LLM tools in the prediction of the stability of perovskite solar cells

TL;DR

This paper investigates the use of Large Language Model tools like ChatGPT and DeepSeek to predict the degradation and stability of perovskite solar cells, demonstrating their ability to suggest models and provide accurate predictions despite incomplete physical information.

Contribution

It introduces a novel approach using LLM tools for PSC degradation prediction, highlighting their capacity to suggest models and justify predictions in an incomplete information setting.

Findings

LLM tools can suggest and justify PSC degradation models.

ChatGPT can produce accurate degradation trend predictions.

The approach integrates environmental data for time series analysis.

Abstract

Predicting degradation rates is an important task in the development of new perovskite solar cells (PSCs). In this paper, we explore the feasibility of solving this problem using Machine Learning models supported by LLM tools. We consider both the "lifetime" prediction of the device and the prediction of degree of its degradation at specific time intervals. We demonstrate the ability of common LLM tools (ChatGPT, DeepSeek) to suggest and justify prediction methods in a dialogue with the user under conditions of incomplete information about the physical models of PSC degradation and the influence of the environment, providing rather accurate prediction. The results cover the formation of time series of efficiency with a given architecture, calculated using various available mathematical models together with environmental characteristics archived in various meteorological databases (illumination, temperature, humidity, UV level). We conclude that ChatGPT currently has sufficient access to training samples, can suggest to PSC designer various PSC models in the literature, and has adequate solutions for predicting degradation trends.