Knowledge Base

Safdar, M. (2026). Toward machine learning assisted directed energy deposition of Ni-WC metal matrix composites. Thesis.

Laser powder directed energy deposition (DED) additive manufacturing (AM) allows material deposition on complex geometries for hybrid manufacturing and repair. It supports tailored compositions such as nickel tungsten carbide (Ni-WC) metal matrix composites (MMCs), enhancing properties like wear resistance. However, complex thermal conditions often induce co-occurring microstructural defects, for which predictive first-principles models are inadequate. This thesis advances machine learning (ML) methods for defect detection and microstructural analysis in DED-based MMCs to accelerate process development. The application of ML in DED Ni-WC MMCs faces in-process, post-process, and deployment challenges. This work addresses all three using tailored methodologies. In-process challenges included monitoring deposition phenomena and detecting co-existing anomalies. ML models were trained using datasets from single tracks. Video-based deep learning (DL) models were fine-tuned on melt pool monitoring data from dual camera system, and explainable feature selection and fusion were applied for defect classification. Post-process efforts involved semantic segmentation of microstructure images using both optical and electron microscopy to quantify anomalies and support high-throughput analysis. Deployment focused on bridging the gap between ML and AM, identifying roles, systems, and requirements for real-world implementation. A key challenge in defect detection was the extraction of ground truth from characterization data. DL models were developed for automated segmentation of optical metallography. A fusion method combining convolutional and transformer model predictions was proposed, leveraging their respective strengths on majority and minority classes. The approach achieved 93% accuracy across datasets and drastically reduced processing time. In a second study, DL models were developed for scanning electron microscopy (SEM) to address optical limitations in capturing reinforcement degradation. Vision transformers (UPerNet, SegFormer) outperformed convolutional models in detecting carbide-related defects, with UPerNet recommended for high accuracy and SegFormer offering efficiency. To capture melt pool dynamics, an experimental setup with dual mid-wave IR (MWIR) cameras (CLAMIR and FLIR) was integrated. Eighty samples were deposited under varied process parameters. A transformer model (VideoMAE) was fine-tuned to extract spatiotemporal features from videos, surpassing image-based methods in distinguishing process states. Explainable AI techniques identified effective MWIR features for classifying the extent of six defect categories, with feature reduction improving model performance. FLIR features proved robust to noise, while CLAMIR performed better under instability, validating multi-camera fusion for practical use. A novel framework that links microstructure informatics with a range of expert characterization knowledge using customized and hybrid vision-language representations (VLRs) was proposed. Finally, an MLOps framework was proposed to integrate ML models into industrial AM pipelines. Fundamental requirements were outlined to guide deployment. A modular tool, MicroSegQ+, was developed for microstructural quantification, and a web tool, DeepBead, was built for defect prediction from process data using developed ML and DL models. A modular tool, MK-Link, was proposed to support fully automated post-process analysis by complementing MicroSegQ+ and by linking segmented microstructure with the expert characterization criteria. This work establishes a comprehensive ML-enabled framework for process monitoring, defect prediction, and microstructural characterization in DED-based processing of Ni-WC MMCs. The results provide foundation for future research directions on real-time defect mitigation and multi-layer monitoring in advanced AM systems.

Read the publication