Emerson E&P has developed a supervised machine learning approach called Democratic Neural Network Association (DNNA). The method reconciles multiple datasets to predict facies away from the wellbore. It employs an ensemble of many neural networks running in parallel that simultaneously learn from the multiresolution wellbore and seismic data using different strategies and associations. This architecture minimizes the possibility of biasing. It includes a secondary training stage where seismic data are introduced away from the wellbore and voted on for training set inclusion to stabilize network training while preventing overlearning. The outcome of this process is a probabilistic facies model description of the reservoir. It predicts the most probable facies distribution and associated maximum probability as well as the probability relative to each facies. This results in less guesswork when quantifying uncertainty in rock type distribution. Results are interactively generated in a 2-D and 3-D environment for in-depth analysis and are reservoir simulation ready. The outcome is critical for reservoir geologists and engineers to better understand reservoir behavior. Once considered nice-to-have technologies, the sheer volume of well and seismic data that need to be analyzed has made machine learning an effective approach for transformation and analysis of subsurface data. Automated machine learning produces outputs in minutes or hours rather than months or years. DNNA provides a practical approach to invert directly for the desired model facies resolution and heterogeneity, including fluid overprint. The method has been demonstrated to predict lithozones in both conventional and unconventional reservoirs.
展开▼
