Uncertainty-aware Bayesian neural networks for FTIR gas-phase process monitoring: a semi-synthetic feasibility study
Year: 2026
Type: article
Abstract: To remain safe and actionable, multi-component gas-phase Fourier-transform infrared (FTIR) process analysis requires a reliability estimate for each concentration prediction. In a semi-synthetic feasibility study for a fixed compound library, we present a Bayesian one-dimensional Convolutional Neural Network (1D-CNN) trained on spectra from single-compound reference data that predicts concentrations, compound presence probabilities and per-prediction uncertainty bounds. The model yields accurate concentration predictions (mean absolute error 0.165 in log 10 -space, within a factor of about 1.5) and uncertainty bounds that remain well-calibrated under severe instrumental artifacts. A pre-inference amplitude guard screens for inputs outside the detector’s valid operating range. Validation on measured spectra is the necessary next step before deployment.
Source: at - Automatisierungstechnik
Authors: André Orth, Hawo H. Höfer, Markus Reischl
Institution: Karlsruhe Institute of Technology
Language: English
Subfield: Analytical Chemistry
Field: Chemistry
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,712
OpenAlex ID: W7225616775