenging due to their high computational and memory demands. Post-Training Quantization (PTQ) is a widely used technique for model compression and acceleration. However, existing PTQ methods fail to consider the cross-attention architecture in the SAM decoder. This degradation primarily stems from the unique challenges posed by SAMs: (1) Attention dissipation, where the attention information in the decoder, which is crucial for representing segmentation masks, collapses into a diffuse and non-semantic form under low-bit quantization; and (2) Reconstruction oscillation, where bidirectional coupling within the two-way transformer introduces cross-branch error interference and destabilizes convergence. To tackle these issues, we propose CAR-SAM, a unified quantization framework tailored for SAMs. Firstly, to mitigate attention dissipation, we introduce MatMul-Aware Compensation (MAC) mechanism that transfers activation-induced quantization errors from MatMul to preceding linear weights. Secondly, to mitigate oscillation in decoder optimization, we develop a Joint Cross-Attention Reconstruction (JCAR) strategy that jointly reconstructs coupled attention branches, suppressing oscillatory behavior and promoting stable convergence. Extensive experiments show that CAR-SAM robustly quantizes SAM models down to 4-bit precision, surpassing existing methods by 14.6% and 6.6% mAP on SAM-B and SAM-L respectively.
ABSTRACT We are developing and operating the laser hammering system (LHS) that is non‐destructive remote sensing device for concrete infrastructures. LHS can digitize the hammering test which is current mainstream of the inspection of concrete by using an impact laser (short‐pulse laser), laser Doppler vibrometer, scanning system and camera. In this study, long‐range LHS (applicable distance of 30 m) and small LHS (160 kg weight, for use on aerial work platforms) are developed for the bridge inspection. Long‐range system succeeded on measurement of the concrete specimens of class 2 and 3 defect which were identified by active inspectors, and real defect of class 2 on bridge on use at the distance of 30 m. Small LHS succeeded on generation the surface vibration of the concrete specimens of class 2 and 3 defect which were identified by active inspectors. It indicates Long‐range LHS and small LHS satisfied the inspection requirement as the support tool of hammering test prescribed by Japanese government. These devices are available for support to inspectors and digitalization of concrete sound status to achieve smart infra‐maintenance.
Chloride-induced corrosion remains one of the main durability concerns for reinforced concrete exposed to marine or de-icing environments. Conventional diffusion-based models often neglect the chemical form of chloride and the role of counter-cations in altering hydrated cement. In practice, chloride transport is a reactive process controlled by simultaneous diffusion, binding, dissolution/precipitation, and pH buffering within the evolving cement matrix. This study investigates how different cations Na⁺, K⁺, Ca²⁺, and Mg²⁺ affect chloride ingress, binding, and hydrate stability in saturated concrete. A reactive transport model is developed that couples diffusion, aqueous speciation, mineral equilibrium, kinetic reactions, and surface complexation on C-S-H. The simulations reproduce and extend the experimental results of literature for four boundary solutions: 0.5 mol/l NaCl, 0.5 mol/l KCl, 0.25 mol/l CaCl₂, and 0.25 mol/l MgCl₂, over exposure periods up to ten years in saturated concrete. Under NaCl and KCl, the pore network remains stable, alkalinity is maintained, and binding is moderate producing deep free-chloride penetration. Under CaCl₂ and MgCl₂, strong near-surface reactions occur: AFm phases convert into Kuzel-type compounds, and portlandite dissolution with C-S-H decalcification produces brucite or M-S-H. These transformations trap chloride near the surface, limit transport, and reduce pH in the outer zone. Consequently, monovalent salts lead to transport-controlled ingress, while divalent salts cause binding/microstructure-controlled accumulation. Reliable prediction of corrosion risk requires evaluating free chloride, total chloride, and alkalinity together. Reactive transport modeling thus provides a physically consistent and predictive framework for performance-based durability design of concrete under diverse chloride environments.
