Petrophysical Open Hole Log Evaluation Using Artificial Intelligence in a Sandstone Gas Field Offshore UK

Abstract In the context of a Carbon Capture and Storage project requiring the petrophysical re-evaluation of approximately 200 well penetrations in a mature offshore UK giant gas field, we assessed the suitability of artificial intelligence (AI) to predict volume of clay, porosity and water saturation. The reservoir is a relatively homogeneous gas-bearing sandstone with porosity varying from 10 to 15% and low clay content. The selected AI model was a supervised learning approach (eXtreme Gradient Boosting algorithm), with Root Mean Square Error (RMSE) used as a model performance indicator. Dataset cleaning and preparation, including selecting the input logs and wells, was paramount to model performance. In the study, 181 wells were used, as they presented at least 3 open hole logs in the reservoir section, Gamma Ray, Density and Resistivity, to be used as inputs to the AI. A petrophysical evaluation of volume of clay, porosity and water saturation was also available. 80% of the selected wells (144 out of 181) were used for AI model training, leaving the remaining 20% for model validation. The initial baseline trained model exhibited a RMSE of 0.62. A significant increase of prediction performance (RMSE=0.2) was observed with dataset augmentation: artificial increase of diversity and size of dataset by data windowing, local gradient and second order interaction. It was noticed that all input parameters do not have the same influence on the prediction performance with GR log having the biggest influence. The optimised trained AI model was applied on the validation dataset (37 remaining wells). Outputs were compared to the manual log evaluation and the difference was low and considered acceptable, therefore validating the model. The AI model was deployed on two fields offshore UK in the same geologic setting but with poorer reservoir properties to test its suitability on a regional scale in a blind test. A decreased prediction performance was observed, therefore highlighting the importance of diversifying the composition of the training dataset, in order to make models more robust to regional variations. As a conclusion after data processing and optimisation, the AI model provided meaningful and corroborated outputs. Sensitivity tests to the number of wells in the training set were also carried out and it was found that the AI model only required 25 wells for training, in combination with dataset augmentation, to predict with confidence the reservoir properties. A potential use case is AI prediction in data room contexts for quick evaluations.