Ontology-guided attribute learning to accelerate certification for developing new printing processes

Abstract

Identifying printing defects is vital for process certification, especially with evolving printing technologies. However, this task proves challenging, especially for micro-level defects necessitating microscopy, which presents a scalability barrier for manufacturing. To address this challenge, we propose an attribute learning methodology inspired by human learning, which identifies shared attributes among seen and unseen objects. First, it extracts defect class embeddings from an engineering-guided defect ontology. Then, attribute learning identifies the combination of attributes for defect estimation. This approach enables it to recognize previously unseen defects by identifying shared attributes, even those not included in the training dataset. The research formulates a joint optimization problem for learning and fine-tuning class embedding and ontology and solves it by integrating natural language processing, metaheuristics for exploration and exploitation, and stochastic gradient descent. In a case study involving a direct-ink-writing process for creating nanocomposites, this methodology was used to learn new defects not found in the training data using the optimized ontology. Compared to traditional zero-shot learning, this ontology-based approach significantly improves class embedding, outperforming transfer learning in one-shot and two-shot learning scenarios. This research represents an early effort to learn new defect concepts, potentially reducing the need for extensive measurements in defect identification.