AI tools accelerate the race toward next-generation solar cells

A research team from Peking University and Peking University Shenzhen Graduate School have used artificial intelligence (AI) to quickly and accurately predict the properties of materials that could improve solar energy devices. Their algorithms were able to predictic important properties such as: conduction band minimum (CBM), valence band maximum (VBM), and bandgap of halide perovskites.

This work provides valuable insights into the rational design of halide perovskites with tailored properties. These findings can impulse the discovery of better-performing solar materials, paving the way for more affordable and efficient solar panels.

One of the most popular way to produce electric energy is through the use of photovoltaic technology. It offers a sustainable and environmentally friendly way to generate electricity. Among various materials studied for solar cell applications, halide perovskites have garnered significant attention due to their remarkable photovoltaic properties, simple and low cost fabrication. These materials, characterized by their ABX₃ crystal structure, can incorporate various organic and inorganic constituents, which influence their optoelectronic properties such as bandgap, charge transport, and stability.

Recent advancements have pushed the power conversion efficiency (PCE) of perovskite solar cells to over 27%, with tandem configurations reaching above 30%, making them competitive with traditional silicon-based solar panels. Despite these successes, there are still challenges, including toxicity concerns related to lead content and issues with material stability. Addressing these challenges requires discovering new perovskite compositions with optimal properties, such as suitable bandgaps and energy level alignment, to improve performance and longevity.

Understanding and tailoring the electronic band structure of perovskite materials, particularly the bandgap, conduction band minimum (CBM), and valence band maximum (VBM), is critical to optimizing their performance. The bandgap determines the spectral range of sunlight that the material can absorb, whereas the alignment of CBM and VBM affects charge separation and transport efficiency. Precise control over these parameters is critical for reducing recombination losses while increasing device efficiency.

A key aspect of optimizing perovskite performance lies in understanding and engineering their electronic band structure. The bandgap, conduction band minimum (CBM), and valence band maximum (VBM) dictate light absorption, interfacial alignment, carrier dynamics and device efficiency. Precise control over these parameters is crucial for minimizing recombination losses and maximizing device efficiency.

Traditional experimental and theoretical approaches, like high-throughput screening and density functional theory (DFT) calculations, though effective, are often time-consuming and resource-intensive. Therefore, there is a growing need for efficient, data-driven strategies to accelerate the identification and design of promising perovskite materials. Machine learning (ML), with its ability to analyze large datasets and uncover complex patterns, has emerged as a powerful tool in this context, enabling rapid prediction of key properties and facilitating rational materials design for photovoltaic applications.

Data-driven ML is efficient, eco-friendly, and cost-effective, yet prior ML research has largely been limited to inorganic halide perovskites, lacking comprehensive predictions of CBM and VBM energy levels.

The Solution: A research team Peking University and Peking University Shenzhen Graduate School built high-accuracy ML models to predict CBM, VBM and bandgaps of halide perovskites, applicable to both inorganic halides and organic-inorganic hybrid halides.

The XGB (Extreme Gradient Boosting) model achieved test set R² = 0.8298 (MAE = 0.151 eV) for CBM, R² = 0.8481 (MAE = 0.149 eV) for VBM, and R² = 0.8008 (MAE = 0.285 eV) for bandgaps computed with the Heyd-Scuseria-Ernzerhof (HSE) hybrid functional. The XGB approach also delivered test set R² = 0.9316 and MAE = 0.102 eV on a larger set of bandgaps computed with the Perdew-Burke-Ernzerhof (PBE) functional. SHapley Additive exPlanations (SHAP) analysis of the optimal models identified the dominant chemical and structural features controlling these energy levels, providing practical design guidelines for tailoring halide-perovskite band structures.

The metrics R² and MAE are standard measures for assessing predictive performance: R² indicates the proportion of variance in the data explained by the model, with values closer to 1 representing better fits. MAE quantify the average prediction errors in electronvolts, with lower values reflecting higher accuracy. Overall, these results highlight the models’ strong ability to reliably predict complex electronic properties, facilitating accelerated discovery and rational design of new halide perovskite materials for photovoltaic applications.

The Future: A future direction is to leverage both the explainability of shallow ML models and the powerful learning capabilities of deep learning models to enable efficient and eco-friendly discovery of outstanding photovoltaic perovskite materials.

The Impact: This work developed a machine learning approach for predicting comprehensive energy band properties of halide perovskites, which accelerated the discovery of suitable photovoltaic materials and guides the rational design of high-efficiency solar cells.
The research has been recently published in the online edition of Materials Futures, a prominent international journal in the field of interdisciplinary materials science research.

Reference: Yucheng Ye, Runyi Li, Bo Qu, Hantao Wang, Yueli Liu, Zhijian Chen, Jian Zhang, Lixin Xiao. Machine learning for energy band prediction of halide perovskites[J]. Materials Futures. DOI: 10.1088/2752-5724/adeead

---

---