Physical Inspection of Electronics · Hardware Assurance · Advanced Packaging Metrology and Failure Analysis
01 – Advanced Packaging Inspection & Metrology
SAM · X-ray · PINN · 2.5D/3D · Metrology · Security · Reliability
This research focuses on inspection and metrology of advanced packaging structures including chiplets, interposers, redistribution layers (RDLs), and fine-pitch interconnects, where 3D stacking and heterogeneous materials limit conventional imaging performance. The work develops physics-informed reconstruction methods that improve subsurface feature recovery from noisy and incomplete measurements. Physics-informed neural networks (PINNs) are used to embed wave propagation and attenuation constraints into Scanning Acoustic Microscopy (SAM) and X-ray reconstruction, enabling improved resolution and reduced artifacts under metal-induced scattering. Cross-modality fusion is applied to incorporate surface-sensitive signals into volumetric reconstruction, enhancing structural fidelity in multilayer packages. The research also establishes metrology and failure analysis workflows for identifying interconnect defects, alignment errors, and structural inconsistencies, with nano-probing used for localized validation of electrical and material behavior.
02 – LLM, VLM & Agentic AI for Physical Inspection
LLM · VLM · Agentic AI · RAG · AI Assistants
This research area focuses on advancing the physical inspection of electronics through the integration of multimodal inspection data with emerging data-driven methodologies. The work targets key inspection modalities including X-ray imaging, scanning electron microscopy (SEM), scanning acoustic microscopy (SAM), and optical inspection, enabling detailed analysis of complex device structures at both package and device levels. A central effort involves the development of retrieval-augmented inspection assistants based on large language models (LLMs), which leverage domain-specific knowledge bases, technical standards, and prior inspection records to support context-aware interpretation and decision-making. In parallel, vision-language models (VLMs) are used to detect imaging artifacts, interpret defect patterns, and provide structured operator guidance during inspection workflows. The research also explores agentic AI frameworks that transform laboratory procedures into structured QA datasets, enabling systems to reason over inspection data, identify failure modes, and recommend next steps in a consistent and reproducible manner. Additional efforts focus on defect taxonomy development, zero-shot recognition of previously unseen defect types, and LLM-assisted root cause analysis, with emphasis on capturing variations across materials, technologies, and manufacturing nodes. These contributions aim to establish scalable, explainable, and systematically validated inspection methodologies for modern microelectronics.
03 – Packaging Ecosystem, Manufacturing & Strategic Infrastructure
Chiplets · OSAT · Workforce · Roadmap
This research analyzes the structural and operational challenges in the semiconductor packaging ecosystem, with emphasis on the evolving role of OSAT facilities, supply chain dependencies, and system-level integration requirements. The primary contribution lies in identifying bottlenecks across the packaging value chain, including limited domestic advanced packaging capacity, fragmentation between design, fabrication, and assembly stages, and heavy reliance on offshore vendors for critical processes and materials . The work further examines the transition toward packaging-centric manufacturing models, where chiplet-based integration and heterogeneous assembly shift performance scaling from transistor-level design to system-level architecture, placing increased demand on packaging infrastructure and coordination . In addition, it outlines key requirements for strengthening the ecosystem, including expansion of OSAT capabilities, tighter integration between industry stakeholders, and development of supporting infrastructure such as materials, equipment, and skilled workforce.
04 – Counterfeit Detection, Hardware Security & Trust
Counterfeit · IC Authentication · OPTICAL · THz · Fingerprinting · Logo · Marking
The unifying goal is to identify tampering and fraud using features that adversaries cannot easily replicate or erase. This research develops non-destructive, physically grounded methods for detecting counterfeit and tampered integrated circuits by leveraging intrinsic package- and process-level features that are difficult to replicate. The primary contribution is the shift from surface-level inspection toward extracting stable physical fingerprints from manufacturing-induced variations, such as defect patterns, encapsulant properties, and fabrication artifacts. Techniques including thermo-reflectance imaging are used to localize defect-induced thermal signatures and generate inherent identifiers tied to package structure and process variability , while terahertz time-domain spectroscopy enables material-level fingerprinting through refractive index and layer property variations in encapsulated devices . In parallel, feature-centric approaches such as pinhole morphology analysis and laser marking characterization establish fabrication-locked signatures that expose resurfacing, re-marking, or process inconsistencies . The work further incorporates machine learning models for multi-modal feature extraction and classification, enabling scalable detection of counterfeit components under limited data and diverse inspection conditions .
05 – Medical Imaging & Broader AI Applications
Medical Imaging · Deep Learning · Segmentation · Survey
This research develops efficient deep learning frameworks for volumetric medical image analysis, with a focus on improving accuracy and computational scalability in 3D clinical settings. A primary contribution is the design of transformer-based architectures that integrate frequency-domain representations to balance global context modeling with fine-grained structural detail. For example, wavelet-driven models decompose volumetric data into low- and high-frequency components, enabling attention mechanisms to operate on compact representations while preserving boundary-level information critical for segmentation tasks . This approach significantly reduces model complexity and memory overhead compared to conventional 3D transformers, making it suitable for resource-constrained clinical environments. Additional contributions include the development of deep learning pipelines for automated detection and classification tasks in medical imaging, such as lymph node segmentation and vertebral fracture identification, where the emphasis is on improving robustness across heterogeneous data and enabling reliable deployment in real-world diagnostic workflows.
