Thomas Dezaire
Research Subjects
Thomas Dezaire has a total of 7 co-authors. Here are the top co-authors by number of shared publications:
Andre van Zomeren
1 shared publicationEleftheria Mavroforaki
1 shared publicationFlorian Meirer
1 shared publicationKatalin Barta
1 shared publication- Marc Bruggeman
TNO, Energy and Materials Transition, Biobased and Circular Technologies Group, 1755 ZG Petten, Netherlands
1 shared publication Pieter Bruijnincx
1 shared publication
Latest Publications
- Open Access
- 25 September 2026
- Published by Elsevier in SSRN Electronic Journal
Quantifying spectral preprocessing effects across regression models for ATR-FTIR-based prediction of lignin structure
Characterization of major lignin structures and the bond types between monolignol units is generally performed by HSQC-NMR, a cost-intensive technique that requires specialized personnel to operate and maintain the instrumentation. ATR-FTIR offers a rapid and experimentally accessible alternative, but its interpretation is complex and correlating spectra with structural data is not straightforward, creating the need for predictive models trained on HSQC-NMR data. Existing FTIR-based models have been calibrated largely on softwood technical lignins, leaving hardwood and less condensed lignins underexplored. Their performance further depends critically on the spectral preprocessing applied before regression, a step usually chosen by convention rather than by systematic comparison. Here, the effects of preprocessing and regression model family on predictive performance were systematically evaluated across 61 isolated lignins from ten botanical origins. Combining background correction, normalization, and derivative transformation in a full combinatorial grid yielded 624 preprocessing configurations. Each configuration was evaluated under repeated stratified low-overlap train/test splits with three regression families spanning classical chemometrics and machine learning, namely partial least squares (PLS), elastic net (ENet), and random forest (RF). Normalization produced the main separation between low- and high-performing configurations, while model family, derivative transformation, and background correction had smaller or more conditional effects. High-performing configurations were concentrated around min-max normalization, first-derivative representations, and ENet regression. For the best internally selected workflow, mean held-out R2 values across the diversified splits were 0.91 for syringyl units, and 0.88 for guaiacyl units, 0.77 for β-O-4 linkages. The benchmark supports ATR-FTIR as a screening-level tool for major NMR-derived lignin descriptors and shows that preprocessing choices shape workflow performance.
- 25 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessQuantifying spectral preprocessing effects across regression models for ATR-FTIR-based prediction of lignin structure
Characterization of major lignin structures and the bond types between monolignol units is generally performed by HSQC-NMR, a cost-intensive technique that requires specialized personnel to operate and maintain the instrumentation. ATR-FTIR offers a rapid and experimentally accessible alternative, but its interpretation is complex and correlating spectra with structural data is not straightforward, creating the need for predictive models trained on HSQC-NMR data. Existing FTIR-based models have been calibrated largely on softwood technical lignins, leaving hardwood and less condensed lignins underexplored. Their performance further depends critically on the spectral preprocessing applied before regression, a step usually chosen by convention rather than by systematic comparison. Here, the effects of preprocessing and regression model family on predictive performance were systematically evaluated across 61 isolated lignins from ten botanical origins. Combining background correction, normalization, and derivative transformation in a full combinatorial grid yielded 624 preprocessing configurations. Each configuration was evaluated under repeated stratified low-overlap train/test splits with three regression families spanning classical chemometrics and machine learning, namely partial least squares (PLS), elastic net (ENet), and random forest (RF). Normalization produced the main separation between low- and high-performing configurations, while model family, derivative transformation, and background correction had smaller or more conditional effects. High-performing configurations were concentrated around min-max normalization, first-derivative representations, and ENet regression. For the best internally selected workflow, mean held-out R2 values across the diversified splits were 0.91 for syringyl units, and 0.88 for guaiacyl units, 0.77 for β-O-4 linkages. The benchmark supports ATR-FTIR as a screening-level tool for major NMR-derived lignin descriptors and shows that preprocessing choices shape workflow performance.