Machine Learning-Driven Design of Experiments for CO2 Mineralization of Steel Slag Materials
- 1CanmetENERGY , , 1 Haanel Drive , , ,
Abstract
The mineralization of CO2 using steel slag represents a promising pathway for simultaneous carbon sequestration and industrial-waste valorization. However, the complex and nonlinear interactions between process variables hinder the identification of the dominant mechanisms and optimal operating conditions. In this study, a hybrid modeling framework integrating artificial neural networks (ANNs) with factorial design analysis was developed to predict and interpret the performance of three key process routes: leaching, direct carbonation, and indirect carbonation. A comprehensive data set compiled from 58 published studies, comprising 171 leaching, 168 direct carbonation, and 22 indirect carbonation experiments, was used to train and validate ANN models. The trained models were subsequently employed to generate a full factorial design (3k), enabling systematic evaluation of parameter significance and interaction effects. The ANN models demonstrated strong predictive capability for leaching and acceptable performance for carbonation processes, with improved accuracy upon inclusion of reaction time, highlighting the importance of kinetic effects. The ANN–factorial analysis suggests that leaching is most strongly associated with feed composition (CaO content) and leaching agent chemistry, while direct carbonation is controlled by reaction kinetics and surface accessibility, with strong interactions between CO2 availability, pretreatment, and time. Indirect-carbonation results, which are based on the smaller data set, suggest dominant associations with solution-phase chemistry, particularly pH, Ca concentration, and precipitation kinetics, with interaction effects playing a critical role in determining carbonation efficiency. Across all processes, chemical and kinetic interactions were found to be more influential than purely physical parameters. The integration of ANN with factorial design enables the creation of a virtual experimental space, allowing comprehensive exploration of complex multivariable systems beyond experimental limitations. Steel slag mineralization exhibits a progressive shift in dominant control mechanisms, from composition-dominated behavior during leaching to coupled composition–gas-supply control during direct carbonation, and finally to solution-phase control during indirect carbonation. Comparison with our previously studied Mg-rich serpentine system further shows that the ANN–factorial framework is transferable across alkaline feedstocks, while the controlling mechanisms remain feedstock-specific: Mg-rich systems are more strongly governed by pretreatment and pH/additive chemistry, whereas steel slag systems are dominated by reactive Ca availability, CO2 supply, and Ca–carbonate precipitation chemistry. These findings provide a screening framework for process optimization and highlight the importance of material-specific modeling in CO2 mineralization technologies.
Related
- Emerging Paradigms in Electrochemical Water Treatment: Advanced Electrode Materials, Resource Recovery, and Machine Learning
- Machine learning-driven prediction of heavy metal pollution for aquatic health surveillance and disease control
- Data-driven machine learning modeling reveals the impact of micro/nanoplastics on microalgae and their key underlying mechanisms
- Using machine learning effectively in marine benthic ecology: a question-driven toolkit
- Ensemble tree-based machine learning models for data-driven AQI prediction
BibTeX
@article{radfarnia2026machine,
title = {{Machine Learning-Driven Design of Experiments for CO2 Mineralization of Steel Slag Materials}},
author = {Hamid Radfarnia and Iris Samputu and Kourosh Zanganeh},
journal = {ACS Omega},
year = {2026},
month = oct,
doi = {10.1021/acsomega.6c07267},
url = {https://pubs.acs.org/acsodf/article/doi/10.1021/acsomega.6c07267/5453987/Machine-Learning-Driven-Design-of-Experiments-for},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Machine Learning-Driven Design of Experiments for CO2 Mineralization of Steel Slag Materials
AU - Hamid Radfarnia
AU - Iris Samputu
AU - Kourosh Zanganeh
PY - 2026
DA - 2026/10/19
JO - ACS Omega
DO - 10.1021/acsomega.6c07267
UR - https://pubs.acs.org/acsodf/article/doi/10.1021/acsomega.6c07267/5453987/Machine-Learning-Driven-Design-of-Experiments-for
AB - The mineralization of CO2 using steel slag represents a promising pathway for simultaneous carbon sequestration and industrial-waste valorization. However, the complex and nonlinear interactions between process variables hinder the identification of the dominant mechanisms and optimal operating conditions. In this study, a hybrid modeling framework integrating artificial neural networks (ANNs) with factorial design analysis was developed to predict and interpret the performance of three key process routes: leaching, direct carbonation, and indirect carbonation. A comprehensive data set compiled from 58 published studies, comprising 171 leaching, 168 direct carbonation, and 22 indirect carbonation experiments, was used to train and validate ANN models. The trained models were subsequently employed to generate a full factorial design (3k), enabling systematic evaluation of parameter significance and interaction effects. The ANN models demonstrated strong predictive capability for leaching and acceptable performance for carbonation processes, with improved accuracy upon inclusion of reaction time, highlighting the importance of kinetic effects. The ANN–factorial analysis suggests that leaching is most strongly associated with feed composition (CaO content) and leaching agent chemistry, while direct carbonation is controlled by reaction kinetics and surface accessibility, with strong interactions between CO2 availability, pretreatment, and time. Indirect-carbonation results, which are based on the smaller data set, suggest dominant associations with solution-phase chemistry, particularly pH, Ca concentration, and precipitation kinetics, with interaction effects playing a critical role in determining carbonation efficiency. Across all processes, chemical and kinetic interactions were found to be more influential than purely physical parameters. The integration of ANN with factorial design enables the creation of a virtual experimental space, allowing comprehensive exploration of complex multivariable systems beyond experimental limitations. Steel slag mineralization exhibits a progressive shift in dominant control mechanisms, from composition-dominated behavior during leaching to coupled composition–gas-supply control during direct carbonation, and finally to solution-phase control during indirect carbonation. Comparison with our previously studied Mg-rich serpentine system further shows that the ANN–factorial framework is transferable across alkaline feedstocks, while the controlling mechanisms remain feedstock-specific: Mg-rich systems are more strongly governed by pretreatment and pH/additive chemistry, whereas steel slag systems are dominated by reactive Ca availability, CO2 supply, and Ca–carbonate precipitation chemistry. These findings provide a screening framework for process optimization and highlight the importance of material-specific modeling in CO2 mineralization technologies.
SN - 2470-1343
ER -