Geometric calibration is a major step in computed tomography (CT) where it provides values for geometrical parameters that later define the system matrix in reconstructing CT images. A standard calibration process usually involves the illumination of an accurate calibration phantom with known coordinates of ball markers using the imaging system, followed by calculation of geometrical parameters by minimizing the errors between reprojected projection of ball markers and its acquired projection image. Although many attempts have been made to estimate the geometrical parameters, little attention has been paid to the optimal structure of calibration phantom. Inspired by the assumption that the larger the regularity of ball markers in the calibration phantom is, the more the stable is, and the better accuracy of estimated geometric parameters is, we propose a method to design phantom that maximize the accuracy of calibration process and mitigate the contribution of errors in indicating the ball centers. The method aims to maximize the regularity of ball markers in the calibration phantom and also in its projection image. The proposed method is applied to different phantom designs with the standard cylindrical holder and is proven to provide more accurate results than the traditional designs. The method can be applied to design scanner-dependent calibration phantoms and potentially free manufacturers and practitioners from manually searching work.