While the traditional processing and interpretation workflows are subjective, inconsistent upon the expertise of Geoscientists and slow in turning around the deliverables, machine learning requires
a large amount of data—either depth or time samples—to effectively span measurement space and
a high number of measurements to deduce a representative, low-dimensional feature set.
The two requirements of machine learning are not generally available in well log data, making the application of machine learning to wellbore data processing and interpretation quite limited.
We proposed a novel Class-based Machine Learning (CbML) approach that alleviates the limitations of machine learning by first reducing training data into a few explainable classes, followed by learning models per class. For new data, the probabilities of a data point belonging to existing classes are computed, and the data point is assigned to the most probable class. Finally, the learnt models per class are applied, and uncertainties are estimated.
The CbML approach acquires knowledge from the training data and propagates, if and where applicable, to the new data. It eliminates the need for large training data and a high number of measurements. In addition, it not only removes the subjectivity and inconsistency but also substantially improves the turnaround time from the receipt of data to the delivery of results. The approach serves as a continuous learning, extraction, and application loop automating the processing and interpretation of wellbore data.
The proposed CbML approach combines the advantage of both traditional petrophysical workflows and machine learning. It provides objective, consistent, and near-instant answers with minimal intervention.
Over the past decade, oil and gas industry has come to see "Machine Learning" as the next evolutionary step to help work better, safer, and smarter. For example, more than 1000 papers related to applications of machine learning and artificial intelligence were published in 2018 alone. The advancement spans a broad range of oil and gas domains such as continuous monitoring, seismic reservoir characterization, drilling data interpretation, and Geostatistics. Petrophysics is one domain where some sporadic successes have been reported, but a generalized machine learning-based framework for well log data processing and interpretation is still in its infancy. It is not yet sufficiently proven that machine learning techniques are superior to domain experts in capturing critical information and using such information at the scale of basins or across the complexity of multi-domain workflows. However, from successful use cases, it is becoming clearer that machine learning can deliver insights more efficiently, objectively, and consistently.
