Deep hierarchical distillation proxy-oil modeling for heterogeneous carbonate reservoirs

This paper presents a novel few-shot proxy modeling approach for the oil and gas industry to reduce reliance on numerical simulators for reservoir analysis. The strategy introduces a regression framework leveraging deep hierarchical self-distillation to construct a meta-model based on ensemble learning. The proposed method employs a cascade training scheme, which uses predictions derived from a superior hierarchical level to distill knowledge into the next predictor. The pivotal idea of self-distillation is to generate "soft targets" rather than hard ones. Soft targets represent probability distributions over potential output curves rather than a correct answer. This smoothing information provides additional guidance to the model during training, helping it generalize better. This architecture utilizes three-dimensional (3D) maps in proxy modeling to forecast cumulative fluid production and generate risk curves indispensable for effective field decisions. The study employed data acquired from a complex oil field characterized by a substantial degree of uncertainty. The hierarchical self-distillation technique outperforms alternative methods, achieving a symmetric mean absolute percentage error (SMAPE) below 2%. It reduces computational overhead by 84% for a probabilistic model with 900 simulations and 62% for a model with 200 simulations. The developed proxy model delivers valuable insights for decision-making in oil field management, offering the potential to decrease expenses and enhance efficiency in field exploration endeavors.