Four different types of baffle arrays were inserted in an unwound channel of a single screw extruder to generate chaotic mixing. The periodic unit of the flow channel was modeled as a dynamic system of complex cavity flow. The finite volume method was used to solve the three-dimensional flow of a purely viscous non-Newtonian fluid obeying the power-law constitutive model. Lagrangian particle calculations along with statistical methods were performed by a fourth-order Runge-Kutta scheme. The effect of the baffle's array mode and characteristic length on the mixing kinematics was investigated numerically. Poincare sections and period points were applied to reveal the different patterns and sizes of the KAM tori. Distributive mixing was visualized by the evolution of passive tracer particles initially located at different positions. The variance index and residence time distribution (RTD) were used to evaluate the statistical results. Compared with case B, where two baffles were arranged side by side in each baffled zone, the staggered alternative modes in cases C and D, where only one baffle was used in each baffled zone, surprisingly produced better mixing. Moreover, the characteristic length was also one of the key factors that influenced the mixing. When the characteristic length equaled 10 mm, case C had better mixing than case D, whereas, when the characteristic length equaled 7.5 mm, the result was reversed and case D-2 had the best mixing, while no obvious KAM islands or period tori were found. (C) 2015 Elsevier Ltd. All rights reserved.