A Multi-Observation-Constrained (MOC) Method For The Electrical Characteristics Prediction Of Reservoir Cores

The establishment of accurate and reliable reservoir core conductivity prediction methods can provide more precise geological information and scientific basis for exploration and development,thereby improving the success rate and efficiency of exploration and development.With the development of digital rock modeling techniques,numerical simulation of rock physics has become an important means to predict rock electrical properties,but it still faces challenges such as consuming computational resources and difficulties in batch processing data.In order to improve research efficiency,artificial intelligence-related technologies have begun to receive attention in the field of digital rocks.However,limitations such as insufficient prior data and difficulties in quantifying uncertainties in classical machine learning regression predictions have posed the application of artificial intelligence in this field.This study proposes a Multi-Observation-Constrained(MOC)method for the electrical characteristics prediction of reservoir cores,based on the Bayesian evidence learning framework using a multi-source data fusion method and uncertainty quantification.To address the lack of robustness of deep learning models in engineering fields with challenging data acquisition,a combination of generative adversarial networks and physics-informed neural networks is used to improve modeling accuracy with limited samples,generating high-precision 3D digital rock models as prior models.Based on these prior models,forward simulations are performed to obtain the pore structure characteristics and conductivity as training data for the prediction models.Multiple machine learning techniques such as canonical correlation analysis,kernel density estimation,XGBoost,and MAPIE interval prediction are used to establish two prediction models,analyzing the underlying relationship between pore structure characteristics and conductivity to constrain the posterior probability distribution of conductivity and reduce uncertainty.The study conducts conductivity prediction experiments based on pore structure information of 3D Bentheimer sandstone and 2D shallow gray medium sandstone,performs outlier detection on the observed data features before prediction,and conducts hypothesis testing on the predicted results to ensure their reliability and validity.Additionally,uncertainty curves are used to evaluate the performance of the two prediction models in terms of accuracy.The results show that generative adversarial networks with physics-informed constraints can quickly generate high-precision prior models.The two established prediction models effectively reduce uncertainty based on the prior models,providing assistance for engineers’ subsequent decision-making.