Production from multi-well and multi-bench unconventional development is time-consuming to model in physics-based simulators and requires multiple runs. Even with the use of high-performance computing or cloud computing, each single run can take several minutes to few hours depending on the model complexity. This challenges the development planning optimization as it is very computationally demanding and almost impractical to perform full subsurface uncertainty and multiple scenario realizations. The objective of this paper is to showcase the use of advanced deep-learning algorithms and AI technology developments to accelerate the calculations by several orders of magnitude while preserving both the accuracy and the physical trends.

Firstly, a large dataset of high-fidelity physics based simulations was created, using modern high-performance computing infrastructure. The input ranges for the simulations were selected based on the geology and reservoir properties of different benches within major unconventional oil plays. We also varied the placement of wells within different benches as well as completion designs to cover a wide range of modern practical operations. We then utilized the Latin-hypercube sampling (LHS) method to generate all the samples for this high-dimensional input problem (100+ parameters). We also developed query tools to post-process the simulation results and assemble the data in a manner that is readily usable by AI algorithms. Secondly, a surrogate (proxy) model was trained using modern deep-learning algorithms. We utilized several error metrics to evaluate and compare the accuracy of the proxy models. Finally, we fully automated the process of comparing the predictions of the trained proxy model on new cases against the outputs from actual simulations for the same cases.

Reservoir simulation is highly mature and complex in terms of both implemented physics as well as the numerics used to solve the governing partial differential equations. Production from typical multi-well and multi-bench unconventional developments is a spatio-temporal problem and highly dynamic in nature. The field of machine learning/deep learning is rapidly evolving and showing significant promise and value across a wide range of applications. For this specific application, we concluded that it is very important to have a larger training dataset for the deep learning algorithms to meaningfully learn all the highly non-linear input-output relationships. Additionally, we learned that enforcing physical relationships in deep learning algorithms is critical to obtain the correct physical trends from the proxy model predictions. We found that the proxy model predictions accurately match the trends as well as magnitudes when compared to computationally expensive, high-fidelity numerical simulations across several real examples of multi-well and multi-bench developments.

A lightning-fast reservoir proxy model significantly reduces the cycle-time for using physics-based models and workflows and captures subsurface uncertainty more holistically. The proxy model workflow benefits from standard features of machine learning systems, including interpretability and confidence scores that provide the user with richer information prior to deciding. Therefore, the physics-based proxy model is a powerful addition to an engineer's toolkit who is involved in optimizing unconventional development. Nonetheless, results of the AI-based proxy model need to be used diligently. The significant speed-up does come with minor accuracy degradation, and so model outputs require more engineering judgement than traditional solvers that are more expensive to run. This can be addressed by adopting a hybrid workflow (e.g., using the proxy model to narrow down the design space and then using the physics-based model to validate the final decision).

You can access this article if you purchase or spend a download.