AI Algorithms for Perovskite Solar Cells Fabrication

The problem

Perovskite solar cells are among the most promising photovoltaic technologies, but finding stable, efficient compositions takes slow, expensive synthesis, and density functional theory (DFT) calculations are computationally heavy. Machine learning can screen thousands of candidates and predict their properties before anything is made.

My models

I developed models for perovskite solar-cell material screening: 99.3% accuracy predicting heat of formation, and 90.5% predicting bandgap.

What the talk covered

• A comparison of more than ten ML and deep learning methods across property identification, stability optimization and lead-free perovskites.
• Perovskites with the ABX3 structure, with models trained on DFT data: crystal structures, band gaps, elastic constants and decomposition energies.
• Gradient boosting, SVM, kernel ridge regression, KNN, neural networks, logistic regression, decision trees, random forests, CNNs and transfer learning.
• Applications beyond solar cells: memristors, luminescent materials and microfluidic synthesis.

Other results discussed in the talk

• About 90% accuracy with neural networks for crystal dimensionality, and about 86% with logistic regression for grain boundaries.
• A GBR, SVM and KRR ensemble narrowed 38,086 hybrid organic-inorganic perovskites to 686 viable candidates.
• Gradient boosting regression and classification identified 2,025 possible perovskites, 151 with band gaps suited to solar cells.
• SVC on 455 materials with 11 attributes predicted formability; KRR, KNN and SVR were trained on 354 DFT decomposition energies validated by experiment.