Dr. Mike Brhlik
Staff Geophysicist, ConocoPhillips
Describing the Reservoir: Seismic Machine Learning and Data Analytics
Quantitative seismic reservoir characterization poses a mathematically ill-constrained inversion problem traditionally solved by methods relying on pre-stack seismic inversion and subsequent rock physics transforms. Alternatively, subsurface models can be matched to field seismic data by seismic forward modeling using wells as calibration points. Both these approaches face practical limitations in sparsity of calibration data and severe non-linearity of the problem requiring multiple simplifying assumptions. Recent extensive developments in machine learning and data-driven model building can provide significant accuracy and efficiency uplift in solving this problem by streamlining seismic attribute analysis and avoiding the need to pass through the elastic domain. We present various approaches to seismic machine learning and their application to both static and dynamic reservoir characterization projects and discuss comparisons to conventional 3D and 4D quantitative interpretation workflows. Emphasis will be given to practical approaches enhancing cross-discipline integration and validation of data analytics methods using both geophysical and data science approaches. We highlight the advantages, challenges and systematic biases encountered in this type of analysis and discuss potential extensions of the data analytics approach using deep learning methods.