Mohammad Samin Nur Chowdhury

Mohammad Samin Nur Chowdhury recently defended his Ph.D. in Electrical and Computer Engineering at Purdue University, where he was advised by Prof. Charles A. Bouman and Prof. Gregery T. Buzzard. He has over eight years of research and industry experience in computational imaging, inverse problems, and machine learning, and has received multiple awards at leading international conferences, including ICIP, ICASSP, and WCNR. His doctoral research focused on developing physics-informed AI algorithms for hyperspectral neutron computed tomography and its extension to strain tensor reconstruction, integrating physics-based modeling with scalable computational methods for large-scale scientific imaging. Prior to Purdue, he worked at Microsoft on next-generation camera ISP pipelines and at UNAR Labs, where he developed computer vision technologies for assistive applications.

Abstract Title: Physics-Informed Algorithms for Hyperspectral Neutron Tomography and Strain Tensor Tomography

Abstract:

Hyperspectral neutron computed tomography (HSnCT) is a 3D imaging technique in which thousands of wavelength-specific neutron radiographs are acquired, enabling spectral characterization of materials. Recent advances in instrumentation, including the VENUS beamline at Oak Ridge National Laboratory, are significantly expanding the potential of HSnCT across diverse high-impact applications. However, the extremely low signal-to-noise ratio (SNR) limits the reliability of spectral analysis, while the massive data volume makes processing extremely time-consuming.

To overcome these challenges, we developed three algorithms built upon a common unsupervised machine learning framework termed Dehydration–Rehydration: Fast Hyperspectral Denoising (FHD), Fast Hyperspectral Reconstruction (FHR), and Fast Material Decomposition (FMD). These algorithms exploit the inherent low-dimensional spectral structure of neutron data while enforcing consistency with attenuation physics. FHD improves SNR by over 30 dB, and FHR and FMD achieve similar SNR gains with 10× faster computation.

Building on this foundation, we extended HSnCT to strain tomography and designed the Model-Oriented Neutron Strain Tomographic Reconstruction (MONSTR) algorithm for strain tensor recovery. To address the severe ill-posedness of strain tomography, MONSTR incorporates physics-based constraints directly into the inversion framework, enabling recovery of strain tensors with about 99% accuracy in controlled studies.