Physics-guided neural networks for predicting unconventional well performance — using Unconventional Rate Transient Analysis for smarter production data analytics

Abstract

Geoscience modelling poses many challenges due to the limited sampling and the complexity of the phenomena that create the resulting rock and its properties. When adding to the geologic and geomechanical complexity the various possible fluid flow mechanisms that are often not fully understood, one realizes quickly how daunting geomodelling could be. As a result, oil and gas fields are often bought and sold using reserves computed with simple decline curve analysis tools. Unfortunately, these simplified production analysis tools contain no physics and do not help develop oil and gas assets which require the knowledge of 1) rock properties distribution and 2) the impact of the rock properties on the selected production mechanism. For example, if one has a naturally fractured reservoir, the presence or absence of the natural fractures and the way the wells are drilled to encounter or avoid these rock properties will determine the reserves and the future of the company developing such an asset. Very often the future of many companies is not very bright due to their lack of knowledge of the distribution of the rock properties such as natural fractures and the subsequent fluid flow resulting from drilling and fracking into these heterogeneous rock properties. For the companies that want to prosper by intelligently developing their assets, niche companies and technologies have been used throughout the years to address all these challenges. Among these technologies we find the use of artificial intelligence and machine learning, which were introduced in the oil and gas industry in the early 1990s with successful application to petrophysics, well test analysis and reservoir modelling. Three decades later, the oil industry and its unconventional revolution is facing frequent challenges ranging from frac hits and well interferences, casing deformation and collapses, expensive cube development with meagre returns on investment and many other puzzling issues with no easy solution. Suddenly, under the stress of financial constraints, the machine learning tools criticized for three decades as ‘black boxes’ are now becoming the preferred solution to all problems. Unfortunately, this love-hate relationship the oil industry has with data-driven approaches may hit a rough patch very soon due to its inability to recognize the limitations of the available data used in the various oil and gas challenges.