Predicting rare earth elements concentration in coal ashes with multi-task neural networks

Yu Song; Yifan Zhao; Alex Ginella; Benjamin Gallagher; Gaurav Sant; Mathieu Bauchy

doi:10.1039/d3mh01491f

Predicting rare earth elements concentration in coal ashes with multi-task neural networks

Mater Horiz. 2024 Mar 18;11(6):1448-1464. doi: 10.1039/d3mh01491f.

Authors

Yu Song^{1

2}, Yifan Zhao¹, Alex Ginella¹, Benjamin Gallagher³, Gaurav Sant^{2

4

5

6}, Mathieu Bauchy^{1

4}

Affiliations

¹ Physics of AmoRphous and Inorganic Solids Laboratory (PARISlab) 5731B Boelter Hall, Department of Civil and Environmental Engineering, University of California, Los Angeles, CA 90095, USA. yusong@ucla.edu.
² Laboratory for the Chemistry of Construction Materials (LC2) 5731J Boelter Hall, Department of Civil and Environmental Engineering, University of California, Los Angeles, CA 90095, USA.
³ Electric Power Research Institute (EPRI) 3420 Hillview Avenue, Palo Alto, CA 94304, USA.
⁴ Institute for Carbon Management (ICM), University of California, Los Angeles, CA, USA.
⁵ Department of Materials Science and Engineering, University of California, Los Angeles, CA, USA.
⁶ California Nanosystems Institute, University of California, Los Angeles, CA, USA.

PMID: 38214154
DOI: 10.1039/d3mh01491f

Abstract

The increasing demand for rare earth elements (REEs) makes them a scarce strategic resource for technical developments. In that regard, harvesting REEs from coal ashes-a waste byproduct from coal power plants-offers an alternative solution to conventional ore-based extraction. However, this approach is bottlenecked by our ability to screen coal ashes bearing large concentrations of REEs from feedstocks-since measuring the REE content in ashes is a time-consuming and costly task requiring advanced analytical tools. Here, we propose a machine learning approach to predict the REE contents based on the bulk composition of coal ashes, easily measurable under the routine testing protocol. We introduce a multi-task neural network that simultaneously predicts the contents of different REEs. Compared to the single-task model, this model exhibits notably improved accuracy and reduced sensitivity to noise. Further model analyses reveal key data patterns for screening coal ashes with high REE concentrations. Additionally, we showcase the utilization of transfer learning to improve the adaptability of our model to coal ashes from a distinct source.