Material innovation has driven human progress, with multicomponent materials playing an increasingly important role. While the Inorganic Crystal Structure Database (ICSD) contains approximately 200,000 inorganic compounds, a significant portion of ternary and higher-order compounds remain undiscovered. The existing database also exhibits chemical and synthetic biases, favoring readily synthesized compounds like oxides. Computational prescreening using density-functional theory (DFT) calculations can accelerate material discovery, as demonstrated by various studies on different material classes. However, DFT-based CSP suffers from low throughput due to the exponential increase in computational load with material complexity. Machine-learning potentials (MLPs), particularly neural network potentials (NNPs), offer a promising alternative by providing DFT-accurate energies at significantly reduced cost. The challenge lies in efficiently training these NNPs without prior knowledge of crystal structures. This study addresses this by using melt-quench molecular dynamics (MD) simulations to generate diverse local orders for training the NNP.

Numerous studies have leveraged DFT calculations to accelerate material discovery, predicting stable compositions for various applications including Li-ion batteries, nitride semiconductors, and high-Tc superconductors. However, these studies often focused on known prototypes or limited evolutionary steps due to the high computational cost of DFT-based CSP. Existing approaches for crystal structure prediction often involve data-mining known prototypes or using DFT-based evolutionary algorithms like genetic algorithms or particle swarm optimization. Machine learning potentials have emerged as a promising approach to accelerate CSP, with previous work demonstrating the effectiveness of Behler-Parrinello type NNPs trained using melt-quench annealing trajectories. However, limitations remain in choosing training sets without prior structure information. The present work addresses this by employing melt-quench MD simulations for self-starting training, thereby overcoming the limitations of prior approaches.

SPINNER combines NNPs with evolutionary or random searches for structure prediction. The workflow begins with a melt-quench-annealing MD simulation for a given chemical composition, generating trajectories used to train an initial NNP. The NNP is iteratively refined using low-energy structures identified during the early stages of CSP. The main CSP involves up to 5000 generations using a combination of random generation, crossover, permutation, and lattice mutation algorithms. These algorithms are specifically tuned for NNPs. Minimum distance constraints, derived from the melt-quench process, prevent the generation of unphysical structures. The crossover algorithm uses atomic energies to maintain stable chemical units. Final candidate structures (within 50 meV/atom) are then sorted after full relaxation using DFT calculations with the VASP package and the PBE functional. The SCAN functional and spin-orbit coupling were used in some cases to improve energy ordering. The training of NNPs utilizes the SIMPLE-NN package with Behler-Parrinello symmetry function vectors. The accuracy of the NNP is monitored by comparing DFT and NNP energies of experimental structures and averaged over final candidates.

In blind tests on 60 ternary compounds, SPINNER predicted the experimental (or theoretically more stable) structures for 75% of the materials, mostly within 1000 generations. For 6 materials, SPINNER predicted more stable structures than the experimental phases according to the PBE functional; this discrepancy was mostly resolved by using the SCAN functional. The accuracy of the NNPs was assessed using the energy difference between DFT and NNP calculations, averaging less than 12 meV/atom for materials with accurate structure prediction. Analysis of the evolutionary steps reveals that the ground state structure was often found directly from relaxing random structures without significant mutations. Melt-quench distance constraints and NNP-specific algorithms effectively guided the search toward the global minimum. SPINNER successfully identified experimentally observed metastable phases for some materials. Benchmarks against other CSP methods showed that SPINNER identified lower-energy structures for the majority of the tested non-oxide ternary compounds, often at higher Z values than in previous predictions. The improved performance of SPINNER is partly attributed to the unique NNP-specific algorithms and the iterative retraining scheme.

The computational cost of SPINNER (3-5 days on a 36-core node for 5000 generations) is significantly lower than DFT-based methods. Comparisons with USPEX, using both DFT and NNP, demonstrated the superiority of SPINNER's algorithms and features. The 25% failure rate of SPINNER highlights potential limitations, primarily related to poor NNP accuracy in some materials (possibly requiring longer evolution or improved NNP training), as well as the use of the conventional unit cell in structure generation. Transfer learning showed the ability of NNPs trained on a specific stoichiometry to predict structures for other compositions, even with valence changes. Tests on binary and quaternary compounds showed that SPINNER remains effective.

SPINNER is a highly efficient framework for crystal structure prediction, significantly accelerating the discovery of inorganic materials. Its success in identifying experimental or more stable phases in a large fraction of ternary compounds, along with its speed and scalability, opens avenues for large-scale, open exploration of undiscovered inorganic crystals. Future research could focus on improving NNP accuracy and transferability, extending SPINNER to more complex systems, and automating the entire workflow.

While SPINNER showed high accuracy in identifying equilibrium structures for many ternary compounds, limitations remain. The accuracy of the NNP is crucial; poor accuracy in a subset of compounds led to failures. In some cases, extending the evolutionary process beyond 5000 generations might be necessary. The choice of the conventional unit cell in structure generation could affect search efficiency, especially for compounds with anisotropic structures. Lastly, the current implementation of NNPs does not consider magnetic interactions, limiting its applicability to materials with significant magnetic ordering.