Speaker
Description
Hyperspectral X-ray absorption spectra (XAS) imaging implemented at the ROCK-SOLEIL beamline [1] offers a unique opportunity to combine second time-resolution and micrometer spatial resolution for monitoring electronic and local order transformations occurring during a chemical reaction. Individual spectra collected under operando conditions at each raw pixel of hyperspectral cubes have low signal-to-noise ratio (SNR) due to the speed at which they are collected (5s) and the pixel size (1.625 µm). To improve the SNR of the spectra, cube merging and pixel binning (time and spatial averaging) are used as post-processing of dozens of collected hyperspectral cubes. In this work, an intelligent data collection using noise-aware deep learning was developed to predict the optimal number of cubes to collect to accurately extract structural parameters (e.g. coordination numbers) from EXAFS spectra targeting a spatial or time resolution.
Quantifying the quality of XAS fits remains a bottleneck for hyperspectral imaging, where millions of low-SNR spectra make pixelwise FEFF fitting computationally expensive. We address this by developing a data-driven method that predicts the EXAFS amplitude fitting uncertainty directly from noisy spectra—combining classical FEFF modelling with deep learning.
We synthetically generate large training datasets by perturbing clean μ(E) XAS spectra collected at ROCK for references under controlled noise models and acquisition conditions. Three noise models are compared: Gaussian-only, Poisson-only, and a mixed model combining counting statistics with residual Gaussian noise [2], [3]. Each synthetic dataset spans variations in edge shift (±10 eV), and k-weight (2, 3), producing >600 k labelled spectra.
A multimodal convolutional neural network (CNN) architecture [4], [5], pretrained on synthetic data and fine-tuned [6] with real beamline measurements, predicts the FEFF-derived fitting uncertainty σ_amp directly from spectra. The spectral branch processes χ(k) concatenated with |χ(R)|, while a compact embedding branch encodes the absorbing element, chemical family, edge type, and k-weight, allowing the network to generalize across diverse experimental conditions and absorption edges. At the pretraining stage, the Poisson-only model provides the most realistic calibration and transferability to real data, with the mixed model performing comparably. After fine-tuning, all models converge toward similar high accuracy (R² ≈ 0.9, MAE ≈ 2%), demonstrating that real-data adaptation effectively bridges the synthetic–experimental domain gap.
Throughput. On a CPU workstation we process around 4,000 spectra/s; on a single RTX-class GPU we reach 50,000 spectra/s, yielding ~37 s for a full 600×350 cube at 70% effective pixels (147k spectra) on CPU versus 3 s on GPU.
Sweeps over temporal averaging and spatial binning reveal clear quality–cost trade-offs, showing how acquisition parameters directly influence the predicted fitting uncertainty. The resulting σ_amp maps provide near-real-time feedback during experiments, allowing users to adjust time resolution or binning to reach a desired precision while optimizing acquisition time. Overall, this method delivers a data-driven approach for quantifying spectral reliability across full hyperspectral cubes, enabling adaptive and efficient acquisition strategies at modern synchrotron beamlines.
References
[1] V. Briois et al., « Hyperspectral full-field quick-EXAFS imaging at the ROCK beamline for monitoring micrometre-sized heterogeneity of functional materials under process conditions », J. Synchrotron Radiat., vol. 31, no 5, Art. no 5, sept. 2024, doi: 10.1107/S1600577524006581.
[2] N. Acito, M. Diani, et G. Corsini, « Signal-Dependent Noise Modeling and Model Parameter Estimation in Hyperspectral Images », IEEE Trans. Geosci. Remote Sens., vol. 49, no 8, p. 2957‑2971, août 2011, doi: 10.1109/TGRS.2011.2110657.
[3] A. Foi, M. Trimeche, V. Katkovnik, et K. Egiazarian, « Practical Poissonian-Gaussian Noise Modeling and Fitting for Single-Image Raw-Data », IEEE Trans. Image Process., vol. 17, no 10, p. 1737‑1754, oct. 2008, doi: 10.1109/TIP.2008.2001399.
[4] S. G. Elnaggar, I. E. Elsemman, et T. H. A. Soliman, « Embedding-Based Deep Neural Network and Convolutional Neural Network Graph Classifiers », Electronics, vol. 12, no 12, p. 2715, juin 2023, doi: 10.3390/electronics12122715.
[5] X.-K. Ma et al., « X-ray spectra correction based on deep learning CNN-LSTM model », Measurement, vol. 199, p. 111510, août 2022, doi: 10.1016/j.measurement.2022.111510.
[6] « Neural Network-Based On-Chip Spectroscopy Using a Scalable Plasmonic Encoder | ACS Nano ». Consulté le: 13 octobre 2025. [En ligne]. Disponible sur: https://pubs-acs-org.ressources-electroniques.univ-lille.fr/doi/full/10.1021/acsnano.1c00079
