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

Technical Summary: LLM Tools in the Prediction of the Stability of Perovskite Solar Cells

Problem Statement
The development of Perovskite Solar Cells (PSCs) is hindered by stability issues, specifically the degradation of Power Conversion Efficiency (PCE) over time. Predicting the lifetime (LT), often defined as T80 (time to 80% of initial efficiency), is critical but challenging. Traditional statistical Machine Learning (ML) approaches struggle due to a lack of representative long-term field data (PSCs require >25 years, but research history is short) and the difficulty of replicating complex outdoor environmental factors (weather, spectrum, load) in accelerated laboratory testing. Furthermore, existing multi-agent LLM systems for PSC design often lack integration with physically based degradation models and require significant programming expertise. This paper investigates whether standard Large Language Models (LLMs), such as ChatGPT and DeepSeek, can independently assist in predicting PSC degradation trends and lifetimes under conditions of incomplete information and data scarcity.

Methodology
The authors propose a framework where the LLM acts as a dialogue partner to a PSC researcher, guiding the construction of prediction models without requiring custom multi-agent submodels. The methodology relies on the following components:

1. Ontology Construction: The study defines a subject domain ontology for PSC lifetime prediction. This includes entities such as DegradationPhase, RecoveryPhase, ReversibleMechanism (e.g., ion migration), and IrreversibleMechanism. The LLM is queried to generate this ontology, ensuring it captures both monotonic degradation and non-monotonic behaviors (self-recovery).
2. Data Synthesis and Integration: Recognizing the scarcity of public, high-resolution field data for specific architectures, the authors utilize the LLM to generate "synthetic" time series. The LLM identifies relevant meteorological databases (e.g., DWD, ERA5, PVGIS) for specific locations (e.g., Berlin) and combines them with mathematical degradation models (Arrhenius, Eyring, Ornstein-Uhlenbeck stochastic processes) to simulate daily PCE trajectories.
3. Model Selection and Validation: The LLM is tasked with selecting appropriate ML algorithms (e.g., XGBoost, Random Forest, Hybrid XGB+OU models) and defining feature engineering strategies. It incorporates environmental parameters (irradiance, temperature, humidity, UV) and architectural details (layer composition) as inputs.
4. Dialogue-Driven Workflow: The process is conducted via a chat interface where the "user" (researcher) provides constraints (architecture, location, time interval), and the LLM suggests data sources, model architectures, and validation metrics (RMSE, MAE, $R^2$). The LLM also addresses the non-determinism of its own outputs by leveraging "temperature" settings to explore diverse degradation hypotheses.

Key Contributions

• Feasibility of Standalone LLMs: The paper demonstrates that state-of-the-art LLMs can independently identify subproblems in PSC degradation prediction (e.g., selecting between reversible and irreversible models) without the need for complex, custom-built multi-agent systems.
• Ontology-Driven Querying: The authors show that explicitly defining the subject domain ontology (including concepts like HysteresisEffect and RecoveryPhase) allows the LLM to retrieve relevant physical models and literature, even when specific datasets are not directly available.
• Synthetic Data Generation: The study validates the use of LLM-generated synthetic time series, derived from meteorological archives and physical degradation equations, as a viable substitute for scarce experimental data in training and testing prediction models.
• Handling Non-Monotonicity: The approach successfully incorporates reversible degradation mechanisms (self-recovery), which are often overlooked in standard monotonic degradation models but are critical for accurate outdoor lifetime prediction.

Results
The paper presents a case study predicting the PCE of a specific PSC architecture (ITO | 2PACz | Cs₀.₁₅FA₀.₈₅PbI₂.₅₅Br₀.₄₅ | C₆₀ | SnO₂ | Cu) in Berlin for the period of September–November 2023.

• The LLM successfully located and integrated data from multiple sources (DWD, PVGIS, and a Zenodo Perovskite Ageing dataset) to construct the necessary input features.
• Using a hybrid model (XGBoost + Ornstein-Uhlenbeck process), the system generated time series that aligned with observed degradation trends, including seasonal fluctuations and recovery periods.
• The model demonstrated acceptable accuracy in predicting T80 over both short (1-month) and long (1-year) intervals, correctly capturing the direction of degradation trends despite the use of synthetic training data.
• The LLM provided justifications for model choices and highlighted the uncertainty surrounding specific degradation mechanisms (e.g., ion migration), offering a list of debated topics and references.

Significance and Claims
The authors conclude that current LLM tools possess sufficient access to training samples and literature to suggest and justify prediction methods for PSC designers, even with incomplete physical models. The paper claims that ChatGPT can:

1. Suggest various PSC models found in literature.
2. Provide adequate solutions for predicting degradation trends by synthesizing environmental data and physical models.
3. Act as a robust tool for overcoming the "small data" problem in PSC research by generating physically plausible synthetic datasets.

The significance lies in the ability of these tools to support the PSC design process by identifying subproblems and offering solutions that are explicitly or implicitly absent from the initial user queries, thereby accelerating the evaluation of PSC stability without the immediate need for extensive, costly long-term field testing. The paper maintains a modest tone, acknowledging that while LLMs can guide the process and provide accurate trend predictions, the non-determinism of answers and the ongoing debates in the scientific community regarding specific degradation mechanisms require careful user interpretation and validation.