At a meeting of the SAID, the French chapter of the Society of Professional Well Log Analysts, Emmanuel Caroli* described a trial of machine learning with neural nets to see if it would be possible to screen a large set of well logs in a data room context. In a major acquisition there may be thousands of logs available which precludes a ‘classical deterministic approach’ to formation evaluation. Total tried ‘deep forward’ and ‘deep convolutional’ neural nets in a variety of geological facies. Training was performed on a minimal log suite (GR, Neutron, resistivity, density) against interpretations of poro-perm, water saturation and clay volume. The results were complex but interesting. It emerged that the neural nets performed better when left to their own devices. Separating different facies for training was not successful. It’s better to use all the data and let the machine sort things out. Overall the best ML-driven interpretations were good, with only 5% errors. But there are a few enigmas. The machine can produce spurious physically impossible results (volumes add up to over 100%). When constraints are added to mitigate such aberrations, the error rate rises. But the results were deemed encouraging, particularly in the light of a very large amount of rather poor data where a massive amount of preparation would be required to perform a petrophysical analysis. ML provides a quick look interpretation and could be seen as a pre-processor for a physics-based workflow.
* With help from Quentin Groshens of France’s Supelec.
© Oil IT Journal - all rights reserved.