AI-Guided Recipe Tames the Chaos of High-Entropy Perovskite Solar Materials

Solar cells made from perovskite crystals have dazzled researchers for more than a decade, but the materials’ most tantalizing frontier, high-entropy perovskites, has long been a computational nightmare. These compounds cram five or more different elements onto a single crystallographic site, creating an almost unsearchable space of possible recipes. A new study published in the Journal of Materials Science by Wenjing Hu, Jiajun Jiang, Lin Peng, Tingting Shi, Xiaolin Liu and Jia Lin of Shanghai University of Electric Power and Jinan University describes a way to bring order to that chaos, using a strategy the authors call an Optimize-to-Optimize, or O2O, paradigm. Instead of letting machine-learning algorithms wander blindly through millions of possible compositions, the workflow first distills hard-won chemical rules from simpler, well-understood materials and then uses those rules to steer the search through the far larger high-entropy landscape.

The core insight of the paper is that knowledge gained from low-entropy systems, in which each crystal site is occupied by just one or two elements, can be transferred to high-entropy systems where the same sites host a blend of many elements. The team began by enumerating 1596 possible combinations of cesium-based chloride double perovskites, a lead-free family of semiconductors with the general formula Cs2BB’Cl6, where two different metal cations share the octahedral B-site of the crystal lattice. Rigorous stability screening narrowed this list to 275 viable semiconductors. From that curated set, the researchers extracted combinatorial rules and stability relationships, such as which pairings of cations are geometrically and thermodynamically compatible, and used them to define a rules-based element pool for constructing five-element B-site alloys.

On top of that constrained design space, the team layered an explainable machine-learning architecture that combines a Stacking ensemble with a Generalized Additive Model. Stacking blends the predictions of multiple base learners through a meta-model, capturing nonlinear interactions between elemental descriptors, while the additive component expresses the bandgap as a sum of smooth, interpretable functions of individual features. The result is a model that achieves a coefficient of determination, or R-squared, of 0.91 on its test set, meaning it accounts for roughly ninety-one percent of the variance in bandgap values across unseen compositions. Crucially, the additive framework does not behave like a black box: it yields closed-form formulas that quantitatively describe how each elemental property, most notably Pauling electronegativity, denoted by the Greek letter chi, shifts the energy gap between the material’s valence and conduction bands.

Bandgap is the single most important number in photovoltaics. It determines which slice of the solar spectrum a material can absorb, and the theoretical maximum efficiency of a single-junction solar cell peaks when the bandgap sits near 1.3 to 1.4 electron-volts. Most lead-free chloride double perovskites, however, have gaps far too wide to harvest sunlight efficiently, often exceeding two electron-volts. The new workflow revealed that by mixing five cations on the B-site, designers can either compress the gap dramatically or hold it steady, depending on how the electronegativities of the chosen elements are balanced. Because the descriptor-response functions are explicit, chemists can read off, in advance, roughly what bandgap a given five-element cocktail should produce, rather than synthesizing it and hoping for the best.

To confirm that the machine-learning predictions correspond to real physics, the authors turned to density-functional theory, the workhorse quantum-mechanical method of computational materials science. Two candidates emerged with strikingly different behaviors. The first, Cs2{ScNaInBGa}2Cl6, a compound whose B-site hosts scandium, sodium, indium, boron and gallium, showed bandgap compression from 2.96 down to 0.49 electron-volts, a reduction of nearly two and a half electron-volts that would allow the material to absorb light across almost the entire visible and near-infrared spectrum. The second, Cs2{BaZnCdSiGe}2Cl6, combining barium, zinc, cadmium, silicon and germanium, exhibited what the authors call compensatory alloying: its bandgap was retained almost perfectly, moving only from 1.56 to 1.57 electron-volts as the five elements’ opposing contributions canceled one another out. That 1.5-electron-volt regime is squarely within the sweet spot for high-efficiency single-junction photovoltaics.

