10:40 AM - 11:00 AM
[DX-01] Machine Learning application in creating a model for property prediction in area with limited well data.
A machine learning project has been conducted in effort as part of the workflow to search for a suitable site for carbon capture sequestration. The main objective of this study is to acquire a high confidence earth model without incurring additional costs acquiring and interpreting new data. Hence, a data science approach was taken to map the subsurface distribution properties from vintage seismic data within a short time frame, without going through conventional seismic interpretation and earth model building.
A deep-learning model was trained using 8 basic seismic attributes, correlating the amplitude, frequency, and phase, to predict various petrophysical properties, observed in coincident well data. Once satisfactorily performant, the resulting machine learning model was used to predict petrophysical properties within the entire seismic volume. A major challenge in this task was to handle the variations in amplitude, phase and frequency of the acoustic wave due to differences in acquisition and processing parameters. Various machine learning techniques such as batch normalization, feature normalization, scaling, and regularization. were used to handle these scenarios. The trained machine learning model when used for an area with similar geology performed well, achieving the R2 score at about 0.93. For new areas with different lithologies and facies, the model was retrained with new data. A workflow had also been developed to balance the seismic cube from different regions to ensure effective prediction. Through continuous training and retraining of the machine learning model with data covering various subsurface environments, a generic and accurate ML model can be obtained for nationwide prediction.
(Keyword: Machine Learning, Carbon capture sequestration, Earth property model, Property prediction, Seismic attributes)
A deep-learning model was trained using 8 basic seismic attributes, correlating the amplitude, frequency, and phase, to predict various petrophysical properties, observed in coincident well data. Once satisfactorily performant, the resulting machine learning model was used to predict petrophysical properties within the entire seismic volume. A major challenge in this task was to handle the variations in amplitude, phase and frequency of the acoustic wave due to differences in acquisition and processing parameters. Various machine learning techniques such as batch normalization, feature normalization, scaling, and regularization. were used to handle these scenarios. The trained machine learning model when used for an area with similar geology performed well, achieving the R2 score at about 0.93. For new areas with different lithologies and facies, the model was retrained with new data. A workflow had also been developed to balance the seismic cube from different regions to ensure effective prediction. Through continuous training and retraining of the machine learning model with data covering various subsurface environments, a generic and accurate ML model can be obtained for nationwide prediction.
(Keyword: Machine Learning, Carbon capture sequestration, Earth property model, Property prediction, Seismic attributes)
Abstract password authentication.
The password was sent to the registered participants by e-mail in advance.