Uncertainty Aware AFM-IR Hyperspectral Cube Reconstruction via Convolutional-Attentive Conditional Neural Processes
Year: 2026
Type: preprint
Abstract: Atomic force microscopy-infrared spectroscopy (AFM-IR) maps chemical composition at nanoscale spatial resolution, but the time required to acquire individual spectra limits hyperspectral imaging. Here, we introduce a convolutional-attentive conditional neural process (CACNP), a neural process architecture for reconstructing AFM-IR hyperspectral cubes from sparsely sampled spectra and a small set of structure images, each acquired at a selected wavenumber. The model is trained entirely on synthetic cubes generated from experimental infrared spectra and predicts both spectral intensities and their associated uncertainty. Validation on experimental datasets of poly(methyl methacrylate) and polystyrene particles embedded in epoxy demonstrates accurate reconstruction using only 5% of the available spectra, with competitive performance when compared to other methods. Inference takes less than 0.1 s on 128 X 128 hyperspectral cubes, supporting rapid reconstruction updates, while predictive uncertainty provides a basis for adaptive sampling of localized chemical features. This approach advances autonomous, high-throughput AFM-IR workflows and offers an architecture for accelerating other hyperspectral imaging methods.
Source: arXiv (Cornell University)
Language: English
Subfield: Analytical Chemistry
Field: Chemistry
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225997157
