硅酸盐学报

预处理方法和机器学习算法对玻璃材料性能预测的影响：比较研究和优化策略

PDF下载

王思雨 ¹^,² , 孙元成 ¹ , 余睿 ² , 杜秀蓉 ¹ , 周利生 ¹ , 刘星 ¹ , 陈志威 ¹^,³

硅酸盐学报 | 先进玻璃材料技术与产业发展专题—研究论文 2025,53(10): 2870-2881

收起

硅酸盐学报 |先进玻璃材料技术与产业发展专题—研究论文 2025 , 53 (10) : 2870 -2881

预处理方法和机器学习算法对玻璃材料性能预测的影响：比较研究和优化策略

全屏

王思雨¹^,², 孙元成¹, 余睿², 杜秀蓉¹, 周利生¹, 刘星¹, 陈志威¹^,³

作者信息

^1．中国建筑材料科学研究总院有限公司，北京 100024

^2．武汉理工大学材料与微电子学院，武汉 430070

^3．武汉理工大学材料科学与工程学院，武汉 430070

通讯作者:

孙元成（1984—），男，博士，正高级工程师。

作者简介:

王思雨（2000—），女，硕士研究生。

Impacts of Preprocessing Methods and Machine Learning Algorithms on Glass Material Property Prediction: Comparative Study and Optimization Strategies

Siyu WANG¹^,², Yuancheng SUN¹, Rui YU², Xiurong DU¹, Lisheng ZHOU¹, Xing LIU¹, Zhiwei CHEN¹^,³

Affiliations

^1.China Building Materials Academy, Beijing 100024, China

^2.School of Materials and Microelectronics, Wuhan University of Technology, Wuhan 430070, China

^3.Materials Science and Engineering, Wuhan University of Technology, Wuhan 430070, China

Corresponding Author:

SUN Yuancheng (1984-), male, Ph.D., Professor. E-mail: sunych@foxmail.com

WANG Siyu (2000-), female, Master candidate. E-mail: wsiyu@whut.edu.cn

出版时间: 2025-10-20 doi: 10.14062/j.issn.0454-5648.20250283

文章导航

摘要

收起

系统研究了不同预处理方法和机器学习算法对玻璃材料性能预测的影响。实验基于一组包含4种组分作为输入、7种性能作为输出的数据集，首先通过标准化与3种插值方法因不同顺序生成6组样本，随后分别采用随机森林（RF）、极限梯度提升（XGB）、分类回归树（CART）、K近邻（KNN）、支持向量机（SVM）和人工神经网络（ANN）对每组样本的每一列输出进行单目标实验，并对所有输出进行综合实验。在此基础上，选择RF与SVM进行超参数调优，比较了随机搜索、网格搜索和贝叶斯优化的效果。结果表明，预处理顺序对误差指标（MSE/MAE）的敏感性高于决定系数（R²），建议优先采用“B-RF插值”组合以平衡精度与稳定性；而数据缺失率>10%时，插值方法对R²的影响权重显著增加，高缺失场景（>40%）需特别注意插值方法对SHAP特征排序的干扰；算法性能呈RF＝XGB>KNN=ANN>SVM≥CART的稳定层级，建议首选集成算法构建预测模型；组分分析显示Na₂O主导密度与折射率，B₂O₃主导玻璃化转变温度T_g，SiO₂调控膨胀软化温度M_g，而热膨胀系数存在阶段性；贝叶斯优化可有效提升SVM，而集成算法建议采用遗传算法强化参数搜索。

关键词

机器学习算法比较 / 性能预测 / 超参数调优 / 数据预处理

Abstract

收起

Introduction

The existing research on modeling glass properties has some limitations in critical modeling stages, thus hindering predictive accuracy and interpretability. Primarily, systematic data pre-processing is often neglected, i.e., most literature relies solely on basic data cleaning (e.g., removing outliers and duplicates), lacking thorough investigation into optimal missing value imputation methods and the crucial sequence of imputation versus feature standardization. Secondly, hyperparameter optimization strategies remain simplistic. As some advanced techniques like random search or Bayesian optimization are introduced, studies generally overlook a critical adaptation of tuning methods to specific algorithm characteristics and a lack comparative analysis of hyperparameter optimization approaches across different models. This potentially restricts the performance of hyperparameter-sensitive models like Support Vector Machines (SVM). To address these challenges, this study was to propose a comprehensive, optimized framework for glass property prediction.

Methods

This research established a fully optimized framework for glass property prediction. The methodology comprised four integrated stages, i.e., 1) Systematic Data Preprocessing: For dataset A, the framework rigorously evaluated the combinatorial effects of three missing value imputation strategies (i.e., Median Imputation, Polynomial Imputation, Random Forest Imputation) and two standardization sequences (i.e., Imputation followed by Standardization; Standardization followed by Imputation); 2) Predictive Model Construction and Evaluation: Six distinct machine learning algorithms, which were chosen for their varied inductive biases (i.e., Random Forest-RF, XGBoost-XGB, Classification and Regression Trees-CART, K-Nearest Neighbors-KNN, Support Vector Machines-SVM, Artificial Neural Networks-ANN), were employed. The model predictive efficacy was rigorously quantified using a triple evaluation metric (i.e., Mean Squared Error-MSE, Mean Absolute Error-MAE, Coefficient of Determination-R²). The model generalization was validated by a novel dataset (A); 3) Interpretable Analysis: The SHAP (SHapley Additive exPlanations) framework was applied to elucidate the contribution paths of key glass components towards the predicted properties; and 4) Hyperparameter Optimization Strategy Analysis: For representative models SVM and RF, a comparative experiment was conducted by three hyperparameter tuning methods, i.e., Random Search, Grid Search, and Bayesian Optimization. This stage was to establish adaptable criteria matching hyperparameter tuning strategies to model architectures.

Results and Discussion

The results show that the sequence of standardization and imputation significantly affects model performance metrics, with a substantially larger impact observed on MSE and MAE (up to 49.64% variation), compared to R². Conversely, the choice of imputation method exerts a greater influence on R² (up to 63.68%) under conditions of high missing data rates (i.e., >10%). The optimal preprocessing is standardization-first followed by Random Forest Imputation. The SHAP analysis reveals that preprocessing sequence does not alter the fundamental trends in feature contributions, but significantly changes the numerical distribution morphology of SHAP values, thus introducing systematic bias. Imputation method, particularly at high missing rates (i.e., >40%), directly affects the ranking of feature importance derived from SHAP.

The comparative performance of the six algorithms, evaluated across metrics, follows an order of RF ≈ XGB > KNN ≈ ANN >SVM≥ CART. The RF and XGB demonstrate a superior performance, leveraging their ensemble learning advantages. The predictive accuracy of model (as measured by MSE, MAE, and R²) is robust to varying missing data rates. However, the interpretability of model predictions via SHAP becomes less reliable at high missing rates.

The SHAP analysis indicates distinct governing components for different properties, sometimes diverging from conventional understