Hybrid Physics-Constrained and Data-Driven Approach for Interwell Saturation Estimation from Well Logs
Abstract Combining physics-based models for well log analysis with artificial intelligence (AI) advanced algorithms is crucial for wellbore studies. Data-driven methods do not generalize well and lack theoretical knowledge accumulated in the field. Estimating well saturation significantly improves if predictions from physical models are used to constrain data-driven algorithms in outlined primary fluid channels and other important points of interest. Saturation propagations in the reservoirs interwell region also generalize better under using combination of models. This work addresses combined usage of theoretical and data-driven models by aggregating them into single hybrid model. Multiple physical and data-driven models are under study, their parameters are optimized using observations. Weighted sum is used to predict water saturation at every point with weights being recomputed at each step. Model outputs are compared in terms accuracy and cumulative loss. A synthesized reservoir box model encompassing volumetric interwell porosity, resistivity and saturation data is used for the validation of the algorithms. Aggregated model for estimating interwell saturation shows improved prediction accuracy compared both to physics-based or data-driven approaches separately.