LLM System for Explainable LPBF Defect Analysis

Enhanced Key Takeaways

• •The system addresses the 'black box' nature of traditional deep learning defect detection by utilizing a Knowledge Graph (KG) to provide explainable, causal reasoning for why a specific defect occurred, rather than just identifying it.
• •The multimodal architecture employs a frozen vision encoder (typically CLIP-based) mapped to a Large Language Model via a linear projection layer, specifically fine-tuned on a domain-specific dataset of LPBF process parameters and micrograph images.
• •The framework incorporates a retrieval-augmented generation (RAG) pipeline that queries the ontology to ground LLM responses in metallurgical principles, reducing hallucinations in high-stakes manufacturing environments.

Technical Deep Dive

• •Architecture: Combines a Vision Transformer (ViT) encoder for image feature extraction with a decoder-only LLM (e.g., Llama-3 or Mistral derivative) via a cross-modal alignment layer.
• •Ontology Structure: Utilizes a hierarchical taxonomy covering defect morphology (e.g., lack of fusion, keyhole porosity, balling), process parameters (e.g., laser power, scan speed), and material properties.
• •Inference Pipeline: Implements a two-stage process: (1) Image-to-text semantic embedding alignment for initial classification, (2) KG-based reasoning module that injects causal context into the LLM prompt for mitigation strategy generation.
• •Dataset: Trained on a curated corpus of academic literature and industrial inspection reports, utilizing synthetic data augmentation to balance underrepresented defect classes.

Future ImplicationsAI analysis grounded in cited sources

Timeline