Characterizing Downhole Fluid Analysis Sensors As Digital Twins: Lessons of the Machine Learning Approach, The Physics Approach and the Integrated Hybrid Approach

Abstract

Digital fluid sampling is a technique utilizing downhole sensors to measure formation fluid properties without collecting a physical sample. Unfortunately, sensors are prone to drift over time due to the harsh downhole environmental conditions. Therefore, constant sensor evaluation and calibration is required to ensure the quality of analysis. A new technique utilizes a virtual sensor as a digital twin which provides a calibration that can be utilized by the physical twin. Digital twin technology enables the end-user to operate and collaborate remotely, rapidly simulate different scenarios, and provide improved accuracy via enhanced up-to-date calibrations. With respect to downhole fluid identification, the contribution of harsh environmental conditions and sensor drift can also be mitigated by realizing a virtual implementation of the fluid behavior and the individual sensor components. Historically, the virtual behavior of a digital twin has been constructed by a combination of complex multi-physics and empirical modeling. More recently, access to large datasets and historical results has enabled the use of machine learning neural networks to successfully create digital twin sensors. In this paper, we explore the efficacy of constructing a digital twin on a single downhole optical fluid identification sensor using both the machine learning nonlinear neural network and the complex, multi-physics' based modeling approaches. Advantages and lessons to be learned from each individual method will be discussed in detail. In doing so, we have found a hybrid approach to be most effective in constraining the problem and preventing over-fitting while also yielding a more accurate calibration. In addition, the new hybrid digital twin evaluation and calibration method is extended to encompass an entire fleet of similar downhole sensors simultaneously. The introduction of digital twin technology is not new to the petroleum industry. Yet there is significant room for improvement in order to identify how the technology can be implemented best in order to decrease costs and improve reliability. This paper looks at two separate methods that scientists and engineers employ to enable digital twin technology and ultimately identify that a hybrid approach between machine learning and empirical physics'-based modeling prevails.