OPVPerfPredictor

The fine-tuned adapters in this repository are based on the TinyLlama-1.1B model, a lightweight and efficient language model well-suited for low-resource tasks. The adapters were fine-tuned using natural language descriptions of organic solar cells (OSCs) to predict key performance metrics:

• Power Conversion Efficiency (PCE)
• Short Circuit Current Density (JSC)
• Open Circuit Voltage (VOC)
• Fill Factor (FF)

Required Packages

To use this repository, make sure the following packages are installed:

• torch
• transformers
• peft
• accelerate
• bitsandbytes

Usage Instructions

• Open the notebook predPCE.ipynb located in the prediction/ folder.
• The notebook reads data.json — update this file to include new samples for prediction.
• Enter your Hugging Face token when prompted.
• The model will load the LoRA adapter and predict the PCE.

The model will automatically use GPU if available. If GPU is unavailable or results in an OutOfMemoryError, it will fall back to CPU.

Predicting FF, VOC, or JSC

To predict other performance metrics (FF, VOC, or JSC), simply:

• Use the same prompt format.
• Update the adapter model path in the notebook

For technical details, please refer to the accompanying paper: Link to the paper

License

This project is licensed under the GT License.