Abstract: Classical density functional theory (cDFT) is a powerful framework for describing the intricatethermodynamic equilibrium properties and structural aspects of classical many-body systems, re-lying solely on the one-body density profile. At the heart of cDFT lies the (excess) Helmholtzfree-energy functional, which is not known exactly. Here we introduce a novel neural network ap-proach, aimed at deducing this functional for a supercritical Lennard-Jones system by exclusivelytraining on a dataset comprising of radial distribution functions of homogeneous bulk fluids. Thismethod circumvents the need to sample costly density profiles in a wide variety of external fields.We demonstrate that the neural functional accurately predicts the inhomogeneous particle densityunder various complex external fields in planar geometry, while simultaneously offering precise ex-cess free energy predictions. The method offers a pathway to significantly ease the computationaldemands for future attempts to extend machine learning for cDFT to arbitrary three-dimensionalsystems.