En
期刊文章
文章精选
当期目录
过刊浏览
高被引
高下载
特邀文章
专刊合集
待刊文章
Volume 117
您当前的位置：首页 > 期刊文章 > 过刊浏览 > Volumes 108-119 (2025) > Volume 117
Deciphering process-morphology relationships in magnesite-derived nesquehonite crystallization using interpretable machine learning
Fengling Weng a, Yulian Wang a b *, Jiayi Liu c d, Hanghuan Yu a, Mufan Xiao a, Hongbo Cheng a, Yining Zhang a, Ben Han a
a School of Materials Science and Engineering, Shenyang Ligong University, Shenyang, 110159, China
b Liaoning, Provincial Engineering, Research Center for High-Value Utilization of Magnesite, Yingkou, 115014, China
c School of Resources & Civil Engineering, Northeastern University, Shenyang, 110819, China
d KU Leuven Bruges, Department of Materials Engineering, Surface and Interface Engineered Materials, 8200, Bruges, Belgium
10.1016/j.partic.2026.08.007
Volume 117, October 2026, Pages 463-476
Received 14 June 2026, Revised 25 July 2026, Accepted 9 August 2026, Available online 13 August 2026, Version of Record 21 August 2026.
E-mail: wangyulian@sylu.edu.cn
Highlights

• Nesquehonite was prepared from low-grade magnesite-derived Mg(HCO3)2 solution.

• Five regression models were compared for predicting nesquehonite aspect ratio.

• SVR showed the best prediction accuracy among the tested models.

• SHAP identified stirring rate as the dominant variable controlling aspect ratio.

• Favorable ranges for high-aspect-ratio nesquehonite were clarified.


Abstract

The aspect ratio of nesquehonite is critical to its morphology-dependent applications, yet its nonlinear dependence on coupled crystallization parameters remains poorly understood. Here, an interpretable machine-learning framework was developed using 212 valid experimental records of nesquehonite prepared from low-grade magnesite-derived Mg(HCO3)2. Five regression models, namely support vector regression (SVR), gradient boosting decision tree (GBDT), XGBoost, random forest and linear regression, were evaluated. Among them, SVR achieved the best performance, with a held-out test-set R2 of 0.9937 and RMSE of 0.3568, together with a repeated cross-validation R2 of 0.9070 ± 0.0512. Four validation experiments conducted under previously unseen process conditions yielded a mean relative prediction error of 3.11%, with all errors below 5%. SHAP analysis ranked the influential variables as stirring rate > reaction temperature > reaction time > pH > MgCl2 dosage and indicated that high stirring rates, moderate temperatures and reaction times, near-neutral pH, and moderate MgCl2 dosages favored high-aspect-ratio crystals. This framework provides a quantitative and interpretable tool for screening nesquehonite crystallization conditions within the investigated process domain.

Graphical abstract
Keywords
Nesquehonite; Aspect ratio; Magnesite; Mg(HCO3)2; Machine learning; SVR; SHAP
文章导读