Descriptor Design for Perovskite Material with Compatible Molecules via Language Model and First-Principles.

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Direct application of large language models (LLMs) in material design faces challenges in achieving experimental validation accuracy.
• Accurate prediction of material properties is crucial for real-world applications.

Purpose of the Study:

• To develop a novel multimode descriptor design method for enhanced materials prediction and analysis.
• To improve the experimental validation accuracy for predicting material properties, specifically aqueous photocurrents of perovskite materials.

Main Methods:

• Integration of a natural language processing (NLP) literature model with density functional theory (DFT) calculations.
• Utilizing a genetic algorithm (GA) to assist in descriptor design and model optimization.
• Case study involving the prediction of aqueous photocurrents for engineered halide perovskite (CH3NH3PbI3).

Main Results:

• Achieved an unprecedented experimental validation accuracy of 87.5% for predicting perovskite aqueous photocurrents using the GA-assisted multimode descriptors.
• Demonstrated significantly higher accuracy compared to common machine learning models (50% accuracy).
• Developed an accurate "white-box" model for perovskite stability prediction (90.2% test accuracy, 92.3% train accuracy) using genetic algorithms.

Conclusions:

• The proposed multimode descriptor design route offers a feasible and accurate method for predicting complex material properties.
• The combination of LLMs and DFT calculations, optimized by GA, enhances predictive accuracy and provides insights into material behavior (e.g., cation···π interactions, crystallization).
• This approach facilitates a deeper ontological and conceptual understanding of molecule-modified halide perovskite materials.