Currently most coal combustion simulations treat the devolatilization products as a mixture of light gases with a given proportion or a postulate substance, which is obviously different from the reality. To obtain a more accurate treatment on the product distribution from coal devolatilization, an artificial neural network (ANN) model is innovatively developed based on a training database constructed from diverse experimental data for a wide range of coal types under a wide range of heating conditions. The accuracy and applicability of the developed ANN model are validated and compared with that of the chemical percolation devolatilization coupled with the functional group (FG-CPD) model for the validation database, and the relative impact of each input parameter on the evolution of each devolatilization product is evaluated. The results show that the detailed product distributions of coal devolatilization predicted by the proposed ANN model are in good agreement with the experimental data for both the training and validation database, and the ANN model can give a more accurate prediction on both the yield of each component and its evolution compared with the FG-CPD model. The coal composition accounts for the most impact (above 60%) on the product distribution, and the relative impact of C-daf, H-daf, O-daf, coal particle diameter, instantaneous heating rates, particle residence time and particle temperature decrease successively. This ANN model has great potential to be coupled into coal combustion simulations to improve efficiency and accuracy, which will be studied in the future. (C) 2018 The Combustion Institute. Published by Elsevier Inc. All rights reserved.