根据瞳面相位差波前传感器的特点，提出把模式分解和随机并行梯度下降算法结合起来实现波前相位的恢复。以32单元变形镜的初始面形和Roddier提出的相位生成方法随机产生的一帧相屏为研究对象，分析所提出的相位恢复算法的性能。结果表明，两种情况下的畸变波前都得到了很好的恢复。当像差仅含有低阶成分时，使用模式分解法就可获得令人满意的效果。高阶成分较多时，各阶之间的耦合使得模式分解法的误差增大。采用模式分解法得到初始解，在初始解的基础上继续采用随机并行优化算法进行优化，可以得到精确解。对噪声情况下的相位恢复结果表明，该相位恢复算法具有较强的抗噪能力。

A kind of phase retrieval algorithm, which combines Zernike-mode decomposition with stochastic parallel gradient descent (SPGD), is raised to realize wavefront phase recovery based on pupil phase diversity (PPD). The performance of phase retrieval algorithm is analyzed through two different aberrated wavefronts. One is an initial shape of 32-element deformable mirror sampled by the interferometer; the other is generated through the method proposed by Roddier. Simulation results show that above two aberrated wavefronts can be restored successfully. For relatively small aberrations, satisfactory results are obtained only using the mode decomposition. When the aberrated wavefronts include many high frequency components, the algorithm only using the mode decomposition does not fit for this situation because of coupling among different orders. The results of the mode decomposition are used as the initial solution of SPGD algorithm, which searches the optimum solution. Simulation results under noise show the phase retrieval algorithm has strong anti-noise ability.