New approaches to characterize and simulate reservoirs by machine learning and model-coupling strategies.

This dissertation comprises two topics. The first topic introduces an innovative multiphase, multicomponent reservoir simulator to simulate the solvent thermal resource innovation process (STRIP). In this framework, a STARS model injected steam only and governed the model by synchronizing temperature, pressure, and phase saturations for two parallel iterations of GEM models (GEM-1 and GEM-2). GEM models were used to mimic steam-CO2 co-injection by adjusting injection rates and relative permeabilities to match targeted reservoir properties. The updated relative permeabilities representing viscosity reduction by CO2 were then delivered back to the STARS model and the process repeated for the entire simulation. This new framework demonstrated the superior performance of STRIP compared to traditional steam injection. The second topic investigates classification and segmentation of geologic features from image logs. CNNs were trained to identify vuggy facies from a well in the Arbuckle Group in Kansas. The complete dataset was culled by removing poor-quality images to generate a cleaned dataset for comparison. Various types of data were used to label the image log for supervised learning. After hyperparameter optimization, median accuracy for vuggy/non-vuggy facies classification was 0.847 for the cleaned dataset (0.813 for the complete dataset). This study demonstrated the effectiveness of using microresistivity image logs in a CNN to classify facies while highlighting the importance of data quality control and hyperparameter optimization. To characterize a reservoir, geologic features need to be segmented (pixel-wise identification). A modified U-Net, a form of FCN, was used to segment drilling-induced fractures (DIFs) from an image log. The U-Net algorithm was trained with borehole resistivity images (feature) against manually labeled image logs using two datasets (original and augmented where the mirror image of the original dataset was also included). A balanced cross-entropy loss function was used because of the unbalanced label data (60× more non-fracture than fracture pixels). The results demonstrated the robustness of this U-Net model for DIF segmentation. Moreover, the model trained on the augmented dataset outperformed that trained on the original dataset (intersection over union of 0.73 vs. 0.61). Finally, this study was in accord with previous studies that showed how overlapping pixels improved predictions.