Oblisk reads a raw Thomson parabola spectrometer image, detects ion traces, identifies candidate species by their mass-to-charge ratios and builds energy spectra. Automatic cropping, denoising and shared geometry fitting replace the first round of hand-tracing.
Each parabola is matched to a charge state through its fitted m/q, using a physics ladder built from your elements.
Signal is integrated along each fitted track and mapped from position to energy.
Origin, rotation, tilt and optional distortion are fitted once, for all traces together.
No pre-cropping. A neural detector finds the active plate and the zero-deflection spot, then fixes orientation.
U-Net denoising, row-wise peak tracking, then one geometry fit shared by every parabola.
Hover traces for species and curvature, open spectra, download plots and data.
parabolas found and correctly identified per image, on average.
mean distance between fitted parabolas and extracted traces on clean images.
on noisy images, still narrower than the traces themselves.
Evaluation from the bachelor thesis “Analysis of experimental data from the Thomson spectrometer”, CTU 2026.
Noise as bright as the signal, a saturated zero point with ghost spots beside it, a rotated and tilted frame, traces whose brightness wanders. That is why one image can take hours by hand.
Oblisk gets more robust with every new detector and geometry. If you run a TPS, I'd love to test it on your shots and help analyse them.
