Systems, methods, and software for modeling subterranean formation permeability are described. In some aspects, a method of training a subterranean formation permeability model based on NMR data includes accessing relaxation-time distributions generated from NMR measurements associated with a subterranean region. Multiple sets of principal components are generated from the relaxation-time distributions. Each set of principal components represents a respective one of the relaxation-time distributions. Parameters for weighted radial basis functions are computed based on the sets of principal components. A subterranean formation permeability model that includes the weighted radial basis functions and the computed parameters is produced.