The frictional properties of a layer of simulated Westerly granite fault gouge sandwiched between sliding blocks of Westerly granite have been measured in a high-speed servo-controlled double-direct shear apparatus. Most gouge layers were prepared to have a self-similar particle distribution with a fractal dimension of 2.6. The upper fractal limit was varied between 45 and 710 ??m. Some gouges were prepared with all particles in the range between 360 and 710 ??m. In each experiment the sliding velocity was cyclically alternated between 1 and 10 ??ms-1 and the coefficient of friction ??m and its transient parameters a, b and Dc were measured as functions of displacement. In addition to the particle size distribution, the following experimental variables were also investigated: the layer thickness (1 and 3 mm), the roughness of the sliding surfaces (Nos 60 and 600 grit) and the normal stress (10 and 25 MPa). Some of the sample assemblies were epoxy impregnated following a run so the gouge structure could be microscopically examined in thin section. We observed that gouges which were initially non-fractal evolved to a fractal distribution with dimension 2.6. Gouges which had an initial fractal distribution remained fractal. When the sliding blocks had smooth surfaces, the coefficient of friction was relatively low and was independent of the particle distribution. In these cases, strong velocity weakening was observed throughout the experiment and the transient parameters a, b and Dc, remained almost constant. When the sliding blocks had rough surfaces, the coefficient of friction was larger and more dependent on the particle distribution. Velocity strengthening was observed initially but evolved to velocity weakening with increased sliding displacement. All three transient parameters changed with increasing displacement. The a and b values were about three times as large for rough surfaces as for smooth. The characteristic displacement Dc was not sensitive to surface roughness but was the only transient parameter which was sensitive to the normal stress. For the case of rough surfaces, the coefficient of friction of the 1 mm thick gouge was significantly larger than that for the 3 mm thick layers. Many of these observations can be explained by a micromechanical model in which the stress in the gouge layer is heterogeneous. The applied normal and shear stresses are supported by 'grain bridges' which span the layer and which are continually forming and failing. In this model, the frictional properties of the gouge are largely determined by the dominant failure mode of the bridging structures. ?? 1989.