Sequential Multi-Realization Probabilistic Interpretation of Well Logs and Geological Prediction by a Deep-Learning Method

Abstract

The majority of geosteering operations rely on traditional shallow sensing logging tools as sources of information. Many such operations rely on stratigraphic-based steering when the logs from the drilled well are matched to logs from an offset well by modifying the lateral shape of stratigraphy. The match of the logs indicates a plausible interpretation, but due to the scarcity of log data in many situations, this interpretation is not unique. In manual workflows maintaining several likely interpretations is not realistic and in automated workflows, multiple interpretations are seldom used. We describe a deep neural network (DNN) that outputs a selected number of stratigraphic interpretations using a single evaluation of the input log data in two milliseconds. The input data defined prior to training consists of one or several log pairs consisting of one current lateral and one offset-well log. For each of the interpretations, the DNN also estimates the respective probability and can be configured to produce likely ahead-of-data predictions of the geology, which are based on the data mismatches and the likelihood of geological configurations with respect to the training dataset. The described probabilistic interpretation and prediction is enabled by the supervised training of a mixture density DNN (MDN) with a stable multiple-trajectory-prediction loss function. In this paper, we apply the MDN for the sequential interpretation of well logs. We use the interpretations and the probabilities from the previous interpretation step as starting points for the probabilistic interpretations and predictions for the current step. We avoid the curse of dimensionality by discarding the unlikely starting points. The batchable MDN evaluation enables tracking of hundreds of solutions while still maintaining sub-second performance, compared to minute(s) reported in other recent papers. The performance of the method is verified on synthetic test data as well as the realistic well data from the Geosteering World Cup 2020 (based on the Middle Woodford formation, located in the South Central Oklahoma Oil Province in the United States) and stratigraphic configurations provided by geologists. In all cases, the method manages to capture likely interpretations. At the same time, the accuracy of predictions deteriorates for the configurations which were not typical for the training dataset.

Direct Multi‐Modal Inversion of Geophysical Logs Using Deep Learning

“…When trained on noisy data, the presented methodology can handle realistic log‐noise levels. Thus the model's sequential version (Alyaev et al., 2022) can be directly used for automatic interpretation during traditional shallow‐log‐based geosteering with fixed‐thickness stratigraphy and possible faults. We expect the approach to naturally extend to other geological discontinuities such as pinch‐outs if retrained on a representative dataset.…”