AI must link the full perovskite solar-cell research cycle

Artificial intelligence is helping researchers improve perovskite solar cells, but its benefits remain largely confined to the particular systems on which models are trained. A Nature Reviews Electrical Engineering analysis, published by nature.com, says wider progress will depend on connecting AI tools across the entire research cycle rather than relying on isolated prediction or optimisation tasks.

Halide perovskites are attractive for solar technology because their optical and electronic characteristics can be adjusted, while their manufacturing costs can be low. Yet turning those advantages into industrial deployment is difficult. The behaviour of a cell depends on several linked choices: material composition influences how processing proceeds, processing shapes the resulting microstructure, and that microstructure affects device performance.

Because these factors interact, a solution found in one material system or manufacturing setting may not work when the composition, fabrication conditions or device architecture changes. The challenge is therefore not simply to find the best value for one parameter. It is to understand and manage a chain of dependencies that operates across different stages and scales of the technology.

Where AI is already helping

AI is now used in four broad parts of perovskite solar-cell research: discovering materials, engineering devices, refining manufacturing processes and studying stability. In settings where sufficient data are available, models can identify relationships and make optimisation faster. This can reduce the effort required to search through possible choices within a defined research problem.

The gains, however, are uneven. A model that performs well for one combination of materials and procedures may not offer the same value for another. Nor does a strong prediction necessarily reveal the mechanism responsible for the observed result. According to the analysis, current systems are better at finding correlations inside familiar, data-rich conditions than at producing insights that remain useful across different compositions, architectures or operating conditions.

That gap matters because perovskite solar cells are not governed by independent variables. A change made to composition can affect processing; a processing change can alter the material structure; and the structural change can influence performance and reliability. AI that treats each stage as a separate task risks losing information that could determine whether an improvement transfers beyond the original experiment.

Four obstacles to broader impact

The analysis identifies four recurring structural problems. First, relevant information is fragmented and datasets are not readily interoperable. Results generated in different parts of the research process may therefore be difficult to combine, limiting what a model can learn from them together.

Second, models often generalise poorly between systems. Patterns learned from one set of compositions, manufacturing environments or device designs do not automatically carry over to another. This restricts the value of AI outside the conditions represented in its training data.

Third, many models offer limited interpretability and weak physical grounding. A prediction may be useful without explaining why it was made, but the absence of a clear connection to the underlying behaviour makes it harder to judge whether the result should apply in a new setting. It also limits the extraction of mechanism-based knowledge that could guide future work.

The fourth problem is a mismatch between what AI is asked to optimise and the constraints that shape practical design. A model may improve a selected target while failing to account for the broader requirements of a working solar-cell technology. The article presents these difficulties as connected features of the research structure, rather than as evidence that better algorithms alone will solve the problem.

From separate models to an integrated pipeline

The proposed response is to build research infrastructure that joins information and feedback across the field. Multimodal and multilevel datasets could bring together forms of evidence generated at different stages. Physics-informed and explainable models could make predictions more closely related to the behaviour they are intended to represent, while also making their reasoning easier to assess.

The analysis also points to closed-loop systems in which AI predictions are connected to experimental feedback. In such a pipeline, results from experiments would not merely provide a final test of a model; they could help update the system and inform subsequent choices. That arrangement would allow knowledge to be carried forward instead of being left within a single optimisation exercise.

The intended shift is from task-specific prediction towards system-level AI. Under that approach, information from materials discovery, device engineering, processing and stability analysis would be continuously propagated, revised and reused. AI would become part of an integrated programme for discovering materials, improving devices and assessing reliability.

For perovskite photovoltaics, the significance lies in transferability. Faster optimisation within one narrow regime is useful, but industrial progress requires approaches that can cope with linked decisions across compositions, manufacturing conditions and device architectures. The Nature analysis argues that this broader role for AI will depend on preserving knowledge throughout the research cycle, not simply on making individual models more accurate.