Evaluation of machine learning representations for trace asbestos detection in recycled gypsum using wide-angle X-ray scattering
Abstract Asbestos contamination blocks gypsum recycling from demolition waste. Regulatory thresholds near 0.1 wt% lie well below the $$\sim $$ 1 wt% limit of conventional X-ray powder diffraction. We evaluate machine-learning detection of trace chrysotile (0.001–9 wt%) in recycled gypsum from wide-angle X-ray scattering, comparing six feature representations under one Extra-Trees estimator on 1532 spectra from 373 samples, with group-aware splits and 18 seeds. Internally the representations are indistinguishable, so we report the minimum detectable difference (0.030–0.062) rather than a ranking. The untransformed spectrum is the strongest internal detector (AUC $$0.9993 \pm 0.0007$$ , zero false negatives at 9.4% false positives), yet misses all 16 measurements of a commercial fibre-cement board absent from training, which an unsupervised linear projection detects in full. Dimensionality reduction, not internal accuracy, determines whether a decision rule transfers to a new matrix. We therefore select on cross-matrix transfer, reporting a PCA model restricted to the diagnostic chrysotile reflections (accuracy $$0.9450 \pm 0.0187$$ ). A retained production sample independently assayed at 0.0201 wt%, one fifth of the regulatory threshold, and used in no selection step is identified as asbestos-bearing. Quantification remains semi-quantitative (log-MAE 0.62) and does not transfer across matrices, so the framework suits on-line screening in industrial recycling.
Authors
- Christian Scherdel (ORCID: https://orcid.org/0009-0002-8454-332X)
- Jan Schmitt (ORCID: https://orcid.org/0000-0003-4537-7680)
- Henning Förster
- Bastian Engelmann (ORCID: https://orcid.org/0000-0002-2053-0995)
- Frank Lotter (ORCID: https://orcid.org/0000-0002-7872-0484)
- Julia Welscher
- Markus Wilhelm (ORCID: https://orcid.org/0000-0002-9620-466X)
- Alexandra Zeh
Institutions
- Technical University of Applied Sciences Würzburg-Schweinfurt (DE)
- Breakthrough (US)
- Smart Material (Germany) (DE)
- Bavarian Center for Applied Energy Research (DE)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-14
- DOI
- https://doi.org/10.1038/s41598-026-79747-w
- Primary Topic
- Spectroscopy and Chemometric Analyses
- Type
- article
- Field-Weighted Citation Impact
- 0.00