An extended narrow-band k-distribution method for gas radiation characteristics based on a hybrid deep neural network

The calculation of gas radiative properties is a fundamental problem in numerous fields such as combustion diagnostics, aerospace engineering, and atmospheric remote sensing. Developing computational models that combine both high efficiency and accuracy has become a shared goal and central challenge in this field. To this end, this study integrates the physical mechanisms of gas radiation with data-driven approaches, proposing an Extended Narrow-Band k -distribution (ENBK) method based on a Hybrid Deep Neural Network (HDNN). The ENBK method extends the high spectral-resolution k -distribution absorption coefficients of multiple single components into mixed-gas multi-band k -distribution absorption coefficients through probability density function convolution, thereby enabling the calculation of radiative properties of inhomogeneous mixed gases across multiple spectral resolutions. Meanwhile, a narrow-band k -distribution absorption coefficient database for H 2 O and CO 2 was constructed using the NBK method, covering a wide range of thermodynamic states with high spectral resolution (5 cm -1 ), which serves as training data for a designed deep neural network that combines fully connected and transposed convolutional layers. The computational accuracy of ENBK and HDNN was evaluated against the Line-by-Line (LBL) and NBK method for several typical engineering cases. The results demonstrate that ENBK maintains high accuracy in both mixing multiple gas components and extending spectral bands, with average relative errors of less than 0.5% in calculated narrow- and wide-band radiation intensity and transmissivity compared to the LBL benchmark method. Furthermore, the trained HDNN model significantly reduces database storage requirements (saving over 90% of memory) while enabling rapid and accurate prediction of narrow-band k -distribution absorption coefficients under given thermodynamic conditions.