READ ME File For 'Ultrafast Laser Filamentation Classification and Analysis via Neural Networks'

Dataset DOI: https://doi.org/10.5258/SOTON/D3714

ReadMe Author: James A Grant-Jacob  University of Southampton https://orcid.org/0000-0002-4270-4247

This dataset supports the publication:
AUTHORS: James A. Grant-Jacob and  Ben Mills
TITLE: Ultrafast Laser Filamentation Classification and Analysis via Neural Networksg
JOURNAL: JPhys: Photonics
PAPER DOI IF KNOWN:  


This dataset contains:

Figure_Images.zip

(Figure_1.jpg 
Figure_2.jpg
Figure_3.jpg 
Figure_4.jpg
Figure_5.jpg 
Figure_6.jpg
Figure_7.jpg 
Figure_2.jpg
Figure_8.jpg 
Figure_9.jpg
Figure_10.jpg)

Figure_Data.zip

Figure_1_spectra_100_descending_5pct.txt
Figure_2_BroadbandBand_500_700_nm_all_lenses.txt
Figure_4a_spectra_compare_LENS_200.txt
Figure_4b_spectra_compare_LENS_250.txt
Figure_6a_200mm_nmf_component_intensity_vs_power.txt
Figure_6b_250mm_nmf_component_intensity_vs_power.txt
Figure_8b_PerPower_BandMetrics_lens_250.txt Data for Figure 8a.txt
Figure_8b_PerPower_BandMetrics_lens_250.txt Data for Figure 8b.txt

The figures are as follows:


Figure 1. (a) Schematic of the experimental setup showing the femtosecond laser source (1030 nm, 190 fs), focusing lens (200 mm focal length in this instance, but changeable), side-imaging camera, and fibre-coupled spectrometer. The camera and spectrometer were positioned orthogonally to the filament axis. (b) Cropped side-view images of plasma generated using different powers. (c) Corresponding emission spectra for different laser power percentages of the maximum using a 200 mm focal length lens, showing broadband continuum features and relative intensity variation. 

Figure 2. Broadband integrated spectral intensity (500–700 nm) as a function of input power for different focusing lenses. The onset of filamentation for each lens is determined by the inflection point in the energy growth curve, identified from a zero-crossing in the second derivative (d2I/dP2). Vertical dashed lines indicate the filamentation onset power, which defines the ground-truth labels for training the CNN.

Figure 3. Filament classification results for data from using a lens of focal length (a) 200 mm and (b) 250 mm. Ground truth labels (blue, ‘×’), CNN predictions (teal, ‘○’), and linear regression predictions (lime green, ‘□’) are plotted against laser power (%). The predicted onset of filamentation from the CNN and baseline regression models are indicated by dotted teal and dotted green vertical lines, respectively, with shaded regions highlighting the corresponding post-onset regimes.

Figure 4. Representative results comparing the NMF–CNN regression model and a linear regression baseline trained to predict plasma emission spectra from image-derived features. Four test samples are shown for the (a) 200 mm focal length lens and (b) 250 mm focal length lens at input powers of 25%, 50%, 75%, and 100%. The top row shows the resized plasma images used for training (resized to 128 × 128 pixels, which stretches the vertical axis and makes the filament appear wider than in raw frames). The bottom row compares the ground-truth spectra (blue) with predictions from the linear regression model (green) and the CNN-based model (yellow). 

Figure 5. NMF component spectra (first column) and Grad-CAM activations map (columns 2–6, 50% transparency) overlaid on test images for Components 1–5 at 25% laser power (top row) and 100% power (bottom rom) for (a) 200 mm focal length lens and (b) 250 mm focal length lens. 

Figure 6: Integrated intensity of NMF components as a function of power for (a) 200 mm and (b) 250 mm focal length lenses. Component 1 rises sharply near filament onset and saturates at high power, consistent with intensity clamping. 

Figure 7. Grad-CAM intensity and spatial migration trends for NMF Components 2 and 3 across input power. (Top) Sum of the top 20 Grad-CAM pixels increases with power, indicating stronger spatial attribution for continuum-related components. (Bottom) Normalized radial distance of the brightest pixel decreases with power, showing activation shifts toward the filament core at high power. These trends provide quantitative evidence of spatial–spectral coupling beyond simple power scaling.

Figure 8. Median RMSE versus power for five wavelength bands for (a) 200 mm focal length lens and (b) 250 mm focal length lens. The colour bar indicates RMSE magnitude (blue is low error, yellow is high error). 

Figure 9. Power-dependent performance and SNR analysis for (a) 200 mm focal length lens and (b) 250 mm focal length lens. (Left) NMF–CNN: Median R2 (blue, left axis) and RMSE (green, right axis) versus power. (Middle) Linear regression: Median R2 and RMSE versus power. (Right) Image and spectral SNR versus power. 

Figure 10. 3D normalised waterfall comparison between measured spectra (blue) and CNN‑reconstructed spectra (green) at 100% power for lenses 75–300 mm. All data were acquired on different days and under slightly different laser alignment conditions than those used to train the model. For clarity, each true spectrum is normalised to its peak value, and the corresponding CNN prediction is scaled using the same per‑lens normalisation factor. This cross‑day evaluation demonstrates that the NMF–CNN retains predictive 


The graph data are contained in the tab separate data files as follows, with their relevant headings:

Figure_1_spectra_100_descending_5pct
Figure_2_BroadbandBand_500_700_nm_all_lenses
Figure_4a_spectra_compare_LENS_200
Figure_4b_spectra_compare_LENS_250
Figure_6a_200mm_nmf_component_intensity_vs_power
Figure_6b_250mm_nmf_component_intensity_vs_power
Figure_8b_PerPower_BandMetrics_lens_250.txt Data for Figure 8a
Figure_8b_PerPower_BandMetrics_lens_250.txt Data for Figure 8b

Raw_data.zip contains the raw files acquired during experiment.

License: CC-BY

Related projects:
EPSRC grant EP/T026197/1
EPSRC grant EP/W028786/1
EPSRC grant EP/Z002567/1
Date that the file was created: 10, 2026