Discovering multicomponent inorganic compounds is crucial for addressing scientific and engineering challenges, but the vast material space hinders experimental synthesis. Crystal structure prediction (CSP) offers a solution, but its computational cost remains a bottleneck. This paper introduces SPINNER, a structure-prediction framework using neural network potentials (NNPs) that accelerates CSP by 10<sup>2</sup>-10<sup>3</sup> times compared to DFT-based methods. SPINNER incorporates algorithms optimized for NNPs, achieving higher accuracy than conventional algorithms. Blind tests on 60 ternary compositions show SPINNER identifies experimental or theoretically more stable phases for ~80% of materials, outperforming data-mining and DFT-based evolutionary predictions. SPINNER paves the way for large-scale exploration of undiscovered inorganic crystals.
