A novel chemistry reduction strategy based on convolutional neural networks (CNNs) is developed and applied to direct numerical simulation (DNS) of a turbulent non-premixed flame interacting with a cooled wall. The fuel syngas mixture is burning in pure oxygen. The training and the subsequent application of the CNN rely on the processing of two-dimensional (2D) images built from species mass fractions and temperature (CNN input), to predict the corresponding chemical sources at the center of the image (CNN output). This image-type treatment of chemistry is found to efficiently capture intermediate radicals species highly sensitive to the local flame topology. To reduce the CPU cost, a simplified 2D DNS database with detailed chemistry serves as reference and is used for training and testing the neural network. Comparisons are also made a posteriori against the same 2D DNS with a reduced chemical scheme specialized for syngas. Then, three-dimensional (3D) DNS are conducted either with CNN or the reduced chemistry for more a posteriori tests. The CNN reduced chemistry outperforms the reduced Arrhenius based mechanism in the prediction of radical species, such as monoatomic hydrogen, and also in terms of CPU cost.