An automatic fitting method for a variogram based on deep learning

A variogram serves as a crucial tool for quantifying spatial correlations. However, existing variogram fitting methods often yield unstable results. This study proposed an automatic variogram fitting method based on deep learning, aiming to enhance the precision and stability of automatic fitting. The fitting of the experimental variogram is essentially a nonlinear optimization problem, which involves optimizing the matching between the experimental and theoretical variograms. The proposed method generated substantial training datasets using several sets of theoretical variograms with varying parameter values for training and learning in deep neural networks. The trained model was then used for the automatic fitting of the experimental variogram. Multiple sets of experimental results demonstrate that based on the robust fitting capability of deep neural networks, the proposed method manifested superior fitting stability and computational efficiency compared to the least squares method, providing a novel approach for automatic variogram fitting in geostatistics.