Convolutional Neural Networks to Estimate Residual Oil Saturation Through Partitioning Interwell Tracer Tests

Summary Modeling of chemical tracers represents one of the most powerful dynamic tools for reservoir characterization and estimation of oil saturation. However, the continuous monitoring during long times in partitioning interwell tracer tests (PITTs), which extend to months or years in some field tests, limits the use of this technology. The large distance between wells and high partitioning coefficients are some of the main reasons for the slow production of tracers, where time-consuming numerical simulations are required to analyze the tracer data in reservoir models before field applications. Therefore, this work presents an innovative machine-learning (ML) workflow using convolutional neural networks (CNNs) for the estimation of residual oil saturation (Sor) based on the generation of partitioning tracer responses in heterogeneous media. To train the CNN model, a Python-based algorithm was developed to generate permeability and porosity fields providing significant information about the behavior of tracer production data. The CNN model was trained with extensive ideal and partitioning tracer profiles generated from PITTs in reservoirs under Sor using numerical simulations with the UTCHEM software (UTCHEM 2019). The response feature in the CNN model corresponds to partitioning tracer profiles obtained from ideal concentration curves to analyze the tracer arrival delay caused by the volume of oil remaining in the reservoir. Numerical case studies and field data were examined to show the applicability of the CNN model developed under multiple flow conditions, where the estimation of Sor is based on a trial-and-error method to match the early partitioning tracer response, which is a simple process since the only unknown is Sor. In most cases, the partitioning tracer responses were properly estimated from ideal tracer data, resulting in a difference of less than ±0.02 compared with the value of Sor calculated. Thus, we show that CNNs represent potential methods for predicting the concentration behavior as a function of early tracer data measured. The work presented is designed to be a starting point for the development of novel ML algorithms for interpreting tracer tests in reservoirs.