Read moreThe design of grouted corrugated duct connections (GCDCs) in precast concrete structures requires adequate length of corrugated duct and embedded bar. Both parameters have been extensively studied for conventional steel-reinforced concrete (RC). While equations for corrugated duct length may be adapted to other reinforcement types, bar embedded equations cannot be directly applied to glass-fibre reinforced polymer (GFRP) bars due to their different characters. Nevertheless, studies on GCDCs in GFRP-RC remain limited. To address this gap, this study investigates the behaviour and design of GCDCs in precast GFRP-reinforced concrete and proposes appropriate design equations and procedures. A finite element (FE) model was developed and verified against experimental tests of column-footing assemblies. Several bond models were also evaluated to identify the most appropriate representation of GFRP to grout bond-slip behaviour. The verified FE model was subsequently adopted in a parametric study to examine the effects of bar embedded length, bar diameter, and grout strength. The results highlight the combined influence of these parameters on the column performance, including lateral capacity and bar stress development. The findings demonstrate that the predictive equation for the minimum required embedded length developed for steel bars is not applicable to GFRP bars. Based on the FE results, a refined equation was proposed to estimate the minimum required embedded length of grouted GFRP bars. This equation considers the GFRP bar diameter, grout strength, and bar stress. The refined equation reliably predicts the minimum embedded length required to prevent bar pull-out.
Read moreThe durability of reinforced concrete structures in chloride-rich environments remains a major concern in infrastructure design, particularly in coastal regions. While standardized laboratory procedures provide reliable quantification of chloride ingress resistance, they are often time-consuming, costly, and unsuitable for early-stage mix design. This study proposes a data-driven framework for predicting the chloride resistance level of concrete using tree-based machine learning (ML) classifiers. A comprehensive experimental dataset was utilized to train and validate three ML models: Decision Tree Classifier (DTC), Random Forest Classifier (RFC), and CatBoost Classifier (CatBC). Extensive hyperparameter tuning was performed using the Optuna framework with 2000 trials per model to enhance predictive performance. Among the tested models, CatBC outperformed its counterparts with a test accuracy of 0.95 and weighted F1-score of 0.85. Feature importance analyses using SHAP values, Prediction Values Change, and other CatBoost interpretability tools consistently identified the water-to-binder ratio, superplasticizer content, test age, and aggregate proportions as key predictors of chloride resistance. The findings demonstrate that machine learning offers a fast, cost-effective, and accurate alternative for classifying concrete’s chloride resistance, supporting informed decision-making in mix design and service-life assessment.
Read moreSYMBOLIC MATRIX STRUCTURAL ANALYSIS OF TRUSSES: STATIC AND DYNAMIC APPLICATIONS - 10th International Conference on Computational Methods in Structural Dynamics and Earthquake Engineering - 15-18 June 2025, Rhodes Island, Greece
Read moreAbstract Vowel glottalization, as voice quality variation, occurs at phrase edges or under prominence, often in conjunction with resolving onsetless syllables or vowel-vowel hiatus. This study examines whether and how voice quality in two varieties of English, American and Australian, is manifested at phrase edges and under prominence in vowels that are neither strictly initial nor final. Analyses of spectral and noise measures show that both varieties utilize voice quality to signal prominence and boundaries, even in non-edge contexts. Specifically, increased glottalization reflects stronger laryngeal articulation in line with domain-initial and prominence-induced strengthening. Conversely, phrase-final positions are characterized by phrase-final creak, which is often observed with phrase-final weakening. While both varieties show similar vowel glottalization usage, Australian English tends to use voice quality more extensively to mark focus than American English, with a greater overall tendency toward glottalization. This suggests that the impressionistic difference perceived in the general use of voice quality between the varieties (more robust glottalized vowels in Australian English) stems from differences in the use of glottalization in marking prosodic structure. These findings underscore the role of non-contrastive voice quality in shaping prosodic structure across varieties while also revealing dialect-specific interactions between phonetics and prosody in its manifestation.
Read moreReinforced concrete (RC) structures are among the most prevalent in the construction industry. However, various in-service conditions such as loading, environmental exposure, and construction practices can lead to the formation of concrete cracks. In extreme cases, these cracks may propagate through the cross-section of structural members, creating section pre-cracks. This study examines the influence of section pre-cracks on the shear failure behavior of RC deep beams (with a shear span-to-depth ratio of 1.57) through three-point bend-loading tests and 3D RBSM analysis. The primary experimental variables include the number of pre-cracks (one or two) and their width (0.5 mm or 1.0 mm). The findings reveal that pre-cracks significantly reduce the initial stiffness and shear strength of deep beams, primarily by disrupting the transmission of axial compressive stress in the concrete, thereby diminishing the contribution of arch action to shear strength. Furthermore, it is observed that the greater the total width and number of pre-cracks, the more significant the reduction in shear strength. In addition, combining both experimental tests and numerical simulations, a total of 48 deep beams were subjected to shear failure tests. Based on the shear strength data, two degradation models for shear strength (one representing the average trend and the other a conservative lower-bound envelope model) were developed in relation to the total crack width.
