Summary Machine learning offers new opportunities for geophysical inverse problems, yet conventional regularized inversions of potential field data remain limited by global smoothing constraints and low structural resolution. We propose a locally adaptive, data-driven framework that combines synthetic Earth model generation and ensemble learning for joint gravity and magnetic interpretation. Training models are generated using geologically informed Voronoi-based geometries and planar structures, and a random forest classifier is trained on local statistical features of gravity and magnetic anomalies. The method yields geologically consistent subsurface models that reproduce observed anomaly characteristics without explicit regularization or iterative inversion. Compared with nonlinear Bayesian and traditional regularized inversions applied to the same dataset, the approach provides a substantial reduction in computational cost while preserving key structural features. The performance of the method is inherently linked to how representative the training ensemble is with respect to the target structure, and the results should be interpreted within this context. This framework demonstrates a practical and efficient alternative for potential field inversion using machine learning.
Ghalenoei* et al. (Thu,) studied this question.