Accelerating Complex Materials Discovery with Universal Machine-Learning Potential-Driven Structure Prediction

TL;DR

This study tests whether universal machine-learning interatomic potentials can speed crystal structure prediction in complex quaternary oxides by applying M3GNet-DIRECT to Sr–Li–Al–O and Ba–Y–Al–O spaces. The approach uses DIRECT sampling and USPEX-driven searches, with DFT benchmarks (PBE, SCAN, R2SCAN, RPA) and phonon analyses to validate predicted phases. It rediscovered known materials outside the training set and uncovered seven new thermodynamically and dynamically stable compounds, including a novel Sr2LiAlO4 polymorph and a disordered Sr2Li4Al2O7, while revealing that PBE-based training can misorder phase stability and that CSP efficiency is currently dominated by search algorithms. The work highlights the promise and current limitations of uMLIP-driven CSP for materials discovery, emphasizing cross-validation with higher-level methods and the need for more efficient structure-search strategies to fully exploit surrogate models in complex chemical spaces.

Abstract

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.

Figures (4)

• Figure 1: Evolution of the lowest formation enthalpy as a function of generation during the crystal structure prediction searches for (a) Sr2LiAlO4 and (b) Ba2YAlO5. The red star marks the generation in which the correct experimental structure, shown in the inset, was first identified.
• Figure 2: (a) Calculated phase diagram of the Sr-Li-Al-O system. The green solid circle represents the stable phase, while colored diamonds denote metastable phases. The lighter the color of the diamond, the more unstable the phase. (b-e) Crystal structures and corresponding calculated phonon spectra for two newly identified stable compounds, Sr2LiAlO4 and Sr2Li4Al2O7.
• Figure 3: (a) Calculated phase diagram of the Ba–Y–Al–O system. The green solid circle represents a stable phase, while colored diamonds denote metastable phases. The lighter the color of the diamond, the less stable the phase. (b-c) Crystal structure and corresponding calculated phonon spectrum of Ba4YAlO7.
• Figure 4: Calculated energy differences between the Sr2LiAlO4 polymorphs ($P2_1/m$ and $P3_221$) using various exchange-correlation functionals.