Invited Speakers

His current position is Research Professor at IKERBASQUE, Basque Foundation for Sciences in Bilbao and a Distinguished Professor at the Department of Analytical Chemistry, University of Basque Country, Spain. His current research interests include hyperspectral and digital image analysis and the application of Chemometrics (i.e., Machine and Deep Learning). He has authored over 190 publications (180+ peer-reviewed papers, books, book chapters, proceedings, etc.) and has given over 80 conferences and courses at international meetings. Jose has supervised or is currently supervising several MSc, PhD and Post Docs, and he is an editorial board member of four scientific journals within chemometrics.

Shortwave infrared hyperspectral imaging (HSI-SWIR) is a line-mapping technique that, until now, has been difficult to use for studying the external volumetric shape of objects.

Here, we introduce a new method that combines HSI-SWIR with Structure from Motion (SfM) algorithms to enable 3D projection and analysis of volumetric samples. The proposed framework, called 3D-HSI-SWIR, uses spatial reconstruction from multiple viewpoints and spectral data integration to offer a detailed multidimensional characterization of complex surfaces and internal structures at the voxel level. By aligning photogrammetric 3D point clouds with hyperspectral cubes, each voxel gains spectral signatures, permitting detailed material identification and chemical mapping in three dimensions. This combined approach opens new possibilities for studying heterogeneous samples, where both shape and chemical makeup are important. Reliable co-registration algorithms ensure pixel-level alignment between spatial and spectral data. The framework also allows correction of geometric distortions and normalization of intensity across views. Advanced chemometric techniques, including classification models and curve resolution methods, are used to extract key features or classify the different voxels within the 3D structure, enabling quick and accurate 3D localization of the compounds being studied.

The framework is particularly suitable for applications in cultural heritage, plant physiology, geosciences, and material forensics. We demonstrate the methodology with real case studies involving synthetic materials with complex geometries. The resulting data allow interactive visualization of spectral variability over depth and surface topography. The results emphasize the benefits of combining spatial and spectral data for better interpretability. This study advances toward real-time, in situ 3D chemical sensing with portable HSI-SWIR systems. 3D-HSI-SWIR paves the way for a new generation of spectral imaging tools with high potential in scientific and industrial fields.

Dr. Dimitrios Meimaroglou is currently a professor in the Chemical Engineering Department (ENSIC) of the University of Lorraine. He was awarded the Chair of Excellence in Polymer Reaction Engineering from 2011 to 2016. His main research interests, within the group of "Product Engineering" of the Laboratory of Reactions and Chemical Engineering (LRGP), are focused on modeling the microstructure of polymers and investigating its effects on their final end-use properties and functionalities. A part of his activities also reaches beyond the field of polymers, in a wider perspective of the study of the conditions-structure-property relationships of different products, within a product design framework. Both knowledge-based and data-driven techniques are implemented in this sense, with a special emphasis on the use of stochastic Monte Carlo algorithms and Machine Learning methods. He is the (co-)author of more than 40 peer-reviewed articles, 3 book chapters, and around 60 participations in national and international scientific events.

Addressing the challenges of data-driven approaches in chemical product engineering problems

Artificial Intelligence, especially Machine Learning, has become a vital tool for engineers worldwide over the last decade. In many applications, from the molecular scale to process design, traditional modeling is being replaced by powerful yet complex black boxes that perform the job better, faster, or both. Some engineers resist this change, mainly because they do not fully understand these tools. Others argue that ML methods are merely simple statistical tools, and if used carelessly, may lead to misunderstandings or serious mistakes. To bridge the gap between skepticism and effective use, we must examine how ML tools are applied in practice. A "plug and play" approach can mislead eager users or result in unreasonable conclusions. Accordingly, in this work, we analyze the choices and assumptions made at each stage of implementing an ML approach to an engineering problem, focusing on how these choices impact model performance. We show that decisions often seen as trivial in literature can greatly affect the model. The problem studied is predicting thermodynamic properties of molecules from their structure, a complex problem with direct implications for many fields, such as energy production, reactive systems, and process design software.

I am a Professor at ISAE-SUPAERO, Université de Toulouse, and associated member with AI institute ANITI. My research is in the field of signal and image processing, and representation learning. I am particularly interested in designing or learning new representations for signal and images, able to improve various tasks such as parameter estimation, image restoration, or image compression. I have a specific interest for spectral (or hyperspectral) images, whose processing and analysis requires to combine unsupervised machine learning in the spectral dimension, and image processing techniques for the spatial dimension. I also try to mix Bayesian estimation and machine learning, to get reliable, fast and powerful inversion or parameter estimation methods.

I try to apply these works to real problems in various scientific fields, including Earth observation, astronomy, physics, or biology.

Internationally recognized for her expertise in spectral sensors and emerging technologies applied to food integrity and authenticity, her research focuses on NIRS sensors, combining next-generation devices with advanced data analytics for real-time monitoring, traceability, and authentication of agri-food products. Her work encompasses feeds, meats, dairy, oils, fruits, and vegetables, applying NIRS alone or in combination with complementary sensors. A central strength of her work is the processing of complex datasets using multivariate analysis and nonlinear modeling, bridging fundamental research with practical applications.

Sylvain Le Corff is Professor of statistics and machine learning LPSM, Sorbonne Université, and a leading expert in generative AI. He has supervised numerous thesis and is author of many publications. He develops probabilistic models for complex and structured data mainly with regard to inverse problems and conditional for generative models.