Experiments with olivine and a basaltic melt were conducted to analyze the melt distribution in partially molten aggregates at low melt fractions. Grain size and melt distribution at the start of an experiment are transient; to reach steady state conditions requires run durations from 2 to 3 weeks at 1 GPa and 1300¿ to 1400 ¿C. Quantitative analysis of the melt distribution from backscattered electron images of postrun samples shows that most of the melt resides in low aspect ratio, disk-shaped inclusions on two-grain boundaries at melt contents from 0.8 to 3.3 vol. %. Tubules along three-grain edges most likely interconnect the melt at all melt fractions, but their shape and size varies even at the lowest melt fractions due to the effects of anisotropy and nonuniform grain sizes. These triple junction tubules can therefore not be approximated by uniform, equally spaced cylinders with near constant cross-sectional area for the purpose of calculating the permeability of the matrix. Moreover, the tubules are much smaller than the disk-shaped inclusions; tubule-like geometries contain only ~10% of the total melt content in each experiment. The permeability of the triple junction tubule network only is therefore low (k~10-17 m2), and segregation velocities are less than 1 mm yr-1. The permeability of the aggregate increases substantially only after the disk-shaped inclusions become interconnected. This melt fraction is calculated from percolation theory, using parameters of the melt inclusions obtained from the backscattered electron images. The threshold melt content, as calculated from the experimental melt distribution, lies between 2 and 3 vol. %.¿ 1997 American Geophysical Union