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

Researchers have developed a physics-informed neural network framework that uses Large Language Models to extract scientific knowledge from literature, enabling accurate manufacturing predictions with minimal data. The lightweight student model achieves real-time inference speeds exceeding 6000 Hz while maintaining robust performance even when LLM-derived physics priors are incomplete.

This research addresses a critical challenge in industrial AI: predicting manufacturing outcomes when experimental data is scarce and interpretability is essential. By combining Large Language Models with physics-based knowledge distillation, the framework bridges theoretical understanding and practical machine learning, enabling systems that are both accurate and explainable in data-limited environments.

The approach leverages LLMs to systematically extract analytical physics principles from scientific literature, embedding these insights into a privileged teacher model that guides a smaller, faster student network. Graph-Masked Attention layers capture complex interdependencies between manufacturing variables, handling both static setpoints and dynamic temporal signatures. This architecture addresses a longstanding tension in industrial AI: complex black-box models often perform better but lack interpretability, while simpler physics-based models struggle with real-world complexity.

For manufacturing and industrial sectors, this development carries substantial implications. The framework demonstrates fault tolerance—maintaining predictive accuracy even when LLM-extracted physics knowledge is imperfect—which is critical for real-world deployment where perfect information is rarely available. The inference speed of 6000+ Hz enables genuine real-time edge deployment on standard industrial hardware, eliminating dependency on cloud infrastructure and reducing latency in time-sensitive manufacturing environments.

The validation across five diverse manufacturing processes suggests the framework generalizes beyond single-domain applications. Future development likely focuses on expanding LLM-physics integration across additional industrial sectors, improving the automated extraction of domain-specific knowledge from scientific literature, and integrating these predictions into closed-loop manufacturing control systems for adaptive production optimization.

• →LLM-extracted physics priors enable accurate manufacturing predictions in data-scarce scenarios through knowledge distillation
• →Student model achieves 6000+ Hz inference speed, enabling real-time edge deployment on standard industrial hardware
• →Framework demonstrates robust fault tolerance, maintaining accuracy even with incomplete or suboptimal physics knowledge
• →Graph-Masked Attention architecture captures complex interdependencies between static and dynamic manufacturing variables
• →Approach bridges interpretability and accuracy, addressing the black-box problem in industrial machine learning