A Shale heterogeneity proxy from Supervised Machine Learning?

A real seismic case is shown to validate the use of supervised machine learning to improve the seismic interpretation of unconventional (shale) seismic datasets (input seismic and well log data courtesy of FracGeo LLC). A supervised machine learning flow is trained on 3-D seismic data to extract potential hidden patterns. In the present context, those hidden patterns are visualized as anomalous features departing from a background trend.

The following figure shows at top two arbitrary sections of the input 3-D seismic full stack (all offsets collapsed) along N-S and E-W directions. A dome or anticline is observed along the E-W direction. Some well log datasets are available including petrophysical and elastic logs (Vp,Vs y density) and one of those well logs is plotted on the seismic background as a magenta vertical track representing the shale content (values increasing from left to right). Additionally, two target zones (labeled as Zone “A” and Zone “B”) around of a time interpreted horizon (on green color) are shown. Both zones have significant brittleness for some wells in this field.

At the bottom, it is shown the seismic hidden patterns (as stacked traces) using a machine learning approach along the same E-W and N-S directions. The basic idea is to detect seismic features that departure from a background trend and measure the grade of departing from statistical tools. Under a grey color scale (as shown), the lightest color represents the lowest divergence (or departing) from the background trend, but a dark color the highest divergence. Our machine learning approach is data-driven, trace-by-trace based and stochastic (or random) on generating the input weights. Even though some of those ML hidden patterns can be associated with input noisy traces, it is observed a higher population of patterns having high divergences (dark color) around the top of the dome which could have some correspondence with some 3-D conceptual models proposed for fracturing distribution in anticlines (i.e., Cooper et al., AAPG Bulleting, v 90, No. 12, 2006).

The next figure shows the horizon-based maps from Zone A (left) and Zone B (right) using an averaging statistic. The corresponding seismic amplitudes and divergences were scaled to a -1 to 1 range to make them more easily comparable. The two panels at left compares the input seismic stacked amplitudes versus the average divergence stack for Zone A. It is very evident that the new ML divergence stack adds more complexity and defines better than input seismic lateral connections among drilled producer wells as dots. Also, the ML divergence stack highlights better many discontinuities indicated in dark colors (higher divergence). The same results are obtained for the Zone B but at this level, the ML divergence stack shows more clearly a stronger connection among the higher divergence patterns maybe representing major faults in the northern area of this field. Finally, from the both divergence maps at zones A and B, it is observed that some drilled wells are located in higher divergence zones, but many other well are drilled on lower divergence zones. Without any additional information about production, it indicates that our ML divergence stack could be a proxy of the grade of heterogeneity of the shale environment in this field.

CLS Geo Solutions LLC is developing workflows to extend the functionality of Machine Learning for seismic Oil & Gas reservoir insight. Both real cases are intended to empathize that supervised machine learning can be a very valuable product in an early step of the seismic interpretation by exposing potential hidden features (anomalous amplitudes) that could rocket your final seismic interpretation. Contact us for your next ML seismic experience by web page www,clsgeosolutionsllc.com or directly emailing to nmartinw@clsgeosolutionsllc.com.