Skip to content

Processing & reconstruction

Deep-learning spectral enhancement

Neural reconstruction that decouples, sharpens and denoises spectra that classical processing cannot rescue.

Some spectral problems are not processing problems. A carbon-proton methyl correlation in a non-deuterated sample is broadened by couplings that no window function will remove, and scalar coupling in the indirect dimension halves your resolution before you start.

FID-Net addresses these directly: virtual decoupling, aromatic reconstruction, and sharpening of methyl correlations in non-deuterated protein. It is a trained transformation of the time-domain data, so we treat its output as a derived spectrum with its own provenance rather than as a corrected version of the original — you keep both, and can always show what the network changed.

What it needs

  • Processed or time-domain spectrum
  • The network appropriate to the experiment

What you get

  • Enhanced spectrum, stored separately from its source

What does the work

  • FID-Net

    Open source, redistributed under its licenceMIT

    Trained reconstruction networks

Related in Processing & reconstruction