Accelerating complex materials discovery with universal machine-learning potential-driven structure prediction

Universal machine-learning interatomic potentials (uMLIPs) have become powerful tools for accelerating computational materials discovery by replacing expensive first-principles calculations in crystal structure prediction (CSP). However, their effectiveness in identifying new, complex materials remains uncertain. Here, we systematically assess the capability of a uMLIP (i.e., M3GNet) to accelerate CSP in quaternary oxides. Through extensive exploration of the Sr-Li-Al-O and Ba-Y-Al-O systems, we show that uMLIP can rediscover experimentally known materials absent from its training set and identify seven new thermodynamically and dynamically stable compounds. These include a new polymorph of Sr 2 LiAlO 4 ( P 3 2 21) and a new disordered phase, Sr 2 Li 4 Al 2 O 7 ( P 1 ‾ ). Furthermore, our results show stability predictions based on the semilocal PBE functional require cross-validation with higher-level methods, such as SCAN and RPA, to ensure reliability. While uMLIPs substantially reduce the computational cost of CSP, the primary bottleneck has shifted to the efficiency of search algorithms in navigating complex structural spaces. This work highlights both the promise and current limitations of uMLIP-driven CSP in the discovery of new materials. • Universal machine-learning potentials enable the discovery of seven new stable quaternary oxide materials •M3GNet successfully rediscovers known materials absent from its training dataset •Structural search algorithms become the primary bottleneck due to computational costin crystal structure prediction