Geostatistics and artificial intelligence coupling: advanced machine learning neural network regressor for experimental variogram modelling using Bayesian optimization

Abstract

Experimental variogram modelling is an essential process in geostatistics. The use of artificial intelligence (AI) is a new and advanced way of automating experimental variogram modelling. One part of this AI approach is the use of population search algorithms to fine-tune hyperparameters for better prediction performing. We use Bayesian optimization for the first time to find the optimal learning parameters for more precise neural network regressor for experimental variogram modelling. The goal is to leverage the capability of Bayesian optimization to consider previous regression results to improve the output of an experimental variogram using three experimental variograms as inputs and one as output for network training, calculated from ore grades of four orebodies, characterised by the same genetic aspect. In comparison to artificial neural network architectures, the Bayesian-optimized artificial neural network demonstrably achieved the superior Coefficient of determination in validation of 78.36%. This significantly outperformed a non-optimized wide, bilayer, and tri-layer network configurations, which yielded 32.94%, 14.00%, and −46.03% for Coefficient of determination, respectively. The improved reliability of the Bayesian-optimized regressor demonstrates its superiority over traditional, non-optimized regressors, indicating that incorporating Bayesian optimization can significantly advance experimental variogram modelling, thus offering a more accurate and intelligent solution, combining geostatistics and artificial intelligence specifically machine learning for experimental variogram modelling.