Transformer Model Predicts Lead-Free Perovskite Solar Cell Performance
The study, published in Discover Applied Sciences, introduces a Physics-Informed Feature Tokenizer Transformer, or FT-Transformer, to address the complexity of optimizing multi-dimensional parameters in these photovoltaic devices. Tin-based perovskites, specifically methylammonium tin triiodide (MASnI3) and formamidinium tin triiodide (FASnI3), are considered promising pathways for solar cells with improved efficiency, yet their optimization remains difficult due to the interplay of various physical factors.
The core innovation of this work lies in integrating physical constraints directly into the transformer architecture to maintain scientific integrity. Unlike standard machine learning models that may produce predictions inconsistent with fundamental physics, the FT-Transformer incorporates custom monostaticity constraints. These constraints enforce known semiconductor properties, such as the principle that higher defect densities lead to lower device efficiency. By embedding these physical laws into the model, the researchers aim to ensure that the predictions remain physically plausible, even when extrapolating to new parameter combinations.
To validate the model’s reliability, the team employed a comprehensive 5-fold cross-validation procedure using various random seeds. This approach was specifically designed to avoid partitioning bias, a common issue in machine learning where data splitting can artificially inflate performance metrics. The results demonstrated that the FT-Transformer possessed good architectural stability, producing consistent cross-validated R-squared values across different validation folds. For the MASnI3 material, the model achieved an R-squared value of 0.8024 with a standard deviation of 0.0135, while for FASnI3, it achieved a higher R-squared of 0.9574 with a standard deviation of 0.0024.
The study also compared the FT-Transformer against classical ensemble models, such as Random Forest. While the Random Forest model exhibited better empirical R-squared values in some instances, it failed to maintain physical consistency in its predictions. The FT-Transformer, by contrast, prioritized physical consistency, ensuring that its outputs adhered to the expected behavior of semiconductor materials.
Further analysis using gradient-based inverse design and SHAP explainability methods revealed distinct sensitivities between the two perovskite materials. The findings indicated that FASnI3 exhibits stronger sensitivity to defect densities at the interfaces compared to MASnI3.
The authors conclude that their cross-validated benchmarking framework offers a comprehensive and physics-based suite of tools to accelerate the development of perovskite technology. By combining the predictive power of transformer architectures with the rigor of physical constraints, the study provides a robust method for navigating the complex parameter space of lead-free perovskite solar cells.
Subject of Research: Machine learning modeling of lead-free perovskite solar cells
Article Title: Physics informed transformer modeling of lead free perovskite solar cells
Article References: Abdullah, S., Jamee, A. N., & Ifty, M. S. M. (2026). Physics informed transformer modeling of lead free perovskite solar cells. Discover Applied Sciences. https://doi.org/10.1007/s42452-026-09855-9
Image Credits: AI Generated
DOI: 10.1007/s42452-026-09855-9
Keywords: perovskite solar cells, machine learning, transformer model, lead-free materials, physics-informed, photovoltaics, cross-validation, defect density, SHAP analysis, inverse design, Physics, informed
News Source: Denise Maddox. (November 10, 2026). Transformer Model Predicts Lead-Free Perovskite Solar Cell Performance. Scienmag.
We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.
Enter your email address to subscribe to this blog and receive notifications of new posts by email.
Email Address
Subscribe
Bioengineer.org © Copyright 2023 All Rights Reserved.
Login to your account below


Remember Me

Please enter your username or email address to reset your password.


Bioengineer.org © Copyright 2023 All Rights Reserved.

source