Soma Budai

The first scientifically described dinosaur fossil from Transylvania is a quasi-completely preserved but crushed hadrosauroid skull (NHMUK R.3386), discovered by Ilona Nopcsa in 1895 along the Sibi...

The recognition of clastic parasequences is hampered by vagueness in their definition and rendered difficult by variability in their origin and temporal and spatial expressions. This makes the automation of parasequence identification in subsurface data particularly challenging. This study presents a supervised machine-learning approach for the recognition of parasequence boundaries (flooding surfaces) in borehole data. The approach enabled the classification of facies logs using a training data set consisting of over 9000 facies units in 169 shallow-water clastic parasequences, from both outcrop and subsurface geologic analogs. Data on the thickness, grain size, and vertical stacking trends of sedimentary facies were collated to produce three alternative training data sets based on vertical windows (facies sequences) spanning a variable number of facies units (two, three, and five). Eight alternative machine-learning algorithms for supervised classification were employed to perform picking of parasequence tops, a task that was treated as a binary classification problem. The training of the models was performed by optimizing their ability to detect positive instances; i.e., portions of facies sequences that were correctly identified as occurring at parasequence tops. Model evaluation against a data set reserved for testing revealed that ensemble machine-learning algorithms were able to pick flooding surfaces that matched those recognized by stratigraphers in up to 39% of the cases. Generally, sensible classifications were made by the best performing models where evident thickness and grain-size signatures of flooding surfaces existed in facies sequences. Because of the aforementioned subjectivity in parasequence definition, model training and performance assessment were not straightforward. Nevertheless, promising aspects of this approach include: (1) its applicability to ensemble machine-learning algorithms; (2) the ability to train classifiers on outcrop analogs; (3) the capacity to compile different training data sets reflecting different schools of thought about parasequence definition and hierarchical significance; and (4) its possible integration with computational methods for classifying lithologies in borehole data and for correlating surfaces across wells.