To reduce the atmospheric turbulence-induced power loss,an AlexNet-based convolutional neural network[CNN]for wave-front aberration compensation is experimentally investigated for free-space optical[FSO]communication systems with standard single mode fiber-pigtailed photodiodes.The wavefront aberration is statistically constructed to mimic the received light beams with the Zernike mode-based theory for the Kolmogorov turbulence.By analyzing impacts of CNN structures,quantization resolution/noise,and mode count on the power penalty,the AlexNet-based CNN with 8 bit res-olution is identified for experimental study.Experimental results indicate that the average power penalty decreases to 1.8 dB from 12.4 dB in the strong turbulence.