AI and the World's Hard Problems · Life sciences · arXiv · 2026-10-23
Wavenumber-Aware Deep Learning for Robust Analysis of Raman Spectra
Spectral machine-learning models usually treat a spectrum as a fixed grid, which limits them when measurement conditions change. The authors encoded each point by its intensity and wavenumber, computing Transformer positional encodings from the wavenumber. On Raman spectra of three bacterial species and their mixtures, the approach stayed robust to noise, sparser sampling and partial coverage, and it held up on sampling grids not seen in training.