Researchers at the University of Moratuwa in Sri Lanka and the Sri Lanka Institute of Information Technology have built a machine-learning framework that screens oxide perovskites for band gaps suited to solar absorbers, narrowing a starting pool of over half a million candidate compositions down to 38 lead- and cadmium-free contenders worth further study.
Oxide perovskites are attractive for photovoltaics because their electronic and optical properties can be tuned through composition, but most known compositions are either insulating or have band gaps too wide for efficient light absorption. Identifying the minority with moderate, solar-relevant band gaps has traditionally meant either slow experimental trial and error or computationally expensive density functional theory (DFT) calculations, which become impractical once thousands of candidate compositions are involved.
The team started from a previously assembled database of 551,696 charge-neutral, geometrically formable oxide perovskites, screened using the Goldschmidt tolerance and octahedral factors. From these, 5,450 compounds with existing DFT band-gap values (calculated with the Local Density Approximation, or LDA) were used to train the models, while 515,452 structures, after excluding those with organic A-site cations, formed the pool the models were applied to. Because complete structural data wasn't available for many compounds, the researchers represented each material using composition-based Oliynyk feature vectors rather than structure-based descriptors, then pared the feature set down in three stages: removing low-variance and highly correlated features, using distance correlation and mutual information to surface features linked to band gap, and applying LightGBM feature importance to drop the rest.
Two XGBoost classifiers, built around band-gap thresholds of 0.5 and 2.0 eV, reached test accuracies of 96.6% and 95.7% respectively, and a weighted-voting ensemble of Extra Trees, SVR, CatBoost and LightGBM regression models refined those predictions further, with a mean absolute error of about 0.209 eV. Applying this pipeline surfaced around 15,966 structures with predicted band gaps in a 1.28-1.62 eV window, a range adjusted to account for LDA's known tendency to underestimate band gaps and based on the band gaps of high-efficiency halide perovskite solar cells rather than any demonstrated oxide-perovskite device. Filtering further for charge neutrality and structural formability narrowed the field to 592 compositions, and a final pass for elemental sustainability, toxicity, supply-chain availability and manufacturability cut that down to 38 candidates. Notably, none of the 38 contain lead or cadmium, and predicted band gaps for candidates such as Ba2GeSnO6 and K2MoSnO6 come out close to 1.5 eV, comparable to that of CdTe.
The authors are careful to frame the 38 compositions as a prioritized shortlist for further investigation rather than deployment-ready materials: thermodynamic and dynamic stability, light absorption and charge-transport behavior, defect tolerance, synthesis feasibility and actual device efficiency still need to be evaluated before any of them could be considered viable solar absorbers.
