<![CDATA[Machine Learning (ML) offers an innovative approach to reservoir characterization, yet its widespread application is hindered by poor cross-basin generalization. Specifically, most models validated on intra-basin data fail to account for the geological heterogeneity encountered in new settings. This study addresses this limitation by quantifying the geological domain shift that hinders direct model transferability and by proposing a conceptual framework to overcome it. This approach is designed to systematically address geological variability by identifying local electrofacies structures before attempting predictive modelling. To validate the necessity of this framework, we conducted a preliminary empirical study comparing the Talang Akar Formation in the North West Java Basin and the Menggala Formation in the Central Sumatera Basin (Basins A and B). This study empirically demonstrates that, despite shared fluvial-deltaic depositional environments, the intrinsic statistical structures of the two basins diverge significantly. A quantitative analysis revealed a severe domain shift, evidenced by a large Euclidean distance between cluster centroids and a low adjusted Rand index (ARI) of 0.113 when applying direct analog mapping. These findings empirically demonstrate that direct model transfer is ineffective because of second-order geological controls. Consequently, this study establishes the critical need for the proposed UL-SL strategy to adaptively handle domain shifts, providing a geologically grounded roadmap for accurate characterization in frontier and data-scarce basins.]]>
