SPIN Group
Research
From physics-based sensing to interpretable machine learning: turning raw sensor data into decisions about material health.
Our Research Vision
The SPIN Group develops computational methods that transform raw sensor data into actionable knowledge about material health. Our research sits at the intersection of physics-based sensing, signal processing, and machine learning, enabling intelligent, automated inspection systems that scale across modalities and applications.
Research Themes
Three interconnected pillars that define our approach to computational NDE
Physics-Based Sensing & Signal Processing
We acquire and process data from terahertz, ultrasonic, and X-ray modalities. Our focus is on extracting physics-informed features, time-of-flight, spectral content, attenuation maps, that capture the signatures of subsurface damage.
Machine Learning for NDE
We develop weakly supervised and self-supervised learning frameworks that reduce reliance on manual annotation. Our pipelines combine classification, localization, and quantification, from Grad-CAM pseudo-labels to physics-informed feature maps.
Cross-Modal Validation & Deployment
We benchmark AI-driven inspection across modalities using X-ray ground truth, and build end-to-end pipelines designed for real-world deployment, from laboratory prototyping to field-ready systems.
Sensing Modalities
Multi-modal data acquisition and cross-validation for robust damage characterization
Terahertz (THz)
Time-domain spectroscopy & imaging
Ultrasonics
Phased-array & guided wave inspection
X-Ray
Radiographic ground truth & validation
Application Domains
Sectors where intelligent nondestructive evaluation delivers safety and reliability
Featured Projects
Selected ongoing and recent research
BVID Detection in Woven GFRP Laminates
Transfer learning with a DenseNet-121 backbone classifies barely visible impact damage (BVID) in woven glass-fiber composite plates directly from terahertz time-of-flight B-scans. Ground truth is established by X-ray micro-computed tomography on a controlled low-velocity impact test bed, providing a reproducible benchmark for downstream localization and severity studies.
■DenseNet-121 accuracy = 99.1%
Status: publication link to be added
Physics-Guided Weakly Supervised Damage Localization
A modality-agnostic pipeline pretrains on synthetic data from coupled Abaqus/Explicit impact simulations and MEEP electromagnetic forward models. Gradient-based saliency is converted into pseudo-labels and a quantitative severity output, removing the need for pixel-level human annotation. The same network transfers without modification from terahertz to X-ray imagery of impacted GFRP laminates.
■Sim-to-real transfer · zero retraining
Status: OTST 2026 abstract submitted · manuscript in preparation
Multi-Task Subsurface Profilometry from FDTD-Synthetic THz
A three-tier multi-task network trained on one-dimensional FDTD simulations jointly predicts defect presence, interface index, ply count, defect thickness and diameter, and per-ply thicknesses with uncertainty-weighted losses. A calibrated physics-based predictor lifts interface localization accuracy from ≈14% to 91.6%, and the full pipeline transfers cleanly to an external GFRP benchmark.
■Interface accuracy 14% → 91.6%
Status: manuscript in preparation
Research Directions
Where the group is heading next
Multimodal Sensing
Fusing terahertz, ultrasonic, and X-ray data so each modality compensates for the blind spots of the others.
Physics-Informed Machine Learning
Embedding wave physics, material models, and simulation priors directly into learning architectures for trustworthy, generalizable NDE.
Weakly Supervised Defect Localization
Pixel-level damage localization from scan-level labels, using saliency, pseudo-labels, and simulation-driven supervision.
Real-Time Inspection Systems
Algorithms and architectures that move diagnosis from offline analysis to in-situ, real-time decision making.
Digital Twins for NDE
Simulation-backed digital replicas of structures and sensing systems for predictive maintenance and structural health monitoring.
Cybersecure Intelligent Inspection
Security frameworks for sensing and data pipelines that protect the integrity of diagnostic decisions in deployed systems.
Useful Links
Resources, tools, and references for our research community