Driving Manufacturing Best Practices Using Multimodal AI

Abstract

Multimodal artificial intelligence offers promising solutions for enhancing operational excellence in contract manufacturing, where small job shops typically operate with limited standardization and high process variability. This research develops a part similarity tool that integrates geometric, material, and scale information to improve quoting accuracy and engineering efficiency in high-mix, low-volume production environments. After examining the fragmented manufacturing landscape and reviewing current AI applications in manufacturing, the study introduces an approach based on Variational Autoencoders for encoding 3D geometry alongside material properties and dimensional scale information. The technical implementation addresses challenges of multimodal fusion, missing data handling, and computational efficiency, while a qualitative ablation study demonstrates how this comprehensive approach outperforms single-modal methods in manufacturing relevance. Engineers benefit from improved insights for manufacturing planning, while estimators achieve more consistent cost predictions using the multimodal system. Reinforcement learning with human feedback provides a mechanism for continuous refinement, creating a framework that bridges geometric similarity with manufacturing context and reduces subjectivity in critical business processes. The research contributes both theoretical insights into multimodal learning and practical implementation strategies for standardizing operations in contract manufacturing environments.