ANALYSIS OF ARTIFICIAL NEURAL NETWORK UTILIZATION FOR BOUGUER ANOMALY INTERPOLATION

• Main Author: Luqman Addura, Muhamad
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/87879
• Institution: Institut Teknologi Bandung

This research successfully developed an Artificial Neural Network (ANN) program for interpolating Bouguer anomaly data using Python. The study explored 162 combinations of hyperparameters to optimize the interpolation, varying the number of neurons (256, 512, 1024), activation functions (ReLU and Hyperbolic Tangent), batch sizes (16, 32, 64), learning rates (0.0001; 0.001; 0.01), and the number of iterations (100, 500, 1000). Applys on two simple synthetic models demonstrated that ANN effectively captures non-linear patterns, with results heavily influenced by hyperparameter selection and the amount of training data. For the first synthetic model, assuming a body with anomalies in the Earth's crust and upper mantle layers, the best ANN configuration was obtained with a learning rate of 0.001, ReLU activation function, 1000 epochs, batch size of 32, and 512 neurons. This configuration yielded a Root Mean Square Error (RMSE) of 6.413055 and a coefficient of determination (R²) of 0.989493. In contrast, for the second synthetic model, which consisted of two opposing bodies, the best configuration used a learning rate of 0.001, Hyperbolic Tangent activation function, 1000 epochs, batch size of 32, and 1024 neurons. This resulted in an RMSE of 0.06768 and an R² of 0.9650693. The study also compared the application of the ANN method with the Kriging method. The Kriging method demonstrated superior performance in capturing local spatial patterns in areas with strong spatial correlations. For the Kriging prediction model, the first synthetic model yielded an RMSE of 25.41172 and an R² of 0.99799, while the second synthetic model achieved an RMSE of 0.04794 and an R² of 0.9998. Although the RMSE and R² evaluations between ANN and Kriging are relatively similar, ANN provided more varied predictions beyond the range of observed data, with results closely approaching those of Kriging.