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 Sr2LiAlO4 (P3221) and a new disordered phase, Sr2Li4Al2O7 (P1_bar). 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.

💡 Research Summary

This paper investigates the use of a universal machine‑learning interatomic potential (uMLIP), specifically the M3GNet‑DIRECT model, to accelerate crystal‑structure prediction (CSP) in complex quaternary oxide systems. Traditional CSP couples evolutionary global‑optimization algorithms with density‑functional theory (DFT) calculations, but the high computational cost of DFT limits exploration to relatively simple chemistries. Recent advances in uMLIPs, trained on massive databases, promise DFT‑level accuracy at a fraction of the cost, yet their ability to discover genuinely new, complex materials has not been systematically demonstrated.

The authors select two chemically rich, sparsely explored quaternary systems—Sr‑Li‑Al‑O and Ba‑Y‑Al‑O—as testbeds. M3GNet‑DIRECT is a retrained version of the first‑generation M3GNet, employing the DI‑RECT sampling strategy on the Materials Project (2021.2.8) dataset, which yields an average energy improvement of ~6 meV/atom over the original model. The CSP workflow is built on the USPEX evolutionary code. Each search starts with a population of 100 random structures; subsequent generations are generated using four variation operators (heritage 50 %, random 30 %, mutation 10 %, atomic mutation 10 %). The M3GNet‑predicted enthalpy serves as the fitness function. Convergence is declared when the lowest‑enthalpy structure remains unchanged for 35 consecutive generations or after 80 generations, and each search is repeated three times for reproducibility. Searches are limited to cells with ≤30 atoms and charge neutrality, resulting in 17 distinct compositions for Sr‑Li‑Al‑O and 29 for Ba‑Y‑Al‑O.

Benchmarking against two experimentally known quaternary compounds that are absent from the uMLIP training set—Sr₂LiAlO₄ (P2₁/m) and Ba₂YAlO₅ (P2₁/m)—demonstrates that the uMLIP‑driven CSP can recover the correct structures. Sr₂LiAlO₄ appears in the 11th generation, while Ba₂YAlO₅ emerges in the 44th generation, with a steady decline in enthalpy throughout the runs, confirming the model’s ability to guide the evolutionary algorithm toward the global minimum.

In the Sr‑Li‑Al‑O system, six candidates with formation energies (E_hull) below 30 meV/atom are identified. Two of these correspond to the known Sr₂LiAlO₄ polymorphs; the remaining four are novel. The most striking discovery is a new trigonal polymorph of Sr₂LiAlO₄ (space group P3₂₁) that is 8 meV/atom lower in energy than the experimentally reported monoclinic P2₁/m phase, with a calculated hull of 0 meV/atom. Phonon calculations show no imaginary modes, indicating dynamical stability. Another new compound, Sr₂Li₄Al₂O₇ (space group P 1̅), lies 9 meV/atom above the hull, exhibits a triclinic lattice, and retains its energy upon swapping Li and Al positions, suggesting possible disorder. Additional metastable but dynamically stable phases—SrLiAlO₃ (P4₃, 16 meV/atom), SrLiAlO₃ (C2/c, 25 meV/atom), SrLi₄Al₂O₆ (C2, 26 meV/atom), and Sr₂Li₃AlO₅ (P1, 29 meV/atom)—are also reported.

In the Ba‑Y‑Al‑O system, only two low‑hull candidates emerge: the known Ba₂YAlO₅ (0 meV/atom) and a newly predicted Ba₄YAlO₇ (space group P1) with an E_hull of 28 meV/atom. Phonon spectra confirm its dynamical stability.

A critical part of the study examines the reliability of the widely used PBE functional for phase‑stability ranking. For the two Sr₂LiAlO₄ polymorphs, PBE incorrectly predicts the trigonal P3₂₁ phase to be 71 meV/f.u. more stable than the monoclinic P2₁/m phase. Higher‑level methods—SCAN, R2SCAN, and the random‑phase approximation (RPA)—consistently favor the experimental monoclinic structure, with energy differences of 44–52 meV/f.u. This discrepancy highlights the necessity of cross‑validating uMLIP‑driven CSP results with more accurate exchange‑correlation treatments before experimental synthesis.

Beyond accuracy, the authors identify a shift in the CSP bottleneck: while uMLIPs dramatically reduce the cost of energy evaluations, the efficiency of the global‑search algorithm now dominates overall runtime. The current USPEX settings may not fully explore the vast configurational space of complex quaternaries, risking missed low‑energy structures. The paper suggests future directions such as integrating reinforcement‑learning or Bayesian optimization strategies, employing ensemble‑based uncertainty quantification for the uMLIP, and developing multi‑scale workflows that combine rapid uMLIP screening with selective high‑level DFT or RPA refinement.

In summary, the work demonstrates that a first‑generation universal ML potential can successfully rediscover known quaternary oxides and uncover several new thermodynamically and dynamically stable compounds, thereby validating uMLIP‑driven CSP as a viable tool for complex materials discovery. However, reliable phase‑stability predictions still require higher‑level electronic‑structure validation, and further advances in search‑algorithm efficiency are essential to fully exploit the speed gains offered by uMLIPs.