Platinum (Pt) alloys are expected to overcome long-standing issues of Pt/C electrocatalysts for oxygen reduction reaction (ORR). Entangled with serious uncertainty in configurational and compositional information, the design of a promising multi-component electrocatalyst, however, has been delayed. Here, we demonstrate that a first-principle database-driven machine-learning approach is extremely useful for the purpose via exploring materials beyond the regime of pure quantum mechanical calculations. Guided by a computational ternary phase diagram we indeed experimentally synthesized a PtFeCu nanocatalyst with 2 g per batch capacity and measured its catalytic performance for ORR. Both our computation and experiment consistently demonstrate that PtFeCu is highly active due to the atomic distribution of Cu leading to beneficial modulation of surface strain and segregation. Strikingly, PtFe high Cu low (776 μA cm −2 Pt and 0.67 A mg −1 Pt ) exhibits not only 3-fold better specific and mass activities than Pt/C but also little performance degradation over the accelerated stress test. • First-principle establishment of catalytic property data for ternary nanoparticles • Parametrization of atomistic interaction into neural network potentials • A striking consistency between computational prediction and experimental validation • Exceptional catalytic mass activity of PtFeCu 3-fold higher than that of Pt/C One of the challenging issues in nanocatalyst design for catalysis applications is to introduce the atomic-level functionalities of nanoparticles, determining the activity and stability. Multi-component systems can be promising, but the enormous configurational degrees of freedom and optimization routes for a target reaction remained unsolved. We demonstrate the combination of a machine-learning approach and experimental tests of PtFeCu nanoparticles for oxygen reduction reaction. Promising candidates were efficiently screened theoretically and successfully validated by the experiments. Both our computational and experimental outcomes indicate that the highly active electrocatalytic performance of PtFeCu originates from Cu atomic distribution. Our study suggests that first-principle calculations combined with machine-learning technology can be a promising approach in design of nanocatalysts to reduce the gap between the simulations and experiments. To efficiently design the Pt-based alloy nanocatalysts for oxygen reduction reaction, we utilized first-principle data integrated with machine learning to search for PtFeCu configurational spaces. We identified the promising candidates and revealed the atomic-level understanding via first-principle calculations. Cu atomic distribution remarkably modulated surface strain energies and segregation of the alloying components toward better oxygen reduction reaction performance. Tuning of Cu contents led to the high electrochemical performance of PtFeCu nanocatalysts, which was successfully validated by experiments.