Automated Machine Learning for Optical Surface Inspection: Neural Architecture Search and Hyperparameter Optimization for Self-Designing Inspection Networks

Abstract

Training high-quality deep learning models for optical surface inspection requires large diverse datasets, but manufacturing companies are often unwilling or legally prohibited from sharing raw measurement data across facilities due to competitive concerns, customer confidentiality requirements, and data privacy regulations such as GDPR. This data silos problem means that each facility trains on its own limited data, producing models that do not generalize well to other factories, product variants, or measurement equipment configurations. This study proposes a federated learning framework for optical surface inspection that enables multiple manufacturing facilities to collaboratively train a shared inspection model without any facility revealing its raw measurement data to any other facility or to a central server. Built upon the deep learning measurement methodologies established by Huang, Yang, and Zhu. (2023) in 4D thermal imaging and the optical metrology innovations of Huang, Tang, Liu, and Huang (2026), the framework employs differential privacy mechanisms to provide formal privacy guarantees, a Federated Averaging protocol optimized for optical measurement data distributions, and a contribution scoring system that rewards facilities for sharing informative data without exposing what that data contains. Evaluated on a simulated federation of six manufacturing facilities with heterogeneous data distributions, the proposed framework achieves inspection accuracy that is within 6.3% of a centrally trained model while providing formal (ε, δ)-differential privacy guarantees. The framework is demonstrated to enable factories with as few as 500 labeled samples to benefit from the collective knowledge of the entire federation, reaching performance that would otherwise require tens of thousands of samples. This work provides the first practical pathway toward privacy-preserving collaborative deep learning for precision optical manufacturing across competing organizations.