Optical sensor adaptive calibration

Abstract

The subject disclosure provides for a method of optical sensor calibration implemented with neural networks through machine learning to make real-time optical fluid answer product prediction adapt to optical signal variation of synthetic and actual sensor inputs integrated from multiple sources. Downhole real-time fluid analysis can be performed by monitoring the quality of the prediction with each type of input and determining which type of input generalizes better. The processor can bypass the less robust routine and deploy the more robust routine for remainder of the data prediction. Operational sensor data can be incorporated from a particular optical tool over multiple field jobs into an updated calibration when target fluid sample compositions and properties become available. Reconstructed fluid models adapted to prior field job data, in the same geological or geographical area, can maximize the likelihood of quality prediction on future jobs and optimize regional formation sampling and testing applications.