Intelligent Identification and Quantitative Characterization of Remaining Oil in Low-Permeability Reservoirs Based on a Pore-Prior and Progressive-Sampling Transformer Architecture

This study develops a Pore-Prior and Progressive-Sampling Transformer architecture, termed PPFormer, for the laboratory-scale analysis of microscopic remaining-oil images acquired from photolithographic glass-micromodel displacement experiments. The architecture integrates pore-prior embedding, progressive sampling of morphology-sensitive tokens, multi-scale self-attention encoding, relative position encoding, and boundary-enhanced decoding. PPFormer identifies five microscopic remaining-oil morphologies: cluster-like remaining oil, columnar remaining oil, droplet-like remaining oil, film-like remaining oil, and blind-end remaining oil. Under the investigated experimental conditions, the model achieved an overall pixel accuracy of 93.6%. The resulting morphology identification maps were used for pore-space-normalized area characterization and displacement-efficiency analysis under three permeability conditions and four displacement strategies. Relative to conventional waterflooding, the area-reduction ranges of cluster-like remaining oil, columnar remaining oil, and droplet-like remaining oil were from 2.29% to 12.66%, from −0.46% to 21.86%, and from 0.09% to 10.75%, respectively. Film-like remaining oil and blind-end remaining oil exhibited smaller changes, ranging from −0.50% to 8.19% and from −0.59% to 5.39%, respectively. Uncertainty was evaluated across independent replicate runs and by comparing predicted masks with consensus ground-truth masks.