Katihar Engineering College

Ultra-high-performance concrete (UHPC) has emerged as an advanced construction material due to its superior strength, durability, and resistance to aggressive environmental conditions. However, predicting its compressive strength is complex because UHPC properties depend on multiple interacting material and workability parameters. Therefore, the present study aims to develop an efficient machine learning-based predictive framework for estimating the compressive strength of UHPC. A comprehensive database consisting of 360 experimental records was compiled from previous studies. Eight key input parameters, fly ash, silica fume, T500 flow time, maximum spread diameter, J-ring value, h2/h1 ratio, V-funnel time, and temperature were considered as predictors. Two hybrid machine learning models were developed using Gaussian Process Regression (GPR) optimized with metaheuristic algorithms, namely the Snow Geese Algorithm (SGA) and Horse Herd Optimization (HHO). The dataset was normalized and divided into training, validation, and testing subsets. Model performance was evaluated using several statistical indicators including root mean square error, mean absolute error, correlation coefficient, weighted mean percentage absolute error, and variance accounted for. The results demonstrate that both hybrid models provide reliable predictions of UHPC compressive strength, while the GPR–SGA model achieved slightly higher accuracy than the GPR–HHO model. Furthermore, explainable artificial intelligence techniques such as SHAP analysis and Partial Dependence Plots were applied to interpret the influence of input variables. The analysis revealed that temperature and silica fume content have the most significant impact on compressive strength prediction. The proposed modeling framework provides a practical decision-support tool for engineers to estimate UHPC strength efficiently, optimize mixture design, and reduce the need for extensive laboratory experimentation.

This study utilizes the effectiveness of machine learning techniques, namely multilayer perceptron (MLP), radial basis function neural network (RBFNN), and long short-term memory (LSTM) for predicting the undrained shear strength of clay (SU). 400 datasets were generated using seven inputs, viz., water content, specific gravity, void ratio, liquid limit, plastic limit, height of the clay layer, and water table position. Performance parameters, including coefficient of determination (R2), variance account factor (VAF), Legate and McCabe index (LMI), root mean square error (RMSE), maximum absolute error (UAE), and expanded uncertainty (U95), were used to assess the model’s performance. Among three models, LSTM outperformed in all the three phases due to highest value of R2 (Train = 0.995, Test = 0.992 and Validation = 0.986), VAF (Train = 99.644, Test = 99.405 and Validation = 98.80), LMI (Train = 0.931, Test = 0.911 and Validation = 0.875) and lowest value of RMSE (Train = 0.014, Test = 0.017 and Validation = 0.016), UAE (Train = 0.040, Test = 0.060 and Validation = 0.030) and U95 (Train = 0.386, Test = 0.370 and Validation = 0.270). The model’s performance was also assessed using the convergence curve, reliability index, regression plot, William’s plot, external validation, and comparative analysis.

This study investigates the structural, mechanical, and degradation properties of PVA composites (0.15–0.25 mm) reinforced with cellulose and kaolin clay, fabricated via solvent casting. The composites (1–5 wt.

This study utilizes three machine learning (ML) models, namely artificial neural network (ANN), Gaussian process regression, and adaptive boosting, to predict the efficiency factor of a strip footing resting on layered soil. 400 datasets with six input variables, including the angle of shearing resistance of the upper and lower soil layer, the unit weight of upper and lower soil layers, the depth of the footing, and the thickness of the upper soil layer, were used to compute the output. To legitimize the precision of these ML models, different performance parameters were used. These performance parameters include coefficient of determination (R2), variance account factor, performance index, root mean square error (RMSE), mean absolute error, and maximum absolute error. Results showed that the ANN model had the most suitable outcomes among these three ML models, with the highest R2 = 0.988 and the lowest RMSE = 0.02 in the training phase, and R2 = 0.964 and RMSE = 0.035 in the testing phase, respectively. The first-order second moment technique was used to compute the reliability index. The efficacy of the models was also checked by using score analysis, regression plots, William’s plots, external validation, and comparative analysis. Sensitivity analysis was carried out to check the influence of each input parameter on the output.

Seepage discharge and piping failure are among the most critical concerns affecting the safety and long-term stability of concrete dams, requiring reliable predictive tools for effective risk assessment and design. Traditional empirical and analytical approaches often face limitations in accurately capturing the complex nonlinear interactions among geotechnical and hydraulic parameters. Therefore, this study investigates the applicability of machine learning (ML) techniques for predicting seepage discharge (SD) and factor of safety against piping (FSP) in concrete dams. Three ML models, namely artificial neural network (ANN), extreme gradient boosting (XGBoost), and support vector regression (SVR), were developed using five key input parameters: specific gravity of soil solids, porosity, effective soil particle size, and upstream and downstream heads. The performance of these models was assessed using several evaluation metrics, including the coefficient of determination (R²), Willmott’s index, a20-index, root mean square error (RMSE), scatter index, and mean absolute error, during both the training (tr) and testing (ts) phases. The results reveal that the ANN model achieved superior predictive accuracy compared with XGBoost and SVR, yielding the highest R² values of 0.999 (training) and 0.997 (testing), and the lowest RMSE values of 0.001 (training) and 0.003 (testing) for SD prediction. Similarly, for FSP prediction, ANN demonstrated excellent performance with R² values of 0.998 (training) and 0.977 (testing), and RMSE values of 0.006 (training) and 0.028 (testing). The performance of the models was further evaluated using rank analysis, reliability analysis, regression curve, William’s plot, objective functional criteria, uncertainty analysis, the Anderson–Darling test, the Mann–Whitney test, and external validation. Based on all performance criteria, the ANN model consistently outperformed XGBoost and SVR in predicting both seepage discharge (SD) and factor of safety against piping (FSP). The findings indicate that ML-based approaches, particularly ANN, provide a reliable and computationally efficient framework for predicting SD and FSP, offering significant potential for enhancing dam safety evaluation and supporting data-driven decision-making in geotechnical engineering.