Conference Agenda

Artificial intelligence (AI) is frequently displayed as a mysterious process of computers gaining human-like reasoning capabilities. In this talk, I will give an introduction to the mathematical foundations of AI and machine learning by describing it as a function approximation problem. I will focus on machine learning methods for image reconstruction and analysis, and discuss particular challenges in developing such methods in the area of THz Imaging.

This talk does not requires prior knowledge in the field of machine learning.

THz spectral region of the electromagnetic spectrum offers distinct advantages due to its non-destructive and non-ionizing nature, which makes it suitable for applications in biomedicine, security screening, and quality control. Such favourable characteristics have catalyzed intensive research efforts, advancing THz imaging technologies, notably through time-domain spectroscopy (TDS) platforms and demonstrated near-field [1], synthetic aperture techniques [2], ghost imaging through scattering media [3,4], which provide unique capabilities largely unavailable at other spectral ranges. Despite these innovations, state-of-the-art TDS depend on femtosecond laser-driven THz sources, which introduce considerable experimental complexity and pose limitations regarding source brightness.

Recent progress shows that continuous-wave (CW) THz radiation can be effectively generated and detected via compact, cost-efficient electronic components. However, most CW-THz imaging approaches are limited to measuring only the intensity distribution of the target sample. From the imaging point of view, phase information, while not typically captured, is crucial for enhancing contrast in transparent samples and extracting valuable structural insights such as topographical and refractive index variations within the sample [5]. Within the context of CW-THz phase imaging, digital holography [6] and ptychography [7] have been widely explored; nonetheless, these techniques are often constrained by the reference arm's instabilities, introducing phase noise and reducing image reconstruction fidelity.

Here, we propose a reference-free THz phase retrieval methodology that integrates seamlessly with standard CW THz imaging setups, requiring no substantial hardware alterations. We demonstrate two experimental configurations involving the acquisition of multiple transmitted intensity patterns through the sample. Using a nonlinear convex optimization scheme, we reconstruct the phase profile of a given imaging sample, thereby significantly enhancing imaging performance without increasing system complexity.

We introduce two imaging configurations for recording transmitted intensity patterns. In the first setup, a lensless configuration is employed, where the detection begins at a predefined distance from the sample. The diffracted THz field is then captured sequentially at several axial positions, each separated by a fixed interval. The second configuration incorporates a 4f imaging system, wherein wavefront evolutions are obtained at various planes symmetrically distributed around the focal plane, maintaining a uniform axial spacing. In both imaging configurations, the propagation of the THz wavefield is modelled using the angular spectrum method as a solution to the Rayleigh–Sommerfeld diffraction integral. Phase retrieval is then cast as a nonlinear convex optimization framework to estimate the complex field at the object plane by minimising the discrepancy between the propagated fields and the measured intensities across all planes. This optimization routine is solved using a gradient descent algorithm supported with second-order information from the Hessian matrix to guide the search direction. A backtracking line search strategy is employed to adaptively determine the step size at each iteration, ensuring robust convergence and enhanced stability during phase reconstruction.

Our complete theoretical and experimental analysis will compare reconstruction fidelity, algorithmic performance, and associated challenges in both imaging configurations and will be presented in detail.

Reference:

[1] F. Blanchard, K. T. Improving Time and Space Resolution in Electro-Optic Sampling for near-Field Terahertz Imaging, Optics Letters 2016, 41 (20), 4645–4648.

[2] Kumar, V.; Mukherjee, P.; Valzania, L.; Badon, A.; Mounaix, P.; Gigan, S. Fourier Synthetic Aperture-Based Time-Resolved Terahertz Imaging. Photonics Research 2024, 13 (2), 407–416. https://doi.org/10.1364/PRJ.544076.

[3] Kumar, V.; Cecconi, V.; Peters, L.; Bertolotti, J.; Pasquazi, A.; Totero Gongora, J. S.; Peccianti, M. Deterministic Terahertz Wave Control in Scattering Media. ACS Photonics 2022, 9 (8), 2634–2642. https://doi.org/10.1021/acsphotonics.2c00061.

[4] Kumar, V.; Cecconi, V.; Cutrona, A.; Peters, L.; Olivieri, L.; Totero Gongora, J. S.; Pasquazi, A.; Peccianti, M. Terahertz Microscopy through Complex Media. Sci Rep 2025, 15 (1), 11706. https://doi.org/10.1038/s41598-025-95951-6.