Bandgap tuning alone is worthless if the resulting crystal falls apart, so the team subjected both candidates to a battery of stability tests. Tolerance-factor analysis, a geometric criterion relating the sizes of the ions to the stability of the perovskite framework, confirmed that both compositions fit the structural window in which the three-dimensional lattice can form. Calculated formation energies came out negative for both materials, indicating that they are thermodynamically favored to crystallize rather than decompose into competing phases. The researchers also evaluated the role of configurational entropy, the thermodynamic driver that stabilizes multielement alloys: the temperature-times-entropy term, T-delta-S, contributes a genuine stabilizing free-energy bonus that helps offset the strain of mixing five dissimilar cations on one lattice site.

The most demanding test was ab initio molecular dynamics, a simulation technique that tracks every atom in a crystal as it vibrates at finite temperature. The team ran 8-picosecond trajectories at 300 kelvin, roughly room temperature, for both candidate compositions. Throughout those simulations the perovskite frameworks remained intact, with no spontaneous bond breaking, cation migration or structural collapse. While 8 picoseconds is a fleeting instant compared with the years a solar panel must survive outdoors, such simulations are a standard first filter for weeding out compositions that would never survive synthesis, and both candidates passed cleanly.

What makes the O2O paradigm notable beyond its specific predictions is its philosophy. High-entropy materials research has exploded in recent years because mixing many elements confers remarkable robustness, tunability and, in some cases, entirely new functionalities, from entropy-stabilized oxides to high-entropy thermoelectrics. Yet the combinatorial explosion is brutal: with dozens of plausible cations and five slots to fill, the number of candidate compositions runs into the millions, far beyond what even the fastest high-throughput DFT pipelines can evaluate. By using interpretable rules harvested from the low-entropy regime as a prior, the new workflow effectively shrinks the search space before the expensive models ever run, a design pattern that echoes the way human chemists have always reasoned, but at machine scale and with quantified confidence.

The emphasis on explainability also matters for trust and reuse. Many state-of-the-art machine-learning models for materials, including graph neural networks, deliver impressive accuracy but offer little physical insight, making it hard for experimentalists to know when a prediction can be extrapolated safely. The Stacking-plus-GAM approach used here produces explicit descriptor-response curves, so the electronegativity dependence of the bandgap can be inspected, criticized and refined by any reader. The authors have made the entire pipeline reproducible, releasing source code, sample inputs and documentation through a public repository, which lowers the barrier for other groups to apply the same strategy to different perovskite chemistries, such as bromides, iodides or hybrid organic-inorganic frameworks.

Considerable work remains before any of these compositions light up a rooftop. The study’s candidates are prioritized hypotheses, not synthesized compounds, and the authors themselves frame the outcome as concrete composition rules, interpretable functions and a shortlist for further phase-stability assessment and experimental evaluation. Growing phase-pure five-element double perovskite crystals, verifying their gaps optically, and testing them in actual device architectures will take years. But the study offers a credible answer to one of the field’s most persistent questions: how to search an astronomically large compositional space without drowning in it. If the low-entropy-to-high-entropy knowledge transfer proves general, the approach could accelerate the discovery of lead-free, stable, spectrally matched absorbers, precisely the combination of properties that next-generation photovoltaics have been waiting for.

Subject of Research: Machine-learning-guided bandgap engineering of high-entropy lead-free chloride double perovskites for photovoltaics

Article Title: Intelligent bandgap engineering for high-efficiency perovskite photovoltaics: directing high-entropy optimization through low-entropy knowledge synergy

Article References: Hu, W., Jiang, J., Peng, L., Shi, T., Liu, X., & Lin, J. (2026). Intelligent bandgap engineering for high-efficiency perovskite photovoltaics: directing high-entropy optimization through low-entropy knowledge synergy. Journal of Materials Science. https://doi.org/10.1007/s10853-026-13827-0

Image Credits: AI Generated

DOI: 10.1007/s10853-026-13827-0

Keywords: high-entropy perovskites, bandgap engineering, photovoltaics, machine learning, double perovskite, density functional theory, lead-free solar materials, explainable AI, configurational entropy, ab initio molecular dynamics, chloride perovskite, materials discovery