AFM-IR hyperspectral imaging with just 5% of spectra: CACNP rebuilds the full chemical cube in under 0.1 s with per-voxel uncertainty
Synopsis
Nano-IR hyperspectral imaging collects a spectrum point by point at every pixel, so a complete cube can take hours. CACNP feeds sparsely sampled spectra together with a few single-wavenumber structure images into a convolutional-attentive conditional neural process trained on synthetic cubes, returning the full cube and a per-voxel uncertainty. On experimental data of polystyrene and poly(methyl methacrylate) particles in epoxy, 5% of the spectra suffices, inference takes under 0.1 s, and reconstruction error approaches the sample noise level.
Interpretation
Nano-IR hyperspectral imaging can now rebuild a full chemical cube from a small fraction of its spectra while reporting how much to trust each voxel. Previously a complete cube required collecting a spectrum at every spatial position, so acquisition time scaled with both the number of positions and the number of wavenumbers; in practice users often imaged at a few preselected wavenumbers, risking missed spectral features outside those bands. On two experimental cubes of polystyrene and poly(methyl methacrylate) particles in epoxy, three structure images and 5% context gave an RMSE of 0.0375 and a misclassification rate of 1.57% for the multi-particle dataset, versus 0.0369 and 1.37% at 10% context; these errors are close to the estimated noise standard deviations of the two samples.
Reconstruction is a single forward pass rather than a refit for each set of observations. Gaussian-process methods slow as observations grow, and matrix factorization and SS-Net reconstruct accurately but return no calibrated uncertainty, which limits adaptive sampling during an experiment. For the single-particle dataset at 1% context, cumulative acquisition plus reconstruction time was about 6 minutes for CACNP, comparable to CNMF, versus about 46 minutes for GP and 53 minutes for SS-Net; at 10% context CACNP took about 31 minutes, GP about 7 hours 45 minutes and SS-Net about 11 hours 25 minutes.
The model places sparse spectra onto the spatial grid, then lets reconstruction features query the measured spectra directly through cross-attention. Convolutional conditional neural processes embed observations on a grid and rely mainly on convolution for spatial relations, with no direct route from a query to the measured spectra. An otherwise identical architecture with all cross-attention removed produced blurrier boundaries and poorer registration on the multi-particle dataset, with higher reconstruction uncertainty and RMSE.
Perspective
The demonstrated scope is two AFM-IR cubes acquired on a Bruker standard sample of polystyrene and poly(methyl methacrylate) particles in epoxy, spanning 761 to 1865 cm-1 in 4 cm-1 steps; all training data are synthetic cubes, and the experimental cubes were excluded from training.
Uncertainty-guided acquisition does not always improve reconstruction: on the multi-particle dataset, selecting by uncertainty propagated one outlier intensity into surrounding regions, most visibly for the distance-plus-uncertainty strategy, whose calibrated coverage fluctuated below the nominal 95.4% over the evaluated range. Raw uncertainty is underestimated at low context fractions and requires rescaling calibrated on held-out context, a procedure that carries no conformal coverage guarantee.