The computation of the infrared signature of an aircraft is a challenging problem in modern aircraft design. The most accurate algorithm for the numerical simulation of thermal radiation is the Line-By-Line method (LBL), which solves the radiative transfer for a very large number of spectral lines (order of magnitude 10^5 for aircraft engine plumes). However, due to their high computational cost, limited application of LBL computations can be found in literature, mostly for benchmark analysis. In recent years, several studies have focused on developing data-driven models to predict radiation spectra, such as [1], although only narrow windows of the spectrum have been used for training, hence making the models unfit for the computation of the aircraft infrared signature. To address this issue, we propose a methodology that combines the high fidelity of LBL calculations with the efficiency of Machine Learning (ML). Our multiphysics approach integrates Computational Fluid Dynamics (CFD) into a wider framework to include the simulation of thermal radiation. This framework was developed in-house and is called NOIR (Nozzle Optical Infrared Radiance), and was applied to the analysis of the infrared signature of an aircraft plume. A CFD simulation provided detailed flowfield informations, including temperature, pressure and species mass fractions. Data-driven models are trained on spectra databases of carbon dioxide and water vapour in the wavelength band 3 − 5 μm. The architecture employed is a deep feed forward network, with transformer encoders in a double-pass logic to capture the highly chaotic behaviour of the spectra. The models are then implemented into NOIR, resulting in two different versions: NOIR-LBL (using LBL method) and NOIR-ML (using data-driven models). Figure 1a shows the comparison between the neural network (NN) predictions and the ground truth for a sample spectrum of CO2, while Figure 1b shows the comparison of the radiance intensity maps of the two codes NOIR-LBL and NOIR-ML. The difference of the infrared signature was ≈ 2%, while the NOIR-ML simulation was 52 times faster.