Machine learning for sedimentological and facies reconstructions of hydrocarbon reservoirs: a critical review of deployed and emerging methods

Relevance. The predictive power of a geological model directly determines the efficiency and economic viability of oil and gas field development. However, reconstructing the accurate spatial structure of reservoirs is hindered by the scarcity of core data, uneven well data density, and heterogeneity of geological and geophysical information. Under conditions of digital transformation in the oil and gas industry, machine learning methods are increasingly being integrated into geological modeling, complementing and enhancing classical geostatistical approaches. This integration opens up new opportunities for improving the accuracy of facies and sedimentological reconstructions. Aim. To systematize global experience in the application of machine learning methods for sedimentological and facies reconstructions of oil and gas reservoirs, identifying both already implemented solutions and alternative technologies at an early stage of adoption. Methods. A critical review of peer-reviewed studies covering Bayesian networks, variational autoencoders, generative adversarial networks, hybrid models, and classical geostatistical algorithms was conducted. The maturity of these technologies was assessed based on data types, expert involvement, and validation metrics. Results and conclusions. Machine learning methods go beyond simple spatial interpolation by uncovering complex, multidimensional relationships and reconstructing depositional scenarios. In practice, the use of generative and probabilistic graph models is expanding, underscoring the importance of multi-type data and geological context. Despite increasing automation, expert input remains critical – especially in Bayesian approaches. The main challenge remains the lack of sufficiently labeled training datasets; however, Bayesian methods demonstrate robustness even with limited data. A growing trend towards hybrid approaches that combine the strengths of machine learning and geostatistics has been observed, highlighting the continued importance of domain experts in modeling.