An integrated unsupervised–supervised learning framework for enhanced petrophysical prediction

Well logging remains a cornerstone of hydrocarbon reservoir evaluation, yet conventional inversion techniques struggle to accurately characterize complex, heterogeneous formations. Although machine learning offers promising alternatives, our benchmarking of three dominant paradigms—single-model frameworks, single-level ensembles, and dual-level heterogeneous ensembles—reveals persistent performance ceilings due to limited representational capacity and reliance on scarce labeled data. Moreover, existing approaches largely overlook the abundant but unlabeled well logs available in field archives, leading to significant underutilization of geological information. To address these limitations, we propose a novel ensemble framework that integrates unsupervised and supervised learning paradigms for well-log analysis. Specifically, we employ contrastive learning to pretrain representations from 1480 unlabeled well logs, which are subsequently fine-tuned using 194 labeled total organic carbon (TOC) samples. Applied to the NY1 well in the Jiyang Depression (China), our two-stage ensemble demonstrates superior TOC prediction accuracy and generalization compared to all baseline and supervised-only models. This result highlights the value of leveraging unlabeled data to enhance petrophysical inversion, particularly in label-scarce environments. The proposed framework offers a scalable, data-efficient solution for intelligent reservoir characterization and provides critical support for unconventional hydrocarbon exploration.