High-accuracy machine learning approach for predicting J-V characteristics of perovskite solar cells under variable irradiance – Scientific Reports

Perovskite solar cells (PSCs) have surged from a 3.8% efficiency debut in 2009 to certified records above 26%, reshaping the conversation around next-generation photovoltaics. But real-world deployment still faces a stubborn challenge: performance swings under changing sunlight. Traditional lab measurements aren’t built for the constantly shifting irradiance solar panels experience outdoors, and exhaustive experimental characterization is slow and costly. A new machine learning (ML) approach aims to close that gap by predicting the current–voltage (J–V) behavior of PSCs across variable irradiance with high accuracy, speed, and scalability.

Why this matters

PSCs promise low-cost fabrication, tunable optoelectronic properties, and top-tier power conversion efficiency (PCE). Yet their sensitivity to environmental conditions—especially sunlight intensity that fluctuates hourly and regionally—makes performance evaluation complex. Relying only on fixed-illumination tests can mask how a device behaves in the field, hindering optimization and eroding confidence for scale-up.

By integrating data-driven prediction into the evaluation workflow, researchers can rapidly generate realistic J–V curves across light levels, stress-test designs before fabrication, and prioritize the most promising material and device configurations.

From lab curves to learned curves

Machine learning has become a powerful ally in photovoltaics and materials science, where intricate, nonlinear relationships tie composition, processing, and structure to device performance. Models such as artificial neural networks (ANNs), support vector machines (SVMs), random forests (RF), and gradient boosting (GB) can ingest large, heterogeneous datasets and expose patterns that conventional models often miss.

Recent results underscore the momentum:

• Multi-layer perceptron (MLP) ANNs have achieved low RMSE in predicting PCE, open-circuit voltage (Voc), short-circuit current density (Jsc), and fill factor (FF).
• Explicitly feeding irradiance into neural nets lifts accuracy for Jsc and FF across both dim and bright conditions, reaching R² values above 0.85.
• Ensemble methods (RF, GB) trained on more than 6,000 PSC samples captured the light-intensity dependence of PCE and Voc with accuracies exceeding 90%.
• Hybrid workflows have trained portions of optoelectronic models on irradiance-dependent J–V data for planar p–i–n PSCs, delivering rapid, accurate predictions.

Beyond forecasting, these tools help reveal underlying mechanisms—from carrier dynamics and interface chemistry to degradation pathways—guiding smarter experiments and accelerating materials discovery.

The study: high-accuracy ANN for variable irradiance J–V

The reported work centers on an ANN that predicts the current density across the J–V curve under different irradiance levels. Trained and validated on comprehensive datasets, the model estimates current as a function of voltage and light intensity, then benchmarks its predictions against simulation-derived J–V curves and held-out test data. Performance is assessed across training, validation, testing, and overall splits to ensure generalization.

This approach offers several advantages:

• Speed and scale: Once trained, the model can generate full J–V curves across a range of irradiances in milliseconds, enabling sweeping scenario analysis.
• Cost efficiency: Reduces reliance on extensive experimental campaigns and specialized hardware for every new condition or design.
• Realism: Captures nonlinear performance shifts that occur as sunlight fluctuates through the day and across weather patterns.

What the broader literature says

ML has already made strides in optimizing PSCs via additive and interface engineering to improve carrier transport, exploring stability in lead-free systems, and co-optimizing tin-based devices with unified simulation–ML frameworks. In tandem structures, irradiance-aware modeling supports better energy yield under real skies rather than just standard test conditions. These efforts build confidence that data-driven methods can meaningfully guide chemistry, architecture, and process choices—before expensive iterations in the fab.

Why predicting under real-world light matters

Constant-illumination measurements—while useful—cannot capture the dynamic operating reality of solar cells. Cloud cover, angle of incidence, spectral shifts, and diurnal cycles all alter device response. The ANN-based framework addresses this by providing accurate, irradiance-aware J–V predictions that reflect how PSCs truly behave outdoors. That makes it easier to:

• Rank device architectures for field performance, not just lab peaks.
• Estimate energy yield more credibly across climates and seasons.
• Identify regimes where Jsc, Voc, or FF are most sensitive to light intensity and guide targeted improvements.

Implications for R&D and commercialization

Data-driven prediction shortens development cycles and reduces cost, helping researchers converge faster on high-performing recipes and structures. It can also be embedded into digital twins of PV modules or production lines to flag anomalies and anticipate degradation. Over time, coupling ANN predictors with high-throughput experimentation and physics-informed models could produce hybrid systems that are both interpretable and accurate—vital for bankability and certification.

Bottom line

By accurately modeling J–V characteristics of perovskite solar cells under variable irradiance, this ANN-driven approach delivers a practical bridge between lab excellence and field reliability. It equips researchers and engineers with a faster, scalable, and cost-effective way to evaluate real-world performance—an essential step toward robust, commercial-grade PSCs.