Bayesian Optimization Outperforms Experts for Spectral Model Development
Abstract
Indirect Hard Modeling (IHM) is a physically motivated approach for quantitative spectral analysis that can account for nonlinear spectral changes, provides loweffort calibration, and allows extrapolation. However, realizing the full potential of IHM requires careful adjustment of model parameters and spectral preprocessing settings, which is typically performed manually and relies on expert knowledge. To address this limitation, we investigate Bayesian Optimization (BO) as a strategy for efficient IHM parameter tuning with mixed continuous, discrete, and categorical parameter spaces. We couple six alternative BO frameworks differing in surrogate model and sampling strategy with Peaxact, a state-of-the-art commercial software for IHM. We compare the calibration performance and computation time. We develop four spectroscopic models based on literature datasets. The corresponding IHM models in the form of published parameter settings serve only as benchmarks and are not used to initialize the optimization. We thus demonstrate that IHM models can be automatically optimized without expert intervention nor prior selected configurations. The results show that BO outperforms the published models, ranging from marginal changes to reductions of approximately 50% in cumulative cross-validation RMSE and approximately 60% in cumulative prediction RMSE, 1 with mean runtimes of approximately five hours or less on a standard desktop computer, supporting practical applicability. Moreover, the configurations obtained via BO partially agree with the expert-selected settings, indicating that BO can identify reasonable calibration settings from scratch. Overall, the findings demonstrate that BO can improve IHM model accuracy, reduce manual calibration effort, and support the broader use of IHM in spectroscopy.
Authors 5
-
Affiliation as printed
Process Systems Engineering (AVT.SVT), RWTH Aachen University, Forckenbeckstr. 51, Aachen, Germany, 52074
-
Affiliation as printed
Process Systems Engineering (AVT.SVT), RWTH Aachen University, Forckenbeckstr. 51, Aachen, Germany, 52074
-
RWTH Aachen University · University College London
Affiliation as printed
Process Systems Engineering (AVT.SVT), RWTH Aachen University, Forckenbeckstr. 51, Aachen, Germany, 52074
Sargent Centre for Process Systems Engineering, Department of Chemical Engineering
University College London
-
RWTH Aachen University · Jülich Aachen Research Alliance
Affiliation as printed
Institute of Climate and Energy Systems ICE-1: Energy Systems Engineering
JARA Center for Simulation and Data Science (CSD
Process Systems Engineering (AVT.SVT), RWTH Aachen University, Forckenbeckstr. 51, Aachen, Germany, 52074
-
Affiliation as printed
Process Systems Engineering (AVT.SVT), RWTH Aachen University, Forckenbeckstr. 51, Aachen, Germany, 52074