Session: CIE-06-01: AL/ML AI and ML for System Engineering

Paper Number: 191528

191528 - Physics-Distilled Neural Network With Large Language Models for Process-Property Mapping in Manufacturing Processes 

The accurate prediction of process-property relationships in manufacturing remains a fundamental challenge, particularly in high-precision domains with data scarcity due to the prohibitive temporal and financial costs of extensive experimentation. To address this problem, we propose a novel physics-distillation framework that bridges the gap between theoretical scientific knowledge and sparse experimental observations. This architecture utilizes a Large Language Model (LLM) as an automated knowledge extraction engine to synthesize analytical physics priors directly from existing scientific literature to eliminate the manual expertise burden typically required for model initialization. The core of our framework is a Privileged Teacher model that integrates these literature-derived priors via a specialized Graph-Masked Attention (GMA) layer. This layer ensures structural consistency by embedding physical partial derivatives and privileged features directly into the model’s latent manifold. Through a robust knowledge distillation process, the physical insights of the teacher are transferred to a lightweight Student Predictor to enable high-quality property mapping without the computational overhead of the full physical model. We evaluate this framework across three distinct manufacturing testbeds: FLIPMM, MSLA, and TADCR. Experimental results demonstrate that our framework significantly outperforms traditional regression and standard neural network baselines, particularly in extrapolative regimes. By achieving an inference frequency exceeding 6000 HZ, this work provides a feasible approach for mapping manufacturing relationships with a salient balance of prediction accuracy and real-time operational speed.

Presenting Author: Ge Song University of Connecticut

Presenting Author Biography: Post-doc fellow, University of Connecticut
Ph.D. University of South Carolina (2025)