Understanding uncertainty in deepwater channel seismic facies classification applying random forest on outcrop-constrained 3D models and synthetic seismic data

Abstract

To understand the uncertainties in seismic interpretation, especially for deepwater seismic facies, we apply a novel approach using outcrop-based synthetic data as ground truth. By applying the random forest (RF) supervised machine learning method we can better understand the influence of classifier hyperparameters on facies prediction accuracy. Based on previous analysis, we chose six seismic attributes that are able to differentiate five deepwater architectural facies: shale (thin-bedded turbidites), channel axis, off-axis, margin, and mass transport deposit. Hyperparameter testing indicated that optimization of the RF classifier is sensitive to the (1) choice of training attributes, (2) original facies proportions, (3) similarity in the seismic expression of different facies, and (4) seismic data resolution and seismic data signal-to-noise ratio. A simple classifier using common RF hyperparameters developed for fluid saturation predicted the facies with 74% accuracy. Although computationally more expensive, optimizing the RF hyperparameters provided approximately 89% accuracy, demonstrating the importance of hyperparameter tuning. When applying the best model to an unseen portion of the model, the position of the channel complex set was predicted accurately, but the accuracy was only 58%. This highlights the limitations of universal models and the persistent uncertainties faced when using ML to enhance our interpretations. Although we use a correlation matrix approach, we acknowledge that uncertainties can only be assessed and not quantified. Optimization of attribute choice and hyperparameter tuning reduce the epistemic uncertainty. We encourage future research to explore this in more detail.