Learning stability across disordered oxide alloy space
Group-III oxides are functional materials used in transparent conductors and wide-bandgap semiconductors. Alloying and doping can tune the band gap and other functional properties, but each added degree of freedom also expands the number of possible compositions and atomic configurations that need to be explored. This work used machine learning to extend first-principles data across that space, screening stability before precise atomic structures are known and testing how information can be reused in related doped materials.
A large materials space from composition and disorder
Changing the Al, Ga and In content of Group-III oxides, rearranging atoms on the lattice and introducing dopants can tune stability and electronic properties. The same overall composition can also correspond to many atomic configurations, so the number of candidate structures grows rapidly.
First-principles calculations provide reliable reference data, but evaluating every composition and configuration individually is impractical for large-scale screening. A faster model can use a smaller set of high-quality calculations to rank candidates before more expensive calculations are applied.
Screening before precise structures are known
The model does not require a fully determined atomic structure for every candidate. Approximate structures are generated from composition and lattice occupations, with lattice parameters estimated from composition using Vegard's law.
These structures are not final equilibrium geometries, but they retain enough compositional, site and geometric information to serve as low-cost inputs for stability prediction. This allows screening to take place before a candidate receives a more expensive high-accuracy structural calculation.
Comparing material representations
The study compared representations that encode different levels of structural information. Cluster expansion describes lattice sites and their elemental occupations; crystal-graph n-grams capture local coordination and connectivity; many-body tensor representation also retains continuous geometric relationships between atoms.
The comparison shows why the choice of representation matters when local distortion and structural relaxation affect energy. Representations that retain continuous geometry can use information in the approximate structures that is absent from ideal-lattice descriptions.
Extending from parent alloys to doped systems
Doped materials share much of the lattice and chemistry of the parent Group-III oxide alloy while introducing new local environments. Models built mainly from parent-alloy data were therefore extended with data from target doped systems to test how far existing information remains useful.
Transfer is strongest when the new materials remain close to structural environments already represented in the training data. New local environments require additional target-system data. The same principle later became central to transfer learning and domain adaptation in materials modelling: related material families can share predictive information instead of being treated as independent problems.
Exploring configurational space
The learned energy model can also be combined with statistical sampling to compare many atomic arrangements. First-principles calculations provide reference energies, while the machine-learning model handles the repeated evaluations needed to explore composition, site occupation and temperature-dependent configurational behaviour.
This is particularly useful for disordered alloys and multicomponent materials, where the relevant design space is a population of possible configurations rather than a small set of preselected structures.
From oxide alloys to broader materials discovery
Complex alloys, semiconductors and energy materials face the same imbalance: the design space is large, while reliable experimental and quantum-mechanical data cover only a small fraction of it. Materials machine learning, transfer learning and foundation models increasingly aim to reuse existing data across related chemical and structural spaces.
Group-III oxide alloys provide a concrete example of that workflow. Candidates can be screened from composition and inexpensive approximate structure information before precise structures are available, while data from a parent material family can support prediction in related doped systems.