Potential energy functions (PEFs) parametrized to bulk data are shown to perform poorly for small aluminum nanoparticles and clusters. In contrast, PEFs parametrized to a limited set of cluster and bulk data, but no nanoparticle data, perform well for nanoparticles. This validates a practical scheme for developing PEFs for nanoscale systems. Building on these findings, we optimized five PEFs by minimizing the error in the fit over a broad data set. Two of these PEFs have errors of less than or equal to 0.08 eV/atom for each of three categories of system sizes, i.e., for small clusters, for nanoparticles, and for bulk potential energies.