A machine-learning approach for classifying fluvial sandbody types from vertical facies sequences using geological analogues

Machine-learning tools exist for classifying sedimentary units according to objective and verifiable labels (e.g., lithological types), but the role of geologists in interpretive facies analyses is considered irreplaceable. This study presents a machine-learning method for automating interpretations of fluvial sandbodies following an approach that mirrors the application of classic one-dimensional facies models presented as idealized vertical facies sequences. Ensemble decision-tree models were trained on descriptors of facies sequences (sandbody thickness, average facies thickness, facies proportion, and vertical thickness trend), documented in many geological analogues studied by different research groups, reflecting global sedimentological thinking. The data were synthesized in two alternative datasets of contrasting dimensionality (50 vs 26 features). Up to 1222 facies sequences were used for model training and testing, with data splits operated in different ways: (i) using a partition algorithm, with options to exclude data from two-dimensional outcrop panels and to prevent the same sedimentary bodies from occurring in both splits; (ii) by arbitrarily picking sandbodies that are thought to be easily misinterpreted. Four ensemble machine-learning models were applied to perform binary classifications of (i) fluvial sandbody types according to their general channel or overbank origin, and (ii) channel sandbody types according to the interpreted planform style of their formative rivers (‘low-sinuosity or braided’ vs ‘meandering’). Across all training approaches, the models for general sandbody classification yield accuracy values ranging between 0.76 and 0.87: on average only 16% of sandbodies are misclassified. Comparable predictive power (accuracy: 0.75–0.91) is seen for models classifying formative river patterns from channel-body facies sequences, a result that challenges the much-cited notion that interpreting river planforms from vertical profiles of channel deposits is futile. A benchmark comparison against interpretations by eight sedimentologists demonstrates that cases of model misclassification are in line with errors by human geologists. The outcomes support the possibility of automating sedimentological interpretations of borehole observations (e.g., image logs) using models trained on geological analogues.