Predicting uplift capacity of square anchor plate using machine learning techniques

Abstract

Regarding analysis and design, the way anchors react to uplift forces is quite significant. Engineers require correct knowledge of the uplift capacity of anchors, a crucial characteristic for assessing the uplift force, in order to appropriately assess the stability of anchors. However, because field and laboratory studies are intricate and time-consuming, quantifying uplift force is typically difficult. This paper uses numerical test data from a research study conducted at the Civil Engineering Department at Jadavpur University to predict the uplift capacity using machine learning (ML) techniques. To forecast the uplift capacity, the study used four machine learning (ML) algorithms: Simple Linear Regression (SLR), Random Forests (RFs), Stochastic Gradient (SGD), and eXtreme Gradient Boosting (XGBoost). These ML models were tested for effectiveness and generalizability using the remaining 20% of the datasets after they had been trained on 80% of them. Through three methods, including Bayesian optimisation, random search CV, and grid search CV with k-fold cross-validation, the hyperparameters for each ML model were adjusted. Five alternative metrics, including R2 score, mean absolute error (MAE), mean squared error (MSE), maximum error (ME), and mean absolute percentage error (MAPE), were used to assess each ML model's performance. The results demonstrated that the XGBoost model consistently performed well across all metrics. It achieved high accuracy and the lowest level of errors, indicating superior accuracy and precision in predicting uplift capacity. The RF model exhibited average performance, with slightly higher error metrics compared with the XGBoost model. However, the Linear Regressor and SGD model performed poorly, with very higher error rates and uncertainty in predicting uplift capacity. Based on these results, we can conclude that the XGBoost model is highly effective at accurately predicting uplift capacity of anchors using the data with minimal input features.