Model Card for Perovskite-R1

This model is a domain-specific large language model fine-tuned from QwQ-32B, specialized in the field of perovskite solar cells, particularly focusing on precursor additives. It is designed to assist researchers, engineers, and material scientists by providing knowledge, insights, and suggestions related to perovskite solar cell formulation, additive effects, and experimental design.

Model Details

Model Description

• Model type: Large Language Model
• Language(s) (NLP): English
• License: apache-2.0
• Finetuned from model: QwQ-32B

Uses

Direct Use

The model can be run using LLaMA-Factory or any other framework capable of calling large language models.

Out-of-Scope Use

This model is specifically designed for perovskite solar cell precursor additives and is not intended for use in other fields. Using it outside this domain may produce unreliable or inaccurate suggestions, and could lead to incorrect conclusions if applied to unrelated materials or chemical systems.

Bias, Risks, and Limitations

Bias: Model may reflect trends and materials overrepresented in published perovskite solar cell literature, underrepresenting less-studied additives.

Risks: Suggested experimental conditions are unverified; following them without proper lab validation may cause failed experiments or unsafe reactions.

Limitations: Not a substitute for expert judgment; may be inaccurate for novel additives or compositions; does not include latest research beyond training data.

Recommendations

Use the model as a reference for research support and hypothesis generation in perovskite solar cells. Where possible, verify suggestions experimentally and seek input from domain experts. Consider combining model outputs with peer-reviewed literature and standard lab protocols. Avoid relying solely on the model for safety-critical or regulatory decisions.

Training Details

Training Data

The model is trained on a curated collection of scientific literature, experimental datasets, and publicly available resources related to perovskite solar cell precursor additives. The dataset includes research articles and drug databases, focusing on synthesis, additive effects, and device performance. All training data has been uploaded and is documented for transparency and reproducibility.

Training Procedure

The model is trained using a transformer-based architecture optimized for scientific text. Training is performed on high-performance GPUs with gradient accumulation. Fine-tuning is conducted on curated perovskite precursor additive datasets.

Training Hyperparameters

• Training regime: During instruction tuning, we use QwQ-32B as the base model and apply LoRA on all weight matrices with rank 16, alpha 32, and dropout 0.1. Training is conducted in bfloat16 precision with a per-device batch size of 1 and gradient accumulation of 8, an initial learning rate of 1e-4 decays via cosine annealing with a 5% warmup, for a total of 10 epochs. FlashAttention2 is employed to improve efficiency and memory usage.

Evaluation

Testing Data, Factors & Metrics

Testing Data

Performance is assessed using a benchmark dataset relevant to perovskite solar cell precursor additives

Factors & Metrics

Evaluation focuses on domain-specific factors such as material composition suggestions and additive effects on device performance. Metrics include qualitative alignment with literature findings and consistency in benchmark predictions.

Results

Model performance was evaluated using a benchmark dataset related to perovskite solar cell precursor additives. Compared with other models on the same benchmark, it demonstrates superior alignment with literature findings and provides consistent, reliable suggestions for material composition and additive effects. While the results are based on this benchmark, they indicate strong potential for supporting research in perovskite solar cell chemistry.

Summary

This model is a large language model specialized in perovskite solar cell precursor additives. It can support research by providing literature-aligned suggestions for material composition and additive effects. Evaluated on a benchmark dataset, it shows superior performance compared to other models, indicating strong potential for accelerating research and hypothesis generation. Users are encouraged to verify all outputs experimentally and consult domain experts for critical decisions.

Environmental Impact

Training was conducted on a limited number of GPUs with standard energy consumption, and no significant environmental impact is expected.

Model Architecture and Objective

The model is a transformer-based large language model designed for scientific text understanding and generation in the domain of perovskite solar cell precursor additives. It adopts a decoder-only architecture and is fine-tuned using instruction tuning with LoRA to provide accurate and literature-aligned suggestions for material composition, additive effects, and device performance optimization.

Compute Infrastructure

Training was conducted on high-performance GPUs with bfloat16 precision, utilizing gradient accumulation and FlashAttention2 for efficiency. Users can run the model on GPUs with sufficient memory to accommodate LoRA-adapted weights.

Hardware

Training is performed on high-performance NVIDIA GPUs (e.g., H100 or equivalent). For inference, users should use GPUs with sufficient memory to load LoRA-adapted weights.

Software

The model was trained and can be run using Python (>=3.10), PyTorch, and LLaMA-Factory. Additional libraries such as Transformers and FlashAttention2 are used for optimized performance.

Citation

BibTeX:

@article{wang2025perovskite,
  title={Perovskite-R1: A Domain-Specialized LLM for Intelligent Discovery of Precursor Additives and Experimental Design},
  author={Wang, Xin-De and Chen, Zhi-Rui and Guo, Peng-Jie and Gao, Ze-Feng and Mu, Cheng and Lu, Zhong-Yi},
  journal={arXiv preprint arXiv:2507.16307},
  year={2025}
}