Physics-Distilled Neural Network enabled by Large Language Models for Manufacturing Process-Property Predictive Modeling

The paper presents a novel knowledge distillation framework that integrates analytical physics priors extracted via Large Language Models into a teacher model for predictive modeling in manufacturing processes. Utilizing a Graph-Masked Attention layer, the framework captures complex dependencies and distills this knowledge into a lightweight student model, achieving inference rates over 6000 Hz suitable for real-time applications. The approach demonstrates high predictive accuracy and robustness, even with limited data, making it significant for practitioners aiming to enhance predictive modeling in data-scarce manufacturing environments.