AI Framework Provides Mechanistic Reasoning for Corrosion Prediction

Corrosion causes significant global economic losses, and while machine learning can predict corrosion rates, it often lacks the ability to explain the underlying mechanisms. This mechanistic reasoning is crucial for safety-critical applications in materials engineering, where simple accuracy metrics are insufficient. To address this, a new domain-adapted retrieval-augmented generation (RAG) framework has been introduced. This framework fine-tuned open-weight language models (Llama-3.1-8B, Qwen-2.5-7B, Mistral-7B) on a large dataset of expert-verified question-answer pairs from peer-reviewed corrosion literature. Integrated with a hybrid dense-lexical retrieval pipeline, the system demonstrated substantial gains in Token F1 scores and achieved high faithfulness (0.964) and context recall (0.988). A novel "Reason Map" framework was also developed, which constructs directed evidence graphs from generated answers and retrieved literature. This allows for the systematic detection of causal direction inversions and unsupported inferences, which traditional factuality metrics often miss. The modular architecture is designed to be applicable across various engineering domains, offering a blueprint for trustworthy AI-assisted knowledge synthesis.