Because of the difficulty in locating error bits in polarization code decoding caused by atmospheric turbulence in free-space optical communication, this paper proposes an LSTM-SCFlips decoding method in free-space optical communication. First, one-hot precoding is performed on the log-likelihood ratio (LLR) information sequence of polarization codes in serial cancellation (SC) decoding, and the characteristics of the information sequence are analyzed and learned in different training step sizes. The root mean square error and computational complexity of the neural prediction model are comprehensively considered, and an appropriate training step size is selected. After the accuracy of the prediction results is improved, the phenomenon of overfitting the prediction results is further eliminated. The LSTM neural network model is used to locate the first error bit of SC decoding and single-bit or multi-bit flips of the SC decoding algorithm according to the error probability are performed. The simulation results show that in different weak atmospheric turbulence intensities, the proposed decoding method can better identify the optimal flip position and reduce the computational complexity on the premise of sacrificing few computing resources. In addition, better bit-error-rate performance is achieved. When the bit error rate is 10 ^-4, the correct recognition rate of the optimal flipped bit of the LSTM-SCFlips decoding method is increased by 7 percentage points, and the coding gain of 0.3 dB--1.2 dB is generated.