چكيده لاتين
Accurate prediction of gravity anomalies in inaccessible areas is one of the fundamental challenges in geophysical studies and exploration of subsurface resources such as oil and gas reservoirs. Geographical constraints such as mountains, lakes, and valleys prevent direct acquisition of gravity data in some locations. In such circumstances, the use of numerical interpolation methods and predictive modeling to estimate gravity and elevation values is necessary and inevitable. In this study, using gravity and elevation data within the geographical area of Iran, a set of machine learning and deep learning models were developed to interpolate these values. The models used included deep neural network (DNN), XGBoost, Random Forest, and classical methods such as IDW, Kriging, MLS, Gaussian, and Collocation. The performance of these models was evaluated based on RMSE, MAE, and R² coefficient of determination for both gravity and elevation components. Numerical results showed that the DNN model provided the most accurate performance in estimating gravitational anomalies and height, with a significant difference compared to other models. This model was able to reduce the root mean square error (RMSE) for gravitational anomaly to about 29 units and for height to about 30 units, while its coefficient of determination in both components was more than 0.998. The XGBoost and Random Forest models also had acceptable performance with errors of about 110 to 122 units and coefficient of determination of about 0.974 to 0.978. In contrast, classical methods such as IDW, Kriging, and Collocation provided lower accuracy in estimating values, with errors higher than 126 to 167 units and coefficient of determination less than 0.971. These results indicate that deep learning models, especially neural networks, have a higher ability to extract nonlinear patterns and provide accurate estimates when faced with complex and irregular data. The use of these models not only reduces interpolation errors, but also allows generalization to data-poor areas. As a result, the framework proposed in this study can be effectively used to reduce field sampling costs, increase the accuracy of subsurface modeling, and improve decision-making in the exploration stages of underground resources.
