Abstract
This study aims to demonstrate that EURs in unconventional plays can be accurately predicted utilizing a neural network trained on petrophysical, engineering, and well design data. Additionally, we aim to demonstrate that probabilistic petrophysical workflows can be automated, deployed at the basin scale, and deliver better results at scale than traditional workflows.
The results from the EUR prediction model show that on an individual well basis EURs from the three formations can be predicted within 38-56% of the DCA-derived results in the test data set. This range is tighter than the estimated range between the P90 and P10 results from a multi-well type curve analysis. Furthermore, when the results from any formation are taken in total, the predictions showed significantly better accuracy with cumulative predicted volumes within 1.4-8.7% of the DCA-derived results. This demonstrates the efficacy of the model for both single-well predictions and on a cumulative basis.
The results show that both single-well precision and overall precision are degraded when using deterministic or cutoffs-based workflows for reservoir characterization. This suggests that there is a significant value add associated with performing a more advanced petrophysical workflow utilizing a probabilistic petrophysical workflow.
This work also demonstrates that probabilistic petrophysical models can be deployed at scale and extensively automated, with model run time of 1 minute/100 wells, therefore unlocking the ability of petrophysicists to better integrate best-in-class methods into geoscience workflows.
We predict well performance from the Powder River Basin using automated interpretation of public well logs and production data. We first determine the estimated ultimate recovery (EUR) of 690 horizontal wells using an automated decline curve analysis (DCA). We then construct a series of geological and petrophysical property maps using results from a mineral inversion performed on data from 682 vertical wells. Finally, we train a performance prediction model based on the petrophysical results and available well engineering data. We validate the model via comparisons of performance predictions to DCA results.
