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

2026-06-09

AI Summary

This work addresses the challenges in predicting process–performance relationships in manufacturing—namely high experimental costs, data scarcity, and poor interpretability of black-box models—by proposing a knowledge distillation framework that integrates large language models (LLMs) with graph neural networks. The approach first leverages an LLM to extract physics-based priors from scientific literature, constructing a teacher model enhanced with a graph masked attention mechanism to capture complex physical dependencies among variables. Knowledge is then distilled into a lightweight student model capable of efficient inference. Despite limited training data, the method achieves high accuracy and robustness while demonstrating fault tolerance to incomplete physical priors. Validated across five distinct manufacturing processes, the student model attains inference rates exceeding 6 kHz, enabling real-time edge deployment on standard industrial hardware.

Abstract

Predicting process-property relationships in manufacturing is often challenged by high experimental costs and the limited interpretability of complex 'black-box' models. This paper proposes a novel knowledge distillation framework designed to achieve high-accuracy predictions in data-scarce scenarios. The framework integrates analytical physics priors, which are systematically extracted from scientific literature via Large Language Models, into a privileged teacher model. We employ a Graph-Masked Attention layer to capture the complex physical dependencies among input variables showing strict setpoints or a combination of static and high-frequency temporal signatures. This privileged knowledge is distilled into a lightweight student predictor for inference. The feasibility and robustness of the framework are evaluated through a comprehensive experiment across five diverse manufacturing processes. To ensure statistical reliability, given the small dataset sizes, a repeated K-fold cross-validation technique is employed to quantify model stability and generalization. Results indicate that the proposed framework consistently achieves high predictive accuracy across all evaluated domains. Most importantly, the architecture demonstrates significant fault tolerance by maintaining robust predictive performance even in scenarios where LLM-derived analytical priors are suboptimal or incomplete. Furthermore, the student predictor achieves an inference frequency exceeding 6000 Hz, which facilitates real-time edge deployment on standard industrial hardware. This work provides a scalable solution for bridging the gap between theoretical physics and real-time industrial monitoring in data-limited environments.

Problem

• process-property prediction
• data-scarce manufacturing
• black-box models
• high experimental cost
• predictive modeling

Innovation

• Physics-Informed Machine Learning
• Knowledge Distillation
• Large Language Models
• Graph-Masked Attention
• Real-Time Edge Inference