In recent years, efficient inter-atomic potentials approaching the accuracy of density functional theory (DFT) calculations have been developed using rigorous atomic descriptors satisfying strict invariances, for example, to translation, rotation, permutation of homonuclear atoms, among others. In this work, we generalize the spectral neighbor analysis potential (SNAP) model to bcc-fcc binary alloy systems. We demonstrate that machine-learned SNAP models can yield significant improvements even over well-established, high-performing embedded atom method (EAM) and modified EAM (MEAM) potentials for fcc Cu and Ni. We also report on the development of a SNAP model for the fcc Ni-bcc Mo binary system by machine learning a carefully-constructed large computed data set of elemental and intermetallic compounds. We demonstrate that this binary Ni-Mo SNAP model can achieve excellent agreement with experiments in the prediction of Ni-Mo phase diagram as well as near-DFT accuracy in the prediction of many key properties such as elastic constants, formation energies, melting points, etc., across the entire binary composition range. In contrast, the existing Ni-Mo EAM has significant errors in the prediction of the phase diagram and completely fails in binary compounds. This work provides a systematic model development process for multicomponent alloy systems, including an efficient procedure to optimize the hyper-parameters in the model fitting, and paves the way to long-time, large-scale simulations of such systems. arXiv:1806.04777v2 [cond-mat.mtrl-sci] 16 Aug 2018 Machine learning (ML) models based on robust local environment descriptors have recently emerged as an approach to describe the potential energy surface (PES) of systems of atoms with near-quantum accuracy at several of orders magnitude lower cost than ab initio methods. 1-6 Effective local environment descriptors must be invariant under translation, rotation, and permutation of homonuclear atoms, and have the properties of uniqueness and differentiability. 7 Examples of such descriptors include symmetry functions 1,8 , smooth overlap of atomic positions (SOAP) 4,9 , bispectrum 2,5 , Coulomb matrix 3,10,11 , among others.A typical approach is to fit the PES as a function of these descriptors by machine learning on ab initio data sets, using techniques ranging from simple linear regression 5,12 to kernel ridge regression 6,7 to neural networks [13][14][15][16] .Thus far, the development of ML potentials based on local environment descriptors have largely been limited to elements and oxides. The Gaussian approximation potential (GAP) using the SOAP descriptor has been applied on Si 4 , C 17,18 , W 9 , P 19 , and Fe 20 , and neural network models based on symmetry functions have been fitted for Si 21 , C 22 , Na 23 , ZnO 24 , TiO 2 25 , GeTe 26 , and Li 3 PO 4 27 . Thompson and Wood 5,28 have developed linear and quadratic models based on the SO(4) bispectrum -the Spectral Analysis Neighbor Potential or SNAP -for bcc Ta and W. Chen et al. 12 later showed that a line...