Nano-FTIR stops discarding its noisy-but-clean harmonics: a cross-harmonic network pulls third-harmonic phase noise to about 0.02 rad
Synopsis
Hyperspectral nano-FTIR imaging has relied on the third harmonic, which is nearly background-free but noisy, while the higher-SNR first and second harmonics were discarded because of background contamination. A cross-harmonic neural network (CHNN), trained on experimental data, reconstructs third-harmonic spectra from first- and second-harmonic spectra, reducing noise on a PMMA/PS blend while preserving spectral features and nanoscale spatial contrast, and it suppresses noise better than a single-harmonic autoencoder.
Interpretation
Hyperspectral nano-FTIR imaging can now use the first and second harmonics, previously discarded for background contamination, as inputs to reconstruct a background-free third-harmonic data cube, raising signal-to-noise ratio without increasing acquisition time. Quantitative hyperspectral measurements have generally relied on higher harmonics such as the third and fourth, which are nearly background-free but have low signal-to-noise ratio; the first and second harmonics have much higher SNR but carry substantial background that distorts resonances and shifts peak positions, so they were usually discarded. On an 80×49-pixel, 64-wavenumber hyperspectral cube of a PMMA/PS polymer blend, the averaged standard deviation of third-harmonic phase in the PMMA region rises from 0.13 rad in the central spectral region to 0.47 rad at the spectral edges, while the CHNN reconstruction gives about 0.02 rad in both regimes.
Reconstruction proceeds pixel by pixel, with the network mapping the complete first- and second-harmonic complex spectra at a position onto the corresponding third-harmonic complex spectrum. The cross-harmonic neural network (CHNN) is a fully connected multilayer perceptron whose input is a 256-dimensional vector concatenating the real and imaginary parts of the first two harmonics and whose output is a 128-dimensional vector of the third harmonic's real and imaginary parts. Training samples come directly from the measured hyperspectral cube, one per spatial position, 3920 in total, split 80% training and 20% validation; the single-harmonic autoencoder benchmark uses the same split, normalization, loss and training settings.
Perspective
The demonstrated scope is an 80×49-pixel, 64-wavenumber hyperspectral data cube with 100 nm pixel size on a PMMA/PS polymer blend, acquired over 4 hours; spatial fidelity is referenced to a single-frequency QCL image of the same region at 50 nm pixel size.
Only the mapping from the first and second harmonics to the third was validated; a 1→3 variant using only the first harmonic also reduced noise but did not recover sub-300 nm PMMA particles or voids inside large domains, which the authors attribute to the larger probing volume of the first harmonic.
