Machine learning and neural networks in predicting grain-size of sandy formations

A total of 2520 sand samples, obtained from grain size analysis, were used in this study. The primary aim is to predict sand grain sizes using machine learning and deep learning algorithms. The input data consists of the five grain size fractions of sands based on the Udden-Wentworth scale, while the output data represents the five sand types (very coarse, coarse, medium, fine, and very fine) according to the average grain size. The machine learning algorithms employed are Random Forest and XGBoost, while the deep learning algorithms include MLP (Multilayer Perceptron) and LSTM (Long Short-Term Memory). All algorithms were implemented using Python. The evaluation metrics used for model assessment include K-fold validation, confusion matrix, and accuracy. The dataset was split into training (70 %, 1764 samples) and validation (30 %, 756 samples) sets. The study reveals that LSTM and MLP neural networks are better suited for predicting sand sizes, with MLP achieving the highest accuracy at 99.6 %. While machine learning algorithms performed well, they slightly lagged behind neural networks, with Random Forest achieving 99.07 % accuracy and XGBoost 98.81 %. In terms of sand type classification, all algorithms predicted very fine and very coarse sands with 100 % accuracy. However, fine, medium, and coarse sands showed some susceptibility to misclassification. Coarse sands were the most prone to misclassification, particularly being misidentified as medium or very coarse sands.