The back-propagation neural (BPN) network was proposed to model the relationship between the parameters of the dieless drawing process and the microstructures of the QSi3-1 silicon bronze alloy. Combined with image processing techniques, grain sizes and grain-boundary morphologies were respectively determined by the quantitative metallographic method and the fractal theory. The outcomes obtained show that the deformed microstructures exhibit typical fractal features, and the boundaries can be characterized quantitatively by fractal dimensions. With the temperature of 600-800°C and the drawing speed of 0.67-1.00 mm·s?1, either a lower temperature or a higher speed will cause a smaller grain size together with an elevated fractal dimension. The developed model can be capable for forecasting the microstructure evolution with a minimum error. The average relative errors between the predicted results and the experimental values of grain size and fractal dimension are 3.9% and 0.9%, respectively. The back-propagation neural (BPN) network was proposed to model the relationship between the parameters of the dieless drawing process and the microstructures of the QSi3-1 silicon bronze alloy. Combined with image processing techniques, grain sizes and grain-boundary morphologies were determined by the quantitative metallographic method and the fractal theory. The results obtained show that the deformed microstructures exhibit typical fractal features, and the boundaries can be characterized quantitatively by fractal dimensions. With the temperature of 600-800 ° C and the drawing speed of 0.67 -1.00 mm · s -1, either a lower temperature or a higher speed will cause a smaller grain size together with an elevated fractal dimension. The developed model can be capable for forecasting the microstructure evolution with a minimum error. The average relative errors between the predicted results and the experimental values ​​of grain size and fractal dimensions are 3.9% and 0.9% respec tively.