Denoising Electromagnatic Surveys Using LSTMs

Resistivity readings obtained from electromagnetic crosswell surveys provide insight for reservoir water saturation prediction. Although high resistivity values should map to low water saturation and vice versa, in many cases the readings may not be consistent with this correlation. This is due to factors that add noise to the resistivity reading, such as the borehole effect and the salinity of the injected water. Here, we attempt to treat the resistivity reading to negatively correlate with water saturation, enhancing the accuracy and interperability of water saturation prediction models. We utilize the resistivity readings from locations further from sources of noise to correct the inconsistencies in the resistivity readings using a Long-Short Term Memory (LSTM) Neural Network approach. Our results demonstrate that by addressing noisy inconsistencies in the data, the performance of the water saturation model increases in terms of R2 from 0.62 to 0.70. Moreover, upon deploying model interpretation method, namely, SHAP TreeExplainer, we show that the resistivity-based features in the water saturation prediction model posses higher importance values than before the enhancement, in comparison with porosity features.