Permeability prediction in tight sandstone reservoirs is strongly influenced by diagenesis, multiscale pore structures and multiphase flow effects, resulting in pronounced spatial heterogeneity. Traditional empirical models based on homogenization assumptions suffer from inherent physical limitations, making their predictive accuracy insufficient for practical applications. Although existing data-driven deep learning methods offer advantages in nonlinear modeling, their performance is constrained by few-shot samples, imbalanced distributions, low-density information and the absence of physical constraints, which hinder the extraction of effective representations of inter-features and lead to biased learning and physical deviations. To address these challenges, this study proposes a framework including a data enhancement strategy based on high-order feature construction for quality improving and adversarial generation for quantity expanding, and a deep learning model termed DAFM-PI, which integrates bidirectional attention fusion with physical information constraints. Experimental results show that the test accuracy R 2 of DAFM-PI improves from 0.444 before enhancement to 0.969 after that, representing a 52.2% improvement. Furthermore, compared with conventional POR-LR, ResNet, and Transformer Encoder models, DAFM-PI achieves accuracy improvements of 61.9%, 18.1%, and 6.2%, demonstrating its effectiveness in enhancing both prediction accuracy and generalization for permeability prediction in tight sandstone reservoirs. • Propose a collaborative, low-cost and physics-informed framework in permeability prediction of tight sandstone reservoirs. • Overcome few-shot and imbalance limitations by data enhancement and model optimization. • Enhance data representation through high-order feature construction and generative sample augmentation. • Develop a neural network termed DAFM-PI with dual attention fusion and physics-informed constraints. • Achieve superior accuracy and generalization compared to other baseline models.
Pang et al. (Wed,) studied this question.