The scenario code CLASS relies on infinite assembly simulation to predict fuel actinide inventories at exit burnup. In the current work, we replace these assembly calculations by full-core simulations and evaluate the impact on actinide inventories predicted by CLASS. To achieve this goal, we generate neural network training databanks for CLASS using the lattice code DRAGON5. For UOX fuels, the databanks are sampled stochastically for exit burnup, moderator boron concentration and uranium 235 enrichment while for MOX fuels an eight-dimensional grid is sampled that also accounts for plutonium and americium-241 initial contents. DRAGON5 is used to generate the databases for DONJON5 3D full-core diffusion calculations in CLASS. Results obtained using neural networks CLASS and DONJON5/CLASS calculations are then compared to assess the different assumptions used in classical scenario simulations and determine the major source of errors. A simple UOX scenario involving long-term fuel storage and a more complex scenario involving reprocessed UOX spent fuel and MOX fabrication are then studied. They show that inventories of uranium 235 and minor actinides are sensitive to full-core simulations. Moreover, the neural networks CLASS simulations can be improved using an adapted that depends on the initial fuel composition.