We trained a convolutional neural network (CNN) to quickly and accurately reproduce energies, forces and dipole moments calculated by density functional theory (DFT) for fullerenes ranging from C20 to C80. The obtained reproduction errors per atom are about 0.001 Hartree for energies, 0.0003 Hartree/Bohr for forces and 0.001 Debye for dipole moments. The speedup of calculations with respect to DFT/6-31G is about 105 (for CPU), when using graphics processing unit (GPU) about 106. The obtained machine learned potential (MLP) can be used, for example, for a fast screening of the most stable fullerene isomers.