Machine learning prediction of bismuth recovery from literature data: publication-grouped validation, anchored calibration and chemical interpretation
Year: 2026
Type: preprint
Abstract: Machine learning models trained on published leaching data often report high accuracy from random splits that place runs of one study in both training and test sets. This study tested whether reported operating conditions can predict bismuth recovery for a study that the model has not seen. A data set of 485 runs from 82 sources was compiled across six unit operations; runs designed to retain bismuth, plate part of a bath or recover another metal were set aside, leaving 375 recovery runs from 72 studies. Thirteen models, including a mean baseline, were compared under random, publication-grouped and feedstock-grouped validation, and six of them under temporal validation. Random splits gave R2 up to 0.57, whereas publication-grouped validation gave at most 0.09 (95% interval −0.22 to 0.28); about half of the residual variance lay between publications. Three measured runs from a new study raised R2 for its remaining runs from about zero to 0.48, and Gaussian-process predictions were positively rank-correlated with measured recovery in 83% of studies. Within studies, recovery rose with acid concentration and temperature, although the temperature trend tended to reverse across studies. Literature data can therefore rank conditions for a new feed, but absolute prediction needs a few local measurements; an offline browser tool that applies this approach to leaching is provided.
Source: ChemRxiv
Author: Subash Chandra Bose Rapaka
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7221613611