With a revisit time of 1 h, spatial resolution of 500 m, and high radiometric sensitivity, the Geostationary Ocean Color Imager (GOCI) is widely used to monitor diurnal dynamics of oceanic phenomena. However, atmospheric correction (AC) of GOCI data with high solar zenith angle (>70°) is still a challenge for traditional algorithms. Here, we propose a novel neural network (NN) AC algorithm for GOCI data under high solar zenith angles. Unlike traditional NN AC algorithms trained by radiative transfer-simulated dataset, our new AC algorithm was trained by a large number of matchups between GOCI-observed Rayleigh-corrected radiance in the morning and evening and GOCI-retrieved high-quality noontime remote-sensing reflectance (Rrs). When validated using hourly GOCI data, the new NN AC algorithm yielded diurnally stable Rrs in open ocean waters from the morning to evening. Furthermore, when validated by in-situ data from three Aerosol Robotic Network-Ocean Color (AERONET-OC) stations (Socheongcho, Gageocho and Ieodo), the GOCI-retrieved Rrs at visible bands obtained using the new AC algorithm agreed well with the in-situ values, even under high solar zenith angles. Practical application of the new algorithm was further examined using diurnal GOCI observation data acquired in clear open ocean waters. Results showed that the new algorithm successfully retrieved Rrs for the morning and evening GOCI data. Moreover, the amount of Rrs data retrieved by the new algorithm was much higher than that retrieved by the standard AC algorithm in SeaDAS. Our proposed NN AC algorithm can not only be applied to process GOCI data acquired in the morning and evening, but also has the potential to be applied to process polar-orbiting satellite ocean color data at high-latitude ocean that also include satellite observation with high solar zenith angles. © 2020 Elsevier Inc.