A Comprehensive Approach to Total Organic Carbon Prediction: Dual-Output Neural Networks and Ensemble Learning

Abstract

Abstract This paper explores a neural network-based approach for constructing prediction intervals (PIs) of total organic carbon (TOC) content. In contrast to conventional methods that focus solely on minimizing prediction error, the proposed method utilizes a dual-output neural network optimized by a novel loss function called \({\mathcal{L}}_{QCE}\) that emphasizes overall PI quality through a balanced consideration of coverage probability, interval width, and cumulative deviation. Consequently, this approach facilitates the generation of higher-quality PIs under specified significance levels. Case studies illustrate that, in comparison to prevailing techniques such as Pearce's method and Gaussian process regression, our proposed approach achieves a notable over 40% reduction in invalid intervals, accompanied by an approximate 50% improvement in interval quality. Additionally, we introduce ensemble learning to assess inherent model uncertainties, further augmenting the precision of PIs. In summary, the presented methodology offers a competitive solution for uncertainty quantification and well log data mining, providing an innovative and effective approach to enhance the quality of PIs for TOC content.