Read moreĐặt vấn đề: Sự kết hợp giữa vi khuẩn H. pylori và ung thư dạ dày (UTDD) cùng với sự gia tăng tỷ lệ lây nhiễm trên toàn thế giới, cho thấy sự cấp thiết của việc tìm ra các chiến lược phòng ngừa bệnh. Việt Namhiện nay là một trong những nước có tỷ lệ nhiễm H. pylori cao. Gen cagA, vacA được đặc biệt chú ý trong UTDD. Hiện nay, ở nước ta chỉ mới có một số nghiên cứu làm sáng tỏ một phần mối liên quan chủng H.pylori có cagA, vacA ở bệnh nhân UTDD. Tuy nhiên, cho đến nay còn ít nghiên cứu đề cập đến việc phân tích biểu lộ gen iceA liên quan với các gen cagA, vacA của H. pylori ở bệnh nhân ung thư dạ dày.Mục tiêu: Tìm hiểu mối liên quan kiểu gen iceA, cagA, vacA của H. pylori và mô bệnh học ở bệnh nhân (BN) ung thư dạ dày.Đối tượng và phương pháp: Đối tượng nghiên cứu: Gồm 91 bệnh nhân UTDD (nhóm bệnh) và 92 bệnh nhân viêm dạ dày mạn tính (nhóm chứng), được chọn trong số những người đã đến nội soi dạ dày và đượcchỉ định sinh thiết niêm mạc dạ dày để chẩn đoán xác định tại Khoa Thăm dò chức năng.Kết quả: Các bệnh nhân UTDD có hình ảnh mô bệnh học (MBH) biệt hóa kém chiếm tỷ lệ cao nhất ở cả hai nhóm cagA và vacA dương tính là 55,4% và 54,5%. Không có sự khác biệt các kiểu gen iceA1 và iceA2 giữa thể tuyến ống và thể tế bào nhẫn ở bệnh nhân UTDD với p > 0,05. Không có thể MBH tuyến chế nhày có H. pylori mang gene iceA. Kiểu gen iceA1 chiếm 54% ở nhóm MBH UTDD biệt hóa kém, 32% ở nhóm biệt hóa vừa. Kiểu gen iceA2 chiếm 50% ở nhóm biệt hóa kém và 40% ở nhóm biệt hóa vừa. Sự khác biệt giữa các kiểu gen A1 và A2 ở các nhóm MBH trên bệnh nhân UTDD ở nhóm biệt hóa vừa và kém có ý nghĩa thống kê với p < 0,05.Kết luận: Không có mối liên quan giữa các týp cagA, vacA; các kiểu gen với các đặc điểm mô bệnh học của ung thư dạ dày theo WHO năm 2010. Sự khác biệt giữa các kiểu gen iceA1 và iceA2 ở các bệnh nhânung thư dạ dày ở nhóm biệt hóa vừa và kém có ý nghĩa thống kê (p < 0,05).
Read moreTunnel squeezing is characterized as a significant degree of distortion in the surrounding rock mass that is typically larger than the designed deformation. The squeezing potential of rocks around tunnels can result in support failures, floor heave, and even flood disasters. In this study, the squeezing potential of rocks around tunnels were estimated by employing a hybrid intelligent framework to improve the performance of a classification algorithm. A total of 139 adjacent rock-squeezing patterns were acquired from places such as China, Nepal, and India to form the empirical basis for this study. The data consists of five influential variables, i.e., strength factor, tunnel depth, rock mass quality index, tunnel equivalent diameter and support stiffness. The mechanism of prediction consisted of three steps. Firstly, factor analysis was utilized to reduce the number of influential variables. The resulting factors were then categorized using k-means clustering. Finally, a random forest algorithm was developed to predict various levels of surrounding rock squeezing potential of rocks around tunnels. The proposed hybrid intelligent framework achieved a strong predictive capability of 96%, contributing to safer and more sustainable tunneling practices by reducing operational risks and improving overall structural stability.
Read moreAdvancements in artificial intelligence (AI) predictive models have emerged as valuable tools for predicting survival outcomes in allogeneic haematopoietic stem cell transplantation (allo-HSCT). These models primarily focus on pre-transplant factors, while algorithms incorporating changes in patient's status post-allo-HSCT are lacking. The aim of this study was to develop a predictive soft computing model assessing survival outcomes in allo-HSCT recipients. In this study, we assembled a comprehensive database comprising of 564 consecutive adult patients who underwent allo-HSCT between 2015 and 2024. Our algorithm selectively considers critical parameters from the database, ranking and evaluating them based on their impact on patient outcomes. By utilising the Data Ensemble Refinement Greedy Algorithm, we developed an AI model with 93.26% accuracy in predicting survivorship status in allo-HSCT recipients. Our model used only seven parameters, including age, disease, disease phase, creatinine levels at day 2 post-allo-HSCT, platelet engraftment, acute graft-versus-host disease (GvHD) and chronic GvHD. External validation of our AI model is considered essential. Machine learning algorithms have the potential to improve the prediction of long-term survival outcomes for patients undergoing allo-HSCT.
Read moreIt is plausible that under organic rice farming conditions, with Monochoria vaginalis as the dominant weed, the Indica rice ‘Takanari’ (Tak) outperform the Japonica rice ‘Koshihikari’ (Kos) in terms of nitrogen (N) uptake and biomass production. However, how N uptake, biomass, and yield in Tak and Kos are affected by weeds under organic rice farming paddy fields across multiple growing seasons remains unclear. To investigate this, we conducted a two-year field experiment (2022 and 2023) at Yamagata University Farm, Tsuruoka, Japan. Tak and Kos were transplanted individually (four seedlings/hill) or interplanted as Tak + Kos (2 + 2 seedlings/hill), with weeding and no-weeding as the main treatments. Rice and weed biomass and N uptake, as well as rice yield, were measured at harvest, with significant differences observed among all treatments. Weeding practices and growing seasons significantly affected on biomass and N uptake in both Tak and Kos. In 2023, aboveground rice biomass under weeding conditions was significantly lower than that in 2022, whereas no significant difference was observed under no-weeding conditions. Tak consistently showed higher biomass, N uptake, and yield compared with Kos across all treatments, including both seasons, weeding practices, and planting modes. The aboveground biomass ratios between Tak and Kos in interplanting mode were higher than those in individual planting mode under both weeding conditions across both years. These results suggest that Tak exhibited stronger competitive ability in terms of N uptake, biomass, and yield when interplanted with Kos. In conclusion, this two-year organic rice farming field experiment indicates that Tak may sustain higher N uptake and yield compared with Kos under both weeding and no-weeding conditions.
Read more• Physics-informed neural networks for elastoplasticity in strong and weak forms. • Strong form supports forward prediction and parameter inversion with data loss. • Data-free weak form captures plastic evolution via energy minimization. • Kolmogorov–Arnold network is compared with multilayer perceptron in both forms. Physics-informed neural networks have recently achieved remarkable success in solving elastic problems by embedding governing equations into the training of neural networks. Building upon these advances, this study extends physics-informed neural networks to material nonlinearity and develops two computational frameworks for small-strain von Mises elastoplasticity. The strong-form framework enforces governing equations through pointwise residual minimization, enabling unified forward–inverse modeling of field variables and unknown material parameters. In contrast, the weak-form framework derived from total potential energy minimization allows data-free learning of elastoplastic evolution through incremental loading, yielding stable and physically consistent predictions. The recently developed Kolmogorov–Arnold network is further incorporated and compared with the conventional multilayer perceptron. Results show that the Kolmogorov–Arnold network alleviates the gradient oscillation and convergence instability in the weak-form framework but performs less effectively in the strong form. Furthermore, validations against reference solutions from conventional numerical methods demonstrate the potential of the developed physics-informed neural network frameworks as mesh-free surrogates for elastoplastic analysis.
Read more
