Séminaire

Vendredi 27 Février 2026 à 11h00.

Presentation of three recent projects at iLM using machine learning approaches

Matthias HILLENKAMP, Claire LOISON, Dylan BISSUEL, Gaël HUYNH
(iLM)

Salle de séminaires Lippmann

Invité(e) par
Abdulrahman Allouche

présentera en 1 heure :

''
Unsupervised machine learning for spectral data analysis[RC]
Matthias HILLENKAMP

The elemental analysis of (nano)materials by STEM-EDX (Energy-Dispersive X-ray spectroscopy in the Scanning Transmission Electron Microscope) generates large amounts of data with element-specific transition lines. These data typically suffer from a low signal-to-noise ratio and are highly redundant, i.e., the variations between individual spectra are small.
In our research we use comparably simple unsupervised machine learning tools (Principal Component Analysis, Non-negative Matrix Factorization) in order to exploit these redundancies for denoising, evidencing spatial/spectral correlations and advanced statistical analysis. These same techniques are applicable for related data sets from experiments such as STEM-EELS, LIBS, XPS, plasmon spectroscopy and others.

Machine Learning approaches to predict non-linear optical properties of molecules embedded in a solvent

Claire LOISON & Dylan BISSUEL

Non-linear optical techniques are increasingly used to probe the structural properties of matter, particularly the orientation of molecules.
In collaboration with colleagues from iLM, we aim to model the intensities gathered during second harmonic scattering (SHS) experiments. In this process, a laser is directed through a sample, and two exciting photons interact with the sample to generate a new photon with twice the energy of the incoming photons.
The intensity of the second harmonic light generated in different directions is measured as a function of the incoming and outgoing polarizations.
In the Theochem group, we perform quantum mechanics/molecular mechanics (QM/MM) calculations to predict the second harmonic response of water molecules in liquid water. However, the QM/MM approach requires extensive sampling of the conformations of the liquid surrounding a water molecule, which leads to high computational costs. To overcome this issue, we are working on replacing the costly first-principles quantum chemistry calculations with machine learning predictors based on various models, particularly Message Passing Graph Neural Networks (MP-GNNs).
In this presentation, we will present the specificity of the molecular observables that we want to predict, the MP-GNN architecture, and the hyperparameter optimization that led to precise predictions of the molecular dipole moment, polarizability, and hyperpolarizability of water molecules in a liquid water environment.
Although we have focused on liquid water, our goal is to generalize to other liquids in bulk or at interfaces to quantitatively interpret SHS intensities in terms of short-range molecular ordering.

Identifying local order in complex materials: from geometric Methods to graph neural networks

Gaël HUYNH

Identifying crystal phases and structural defects in molecular dynamics simulations is essential for understanding material properties. Historically, this task has relied on "rigid" geometric methods or physical descriptors coupled with statistical approaches. While efficient, these methods struggle to capture local order in complex systems or under significant thermal fluctuations.
This talk traces the evolution of these techniques toward machine learning to overcome such limitations. After reviewing the constraints of early implementations, we explore the advantages of graph neural networks. We specifically demonstrate how to natively encode rotational equivariance, ensuring that the model respects the physical symmetries of the crystal. Finally, we discuss the trade-off between accuracy and numerical cost, addressing the challenges of computation time and memory consumption for large-scale structural analysis.

''