<p>Estimating mineral resources is a critical aspect of mining decision-making, requiring accurate prediction of ore grades at unsampled locations. Traditional mathematical methods struggle to capture complex spatial relationships in geological data, whereas machine learning (ML) offers a promising alternative. However, most ML-based ore grade estimation approaches neglect the anisotropic characteristics of ore bodies, which can significantly impact prediction accuracy. To address this limitation, we propose an adaptive anisotropic generalized regression neural network (AA-GRNN) method specifically designed to account for spatial anisotropy. This model incorporates multiple smoothing factors and rotation angles to enhance its ability to interpolate ore grades under anisotropic conditions. Particle swarm optimization was employed to optimize these parameters automatically. Experimental results on the simulation and actual datasets indicate that the AA-GRNN model achieved reductions of root mean squared errors of 24.43% and 33.93%, respectively, compared to inverse distance weighting, 23.78% and 26.54%, respectively, compared to ordinary kriging, and 13.64% and 10.02%, respectively, compared to the deep neural network. These results highlight the model’s superior generalization and predictive accuracy. Thus, this paper introduces this novel and effective method for ore grade estimation, advancing the integration of anisotropy into ML-based spatial prediction models and contributing to more reliable mineral resource estimation.</p>