<p>Data-driven permeability prediction models for porous media facilitate rapid reservoir evaluation; however, heterogeneity, anisotropy, and isoparametric heterogeneity significantly influence prediction accuracy. Utilizing numerical simulation for petrophyscial parameters, followed by the lattice Boltzmann method (LBM) simulation and statistical models, we recognize incomplete multicollinearity as a significant factor confounding correlations among pore structure parameters in prior research. Isoparametric heterogeneity may cause both linear and nonlinear, as well as single and multi-parameter empirical permeability equations, to be ineffective under varying conditions. The absence of a controllable generation model has impeded the examination of permeability uncertainty in cores exhibiting equivalent petrophysical properties. We propose a Multi-Condition Rock Generative Adversarial Network (MCRGAN) model that regulates different petrophysical parameters of generated samples and is suitable for diverse rock types. Employing MCRGAN, we quantitatively assess permeability variations in cores with identical characteristics. Despite equivalent low-order moments like porosity, three-directional permeability means may vary by as much as 50%, but data-driven model inaccuracies might span from 3 to 78% of the real permeability. As porosity increases, permeability range ratios markedly grow, surpassing 10,000%, signifying a considerable augmentation of uncertainty. Our results indicate that aside from reservoir heterogeneity, isoparametric heterogeneity in pore space represents a substantial source of inaccuracy in permeability forecasts utilizing data-driven approaches such as the Kozeny–Carman (KC) equation. Moreover, the MCRGAN model provides significant flexibility, allowing researchers to create MCRGAN + LBM methodologies to examine microscopic control mechanisms of various physical properties in rocks.</p>