[5] Petrov, N. V.; Perraud, J.-B.; Chopard, A.; Guillet, J.-P.; Smolyanskaya, O. A.; Mounaix, P. Terahertz Phase Retrieval Imaging in Reflection. Opt. Lett., OL 2020, 45 (15), 4168–4171. https://doi.org/10.1364/OL.397935.

[6] Huang, H.; Wang, D.; Rong, L.; Panezai, S.; Zhang, D.; Qiu, P.; Gao, L.; Gao, H.; Zheng, H.; Zheng, Z. Continuous-Wave off-Axis and in-Line Terahertz Digital Holography with Phase Unwrapping and Phase Autofocusing. Optics Communications 2018, 426, 612–622. https://doi.org/10.1016/j.optcom.2018.06.011.

[7] Valzania, L.; Feurer, T.; Zolliker, P.; Hack, E. Terahertz Ptychography. Opt. Lett., OL 2018, 43 (3), 543–546. https://doi.org/10.1364/OL.43.000543.

Terahertz (THz) imaging is a promising nondestructive evaluation (NDE) technique for glass-fiber reinforced polymer laminates. While THz imaging excels in detecting subsurface defects in composites, the interpretation often requires confirmation from supplementary methods such as X-ray micro-computed tomography (μCT) to definitively identify barely visible impact damage (BVID). To address this challenge, we combine pulsed THz time-of-flight tomography (TOFT) with convolutional neural network (CNN), classifying B-scan images for damage identification.

Unlike A- or C-scans, B-scans capture the cross-section, thus suitable for observing multilayered damage indicators in BVID cases. A transfer-learning approach is implemented using CNN architectures trained on experimentally acquired THz B-scans labeled according to X-ray μCT imaging as ground truth. The use of transfer learning in NDE and composite material analysis improves classification accuracy and generalization, particularly in scenarios with limited training samples. The trained models achieve > 95 % classification accuracy in detecting damaged regions, demonstrating robust performance across CNN classifiers. Once the CNN model is properly trained, the method requires only THz B-scans, eliminating the need for supplementary X-ray μCT validation. This integration of deep learning significantly advances the viability of THz inspection systems for industrial applications.

In recent years, additive manufacturing (AM) has proven to be a flexible technique for the manufacturing of three-dimensional objects, including polymer-based materials [1]. Key AM technologies include Fused Deposition Modelling (FDM), Stereolithography (SLA), and Selective Laser Sintering (SLS), each with unique advantages. As AM technologies evolve, the need for robust identification and authentication methods becomes crucial. Digital fingerprints embedded within components are expected to play a vital role in smart production lines, anti-counterfeiting and logistics [2]. They provide identification over the whole lifecycle of the component, resist tampering and enhance security and traceability, enabling manufacturers to trace individual units and quickly identify anomalies or defects, thus contributing to overall quality control.

In this work, we aim to embed structures in situ beneath the surface of polymer-based components using three major AM techniques and image them using terahertz (THz) frequency domain spectroscopy. We demonstrate the feasibility of embedding these structures during the printing process and validate the measurement method. This research is the basis for the development of robust and reliable tagging mechanisms in additive manufacturing.

Reference:

[1] Saleh Alghamdi, Saad, et al. "Additive manufacturing of polymer materials: Progress, promise and challenges." Polymers 13.5 (2021): 753.

[2] Sola, Antonella, et al. "How can we provide additively manufactured parts with a fingerprint? A review of tagging strategies in additive manufacturing." Materials 15.1 (2021): 85.

Terahertz time-of-flight tomography enables nondestructive evaluation of layered media by analyzing time delays between internal echoes. For thin layers, temporally overlapping echoes complicate stratigraphic reconstruction, necessitating deconvolution to resolve discrete interfaces. This abstract compares the deconvolution commonly with and regularization for a woven glass fiber reinforced polymer composite laminate with delamination. The sample contains eight plies with a total thickness of 1.85 mm. The delamination is introduced to the sample with a Teflon disk with thickness of 250 m.

The deconvolution for the sample is usually analysed with regularization; however, the computation complexity is relatively high when compared to other regularizations such as . Here we compare the result based on orthogonal matching pursuit with other deconvolution methods to provide guidance for better choice of technique for deconvolution such as sparse deconvolution and autoregressive extrapolation. With the help of deconvolution method, the reflections from plies and delamination can be identified easier.