Uncertainty quantification by using Monte-Carlo neural network method for water saturation log prediction

One of the most crucial and challenging components of the reservoir’s property is calculating water saturation. It offers an approximation of the hydrocarbon content. These parameters have been determined for a long time using traditional techniques like measurement from core sample and formula calculation from well logs. On the other hand, thanks to recent developments in computing technology, Machine Learning (ML) has a promising future and the potential to be employed in data prediction. The ML, which was already well known for its outstanding prediction performance in recent years, might have produced an unstable model when a new parameter called uncertainty was incorporated in the performance calculation. The degree of uncertainty describes the model’s inability to produce precise predictions because there is insufficient data or a firm theoretical foundation upon which to do so. The Uncertainty Quantification (UQ) taken into consideration in the parameter helps to clarify and assess the predictability of the model. A powerful prediction model may run a considerable risk of failing when it comes to real implementation if the UQ in the prediction is disregarded or overlooked. A statistical machine learning approach, like the Monte-Carlo Neural Network method, which can quantify the uncertainty in water saturation data prediction, was used to resolve issue. The study uses the elastic characteristics as the input for machine learning, and the results give the uncertainty and water saturation models. It was shown that statistical ML produced accurate predictions of water saturation with correlation coefficient scores up to 83 percent and low uncertainty. The study uses the elastic properties as the input for machine learning, and the results give the uncertainty and water saturation models. It was shown that statistical ML produced accurate predictions of water saturation with correlation coefficient scores up to 83 percent and low uncertainty. This suggests that statistical-based machine learning has a good probability of producing improved results when put into practise and provides more information into the distribution of uncertainty inside the blind test well location.