Learning To Focus Physics-Informed Deep Learning For Super-Resolved Ultrasonic Phased-Array Imaging H/F CEA

Ultrasonic phased-array imaging is a core technology in non-destructive testing (NDT) for detecting defects such as cracks or voids in industrial components. By electronically steering ultrasonic beams, phased arrays generate detailed 3D images of internal structures. The Total Focusing Method (TFM) is the standard reconstruction algorithm, achieving diffraction-limited resolution by coherently summing signals from all emitter-receiver pairs.
However, conventional TFM suffers from key limitations: its resolution is constrained by diffraction and array pitch, grating lobes degrade image quality, and it assumes uniform sound velocity. It also struggles to resolve sub-wavelength defects, limiting its effectiveness in complex or heterogeneous materials.
Recent deep learning methods have improved ultrasonic imaging through denoising and super-resolution, but most operate as black boxes without physical interpretability. They often fail to generalize across array geometries or material conditions.
This internship proposes a physics-informed deep learning framework that integrates physical modeling of ultrasonic propagation into neural architectures. Instead of static delay-and-sum focusing, the approach learns adaptive, reweighted focusing kernels that enhance resolution while maintaining interpretability.
The research is structured around six axes:
Reweighted TFM: learn per-pixel focusing weights through supervised or self-supervised training for adaptive, interpretable imaging.
Grating-lobe analysis: study array pitch effects and compare learned PSFs with theoretical models.
Tiny defect imaging: test the method on sub-wavelength defects using synthetic and experimental data.
Coded excitation: train models for artifact-free imaging under simultaneous transmit-receive schemes for faster acquisition.
Sound speed estimation: incorporate differentiable beamforming to jointly estimate material properties and focus adaptively.
Transformer-based characterization: use multi-angle scattering data and attention mechanisms for defect classification and interpretation.
Expected outcomes include a new interpretable deep model for ultrasonic imaging, quantitative grating-lobe suppression analysis, and demonstration of sub-wavelength defect detection.
This project bridges data-driven learning and physical modeling, leading to more robust, adaptive, and explainable ultrasonic imaging systems. The resulting framework could significantly enhance industrial inspection and structural health monitoring by achieving super-resolution, real-time imaging of complex materials.
Detailed research proposal here.