The study of the relative stability of multicomponent materials and the search for new materials for high performance applications requires extensive samplings of the composition space. This is a demanding task due to the computational effort that is required for the electronic structure calculations. In this work we propose a machine learning approach with descriptors of the local atomic environment using different chemistry heuristics based on smooth overlap of atomic positions, recursive solutions of tight-binding Hamiltonians and atomic cluster expansions. We demonstrate that these descriptors, which retain different levels of domain knowledge of structural and electronic properties of the chemical compounds, can be used to predict formation energies with high accuracy even with simple regression algorithms. We apply the methodology to complex crystal structures in intermetallic systems.