A novel full spectrum correlated k-distribution model based on multiband fusion artificial neural network for gas absorption coefficient prediction.

• A novel multiband fusion artificial neural network (MFANN) was proposed for gas radiative properties prediction. • MFANN enables the prediction of gas absorption coefficient in a sub-band selectable manner. • The detailed datasets building strategies for MFANN were discussed. • MFANN achieves an error of less than 1.83 % compared with the line-by-line method. • MFANN spends about 1.33 s to predict the k -distributions for the gas mixture of CO, CO2, and H 2 O of 40 equidistant sub-bands in a wavelength range of 1.5–5.5 μm. Gas radiation exhibits strong selectivity in the spectrum, with its radiation and absorption capabilities only present at distinct wavelength bands. This implies that calculating gas radiative property across the entire wavelength range is unnecessary. In contrast, a computing strategy oriented to the radiative properties for the wavelength band of interest is of great importance for radiative transfer studies of gases. On one hand, this strategy enables researchers to avoid interference caused by noise from non-interest bands, thereby increasing the signal-to-noise ratio and improving computational accuracy. On the other hand, it reduces the computational burden by curtailing calculations in unnecessary bands. In light of these significances, this study proposed the multiband fusion artificial neural network (MFANN) model to predict the radiation property of CO, CO 2 , and H 2 O mixture at a pressure range of 0.5–4.0 atm, a temperature range of 300–3000 K, and a wavelength range of 1.5–5.5 μm. The distinctive feature of this model lies in its division of the target wavelength band into multiple contiguous sub-bands. A single-layer network is trained on each sub-band, afterword the network obtained from each sub-band is fused to emerge a compact big model which enables fast prediction of the gas absorption coefficient, in a correlated k -distribution format, for the mixture within all sub-bands. This study compared the prediction results of the fusion model with benchmarks calculated by the line-by-line (LBL) and full spectral correlated k -distribution (FSCK) methods under different cases to evaluate the model's performance. The results show that the Symmetric Mean Absolute Percentage Error (SMAPE) of MFANN is less than 1.83 %, and MFANN only spends about 1.33 s to predict the k -distribution absorption coefficient at the wavelength range of 1.5–5.5 μm with a sub-band interval of 0.1 μm. [ABSTRACT FROM AUTHOR]