The review, led by Yifan Li and Guixiang Li of Southeast University in Nanjing, together with Wei Liu and Shimin Di of Southeast University’s School of Computer Science and Engineering, Qing Cao of the University of Notre Dame, and Mohammad Khaja Nazeeruddin of EPFL, takes an unusually candid look at how machine learning has been deployed across the perovskite research cycle. The authors’ verdict is nuanced. AI has genuinely accelerated local optimization, the authors find, reliably identifying correlations and speeding up experimentation in data-rich regimes. What it has not done is extract mechanism-anchored insights that transfer across compositions, fabrication environments or device architectures. A model trained to predict efficiency for one perovskite recipe in one laboratory typically says nothing useful about a different recipe processed in open air across town.

The root of the difficulty, the authors argue, lies in the physics of the material itself. In a perovskite solar cell, composition governs processing dynamics: change the ratio of formamidinium to methylammonium cations, or swap one halide for another, and the crystallization pathway of the film shifts, altering nucleation rates, intermediate-phase formation and grain growth. Processing in turn determines microstructure, the grain sizes, orientations and defect densities that set how efficiently charge carriers are extracted and how quickly the film degrades. Microstructure then sets device performance and stability. Because these links are causal and cross-scale, an improvement identified at one level rarely survives translation to another. The review frames this as the central obstacle to industrial deployment: not that any single dependency is unknowable, but that optimizing them in isolation produces gains that evaporate when the system is perturbed.

Against this backdrop, the authors survey AI applications across four domains of perovskite research. In materials discovery, machine learning models now predict bandgaps, screen passivation molecules and even design organic ligands; AI-generated ammonium ligands have been used to build high-efficiency and stable two-dimensional/three-dimensional heterojunction cells, and inverse-design workflows have discovered hole-transport materials tailored for perovskite devices. In device engineering, machine vision tools quantify microstructure disorder and AI-assisted performance analysis predicts processing parameters rapidly. In process optimization, machine learning with knowledge constraints has guided open-air manufacturing, robotic platforms have explored thin-film parameter spaces, and high-throughput robotic learning has uncovered phenomena such as temperature-induced stability reversal. In stability analysis, big-data approaches have mined ageing datasets, and wavelet-aided models have predicted long-term outdoor performance. Each of these is a genuine advance, the review stresses, but each is also a local one.

From this survey the authors distil four structural bottlenecks that recur across the field. The first is fragmented data. Perovskite results are scattered across papers and laboratories in formats that cannot be merged, with inconsistent reporting of processing conditions, device architectures and measurement protocols. Efforts to build interoperable descriptions, capable of tracking the hundreds of ions that appear across the perovskite family, remain in their infancy. The second is weak generalization. Models excel within the distribution of their training data but fail when asked to extrapolate to new compositions or architectures, precisely the regime where discovery happens. Few-shot learning methods, which aim to extract maximum insight from minimal experiments, are emerging as a partial remedy, but the underlying data scarcity remains acute.

The third bottleneck is limited interpretability. Many high-performing models behave as black boxes, offering predictions without physical grounding, which makes it hard for researchers to distinguish a genuine materials insight from a statistical artifact of the dataset. Explainable AI techniques and physics-informed models, which embed known constraints such as thermodynamic limits on synthesizability or the geometric tolerance factor that predicts structural stability, are proposed as ways to anchor predictions in mechanism. The fourth bottleneck is misaligned design objectives. AI models are typically trained to maximize a single metric, most often power conversion efficiency, whereas practical deployment demands simultaneous optimization of efficiency, stability, cost, manufacturability and environmental safety, including the management of lead content. Multi-objective frameworks that map Pareto fronts, in which no objective can be improved without worsening another, are needed to align machine optimization with real engineering trade-offs.

The review’s most consequential argument is that addressing these bottlenecks requires integrated infrastructures rather than more accurate individual models. Multimodal and multilevel datasets, combining composition, processing logs, microscopic imagery and device telemetry, would attack fragmentation. Physics-informed and explainable models would attack the interpretability and generalization gaps. Closed-loop coupling between AI predictions and experimental feedback, in which a model proposes experiments, a robotic platform executes them and the results update the model, would attack the misalignment between prediction and practice. The authors point to autonomous laboratories as evidence that this vision is technically feasible: closed-loop frameworks have already demonstrated reproducible perovskite solar cell fabrication, and self-driving research workflows are unlocking end-to-end experimentation in the field.

Looking forward, the authors sketch what they call system-level AI integration: a shift from task-specific predictive tools towards frameworks in which knowledge is continuously propagated, updated and reused across the entire perovskite research cycle. In such a framework, an insight about ligand design discovered during materials synthesis would inform process optimization, which would inform stability modelling, which would feed back into materials selection, without the knowledge being lost at each hand-off. Emerging technologies, including large-language-model-based multi-agent systems in which specialized AI agents cooperate on materials tasks, and physical neural networks with self-learning capabilities, are identified as candidate building blocks. The ambition is to reframe AI from a predictive tool into a participant in materials discovery, device optimization and reliability engineering, a collaborator that carries context across the pipeline rather than solving each stage in isolation.

The stakes are considerable. Perovskite-silicon tandem cells are widely seen as the next major efficiency leap for commercial photovoltaics, and methylammonium-free wide-bandgap perovskites are central to that prospect. But the same coupling that makes perovskites scientifically fascinating makes them industrially treacherous: a composition optimized for efficiency in a nitrogen glovebox may fail within months under the temperature cycling, humidity and illumination of a rooftop. Recent work on strain regulation, oriented nucleation and interfacial contact engineering has extended device lifetimes substantially, and large ageing datasets have revealed that stability tends to follow efficiency, but translating those findings into certified, bankable modules requires exactly the kind of cross-scale, cross-lab knowledge transfer that current AI tools cannot yet deliver.

The review’s message to the field is ultimately one of redirection rather than discouragement. The past several years have produced an impressive arsenal of machine learning methods for perovskite research, from generative models for molecular screening to robotic platforms for high-throughput synthesis. What is missing is the connective tissue: shared data standards, physics-anchored models, and closed loops that turn predictions into experiments and experiments into transferable knowledge. If the field can build that infrastructure, the authors conclude, AI could become the system-level engine that finally carries perovskite photovoltaics from record-breaking laboratory cells to durable, manufacturable technology. If it cannot, the field risks another decade of locally optimal, globally stagnant progress, with each laboratory perfecting a device that no other laboratory can reproduce.

Subject of Research: Artificial intelligence applications and structural limitations in halide perovskite solar cell research and development

Article Title: Towards system-level artificial intelligence in perovskite photovoltaics

Article References: Li, Y., Liu, W., Cao, Q., Di, S., Nazeeruddin, M. K., & Li, G. (2026). Towards system-level artificial intelligence in perovskite photovoltaics. Nature Reviews Electrical Engineering. https://doi.org/10.1038/s44287-026-00332-4

Image Credits: AI Generated

DOI: 10.1038/s44287-026-00332-4

Keywords: perovskite solar cells, artificial intelligence, machine learning, photovoltaics, materials discovery, closed-loop automation, data fragmentation, explainable AI, device stability, self-driving laboratories, process optimization, tandem solar cells