Although many machine learning (ML) techniques were created in the last century, only recently have the advancements in computer hardware facilitated the widespread adoption of ML in the geosciences. In seismic interpretation, geoscientists are limited by the number of attributes displayed at once (three or four with co-rendering). In contrast, ML methods are not limited by the number of attributes analyzed simultaneously and can aid and accelerate the interpretation of seismic facies. Oftentimes, ML results depict subtleties or architectural elements overlooked by the interpreter. Ideally, ML results are calibrated against the ground truth provided by well logs. When well logs are not available to calibrate the response, the resulting ML images or clusters rely on the geoscientist’s experience and an in-context interpretation. Thus, we apply a method to quantitatively assess unsupervised ML algorithms, using synthetic seismic data as a benchmark. We demonstrate that 1) by evaluating seismic attributes a priori, we can use meaningful, non-redundant attributes, which improve ML results, 2) synthetic seismic data allows us to gain valuable insights through the experimentation with different unsupervised ML techniques, 3) we can use the difference between the predicted results and the original data to quantify the uncertainty in the facies definition, and 4) our method provides confidence for the interpretation of seismic facies when other data are not available, which allows for better quality in the data used for reservoir modeling and overall reservoir assessment.

This content is only available via PDF.
You can access this article if you purchase or spend a download.