LLM-Extracted Physics Priors Enable Robust, Real-Time Process-Property Prediction Under Data Scarcity
This paper introduces a knowledge distillation framework that uses LLMs to systematically extract physics-based analytical priors from scientific literature, encoding them into a privileged teacher model via a Graph-Masked Attention layer. The teacher's structured physical knowledge is then distilled into a lightweight student model capable of real-time inference. Evaluated across five manufacturing processes using repeated K-fold cross-validation, the framework achieves strong predictive accuracy even when LLM-derived priors are incomplete, and the student model exceeds 6000 Hz inference frequency — making it viable for edge deployment on standard industrial hardware.