A unified modelling framework for all unresolved terms in the filtered progress variable transport equation in large-eddy simulations of turbulent premixed flames is proposed, using convolutional neural networks. A direct numerical simulation database of a turbulent premixed stoichiometric methane/air jet flame is used in order to train convolutional neural networks to predict both the filtered progress variable source term and the unresolved scalar transport terms. A single variable readily available from the large-eddy simulation is required in order to calculate all inputs to networks, namely the Favre-filtered progress variablec. In the context of flame tabulated chemistry (premixed flamelet), the trained networks are shown to produce quantitatively good predictions of all unresolved terms in an apriori study, despite their different nature and irrespective of variations in filter size, without having to resort to solving any additional transport equations. The framework proposed in this study thus opens perspectives for the application of deep learning to the modeling of the non-linear aerothermochemistry equations which involve unresolved source and transport terms.