Prediction of the Residual Compressive Strength of Rice Husk Ash Concrete after Exposure to Elevated Temperatures Using XGBoost Machine Learning Algorithm

Abstract

The study aimed to assess the applicability of XGBoost in determining the residual compressive strength of rice husk ash concrete exposed to elevated temperature, reducing the need for costly, time-consuming laboratory experiments. Data was collected from the available literature, with 75% used for training and 25% for testing. Synthetic data was created using the Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP). The model accuracy was checked using statistical scores: coefficient of determination (R2), root mean squared error (RMSE), and mean absolute error (MAE). SHAP values were used for feature importance analysis. Coding was done in Python using Jupyter Notebook. With the original data, the model produced RMSE, R2, and MAE test values of 3.351, 0.939, and 2.994, respectively, indicating excellent performance. The combined original and synthetic dataset gave RMSE, R2, and MAE values of 0.071, 0.941, and 0.053, respectively, signifying improved performance. The feature analysis identified higher temperature, unheated compressive strength, and water-cement ratio as the most significant factors in the XGBoost prediction. The exposure duration, alumina content, and iron oxide had minimal influence.