JH976/Perovskite-RL

TEXT GENERATIONConcurrency Cost:2Model Size:32BQuant:FP8Ctx Length:32kPublished:May 3, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

JH976/Perovskite-RL is a 32 billion parameter, domain-adapted large language model based on Qwen3-32B, specifically fine-tuned for perovskite solar-cell additive engineering. It is trained to reason about additive molecules, defect passivation, and various stability-related mechanisms in perovskite photovoltaics. This model excels at literature-based reasoning and molecular additive hypothesis generation within this specialized scientific domain. It was developed by JH976 and optimized using supervised fine-tuning and GRPO reinforcement learning.

Loading preview...

Perovskite-RL: Domain-Adapted LLM for Perovskite Engineering

Perovskite-RL is a specialized 32 billion parameter language model, built upon the Qwen3-32B architecture, designed for the complex domain of perovskite solar-cell additive engineering. It has undergone a unique training pipeline involving supervised fine-tuning (SFT) and GRPO reinforcement learning to enhance its reasoning capabilities in this field.

Key Capabilities

  • Mechanistic Reasoning: Trained to understand and reason about additive molecules, defect passivation, crystallization modulation, interfacial protection, ion migration, electronic effects, and stability mechanisms in perovskite materials.
  • Additive Discovery Workflow: Integrates into a closed-loop discovery workflow for perovskite precursor additive identification, connecting literature-derived reasoning with additive candidate generation.
  • Specialized Training Data: Fine-tuned on 90,749 SFT examples and 5,800 GRPO examples of curated perovskite-additive reasoning data, including literature-derived mechanism and molecular-property reasoning.
  • Reinforcement Learning Optimization: Utilizes GRPO with reward signals focused on answer correctness, format compliance, content recall, and reasoning quality for mechanism-aware additive selection.

Performance

On a mechanism-consistency benchmark, Perovskite-RL achieved 78.1% accuracy (25/32), demonstrating its ability to identify paper-specific mechanistic explanations.

Intended Use Cases

  • Perovskite additive mechanism analysis
  • Molecular additive hypothesis generation
  • Mechanistic descriptor generation
  • Literature-based reasoning for perovskite photovoltaics
  • Assisting computational screening workflows

Limitations

It is crucial to note that Perovskite-RL is a research tool and not a substitute for experimental validation. Generated suggestions may be chemically invalid or experimentally unsuitable, and the model may overstate mechanistic confidence. Outputs should be treated as hypotheses, not final scientific conclusions.