A Physics-Constrained Machine Learning (PCML) Method for Predicting Production from Unconventional Reservoirs: An Introduction

Abstract:

Reliable prediction of well performance is one of the most important technical issues for developing unconventional resources. There are generally two prediction methods. One is physics-based method (including reservoir simulation) that can predict the long-term production based on the early production data and estimated reservoir properties. However, the physics of fluid flow in unconventional reservoirs is not fully clear at this point and the linkage between the physics and some important data types is not understood yet. Thus, the method is not able to integrate all the relevant data. The popular alternative is the machine learning method based on a rationale that the data implicitly contains the relevant physics. The machine learning method, however, does need a significant amount of data that are not always available, especially for newly developed unconventional resources. This communication will briefly review the application of the machine learning methods on predicting hydrocarbon production from unconventional reservoirs in the literature and present the concept and general methodology of an innovative physics-constrained machine learning method (PCML) that combines the advantages of both machine-learning and physics-based methods for the production prediction.