Downhole Fluid Characterization with Validated Mapping from Optical Sensor Measurements

Abstract

Abstract Robust downhole fluid analysis highly relies on quality measurements and data mapping from various optical sensors implemented into modern formation testers. The direct modeling used to determine a multivariate correlation between optical sensor responses and diverse fluid compositions or properties is often cost-prohibitive with sensor-based calibration. This paper presents a novel method based on concatenated optical computing neural networks (COCN), which link sensor-specific signal transformation to generic fluid characterization through validated data mapping. The COCN models are built separately in the in-house laboratory using optical data transformation networks and multi-analyte fluid characterization networks. To evaluate the uncertainty of sensor signal mapping, additional sensor data prediction networks, which exchange inputs and outputs of sensor data transformation networks to provide reverse data mapping, can be built to produce the simulated sensor responses. The actual sensor measurements are then compared to the simulated sensor responses to validate the optical signal transformation. The degree of matching can be used to identify issues associated with the quality of the sensor measurements and the range of calibration and testing data, and to select the transformation networks in ruggedizing predictions of fluid compositions and properties. This paper presents the method used and case studies to exemplify the application of the proposed approach for reliable fluid sample characterization using laboratory and field measurement data. Calibrating sensor-independent fluid predictive networks on a large database usually pertains to low uncertainty because the analyzed measurements can be collected from diverse reservoir fluid samples. The construction of data transformation networks, however, is only realistic by using a small number of reference fluids because of their high calibration cost, which conversely challenges the quality of data mapping with new field sensor measurements. In this study, sensor data transformation networks and sensor data prediction networks form two groups of mutually complementing models that can validate one another, making the quality of data mapping observable before applying fluid characterization models. The results of the case studies demonstrate that the consistency between the actual and simulated sensor responses is a reasonable performance index in the prediction of fluid compositions and properties. The laboratory and field examples also justified the importance of using simulation data appropriately to overcome the limitation of sensor measurement data in the calibration of transformation networks.