Application of Machine Learning Methods for Gravity Anomaly Prediction

Gravity anomalies play critical roles in geological analysis, geodynamic monitoring, and precise geoid modeling. Obtaining accurate gravity data is challenging, particularly in inaccessible or sparsely covered regions. This study evaluates machine learning (ML) methods—Support Vector Regression (SVR), Gaussian Process Regression (GPR), and Ensemble of Trees—for predicting gravity anomalies in southeastern Kazakhstan and compares their effectiveness with traditional Kriging interpolation. A dataset, consisting of the simple Bouguer anomaly values, latitude, longitude, elevation, normal gravity, and terrain corrections derived from historical maps at a scale of 1:200,000, was utilized. Models were trained and validated using cross-validation techniques, with performance assessed by statistical metrics (RMSE, MAE, R2) and spatial error analysis. Results indicated that the Exponential GPR model demonstrated the highest predictive accuracy, outperforming other ML methods, with 72.9% of predictions having errors below 1 mGal. Kriging showed comparable accuracy and superior robustness against extreme errors. Most prediction errors from all methods were spatially associated with mountainous regions featuring significant elevation changes. While this study demonstrated the effectiveness of machine learning methods for gravity anomaly prediction, their accuracy decreases in complex terrain, indicating the need for further research to improve model performance in such environments.