Automating interpretations of fluvial facies sequences: machine-learning models based on geologic analogs

Machine learning is increasingly used for classifying sedimentary deposits according to objective lithological categories. However, facies analysis of one-dimensional facies sequences requires considerable interpretation efforts. A novel method is presented for automating interpretations of reservoir-forming fluvial sandbodies following an approach that mimics the application of facies models presented as idealized vertical facies sequences. Machine-learning models are trained on attributes of facies sequences (sandbody thickness, and facies-type mean thickness, proportion and vertical thickness trends) documented in a global compilation of many geologic analogs. Over 1200 facies sequences are used for model training and testing, with data splits produced in two ways: (i) according to 80–20% training versus testing divisions; (ii) by arbitrarily selecting test sandbodies that are deemed difficult to interpret. Alternative training datasets additionally reflect options to exclude data from two-dimensional outcrop panels, and to prevent sedimentary bodies from occurring in both splits. Four types of ensemble machine-learning models are trained to operate binary classifications of (i) fluvial sandbody types according to their channel versus overbank origin, and (ii) channel sandbody types according to interpreted planform styles of formative rivers (low-sinuosity or braided vs meandering). Across all model-training approaches, models for general sandbody classification exhibit accuracy ranging between 0.76 and 0.87: only 16% of sandbodies are misclassified, on average. Models classifying formative river patterns from channel-body facies sequences—a task that is commonly considered futile even when attempted by expert sedimentologists—exhibit similar predictive power (accuracy: 0.75–0.91). Manually selected sandbodies that are thought to be prone to misinterpretation have been employed to perform a comparison of model performance against interpretations by expert sedimentologists; results demonstrate that model misclassification is in line with human errors. The proposed approach needs refining but holds promise as a way to automate sedimentological interpretation of subsurface datasets documenting sandbody facies sequences (e.g., FMI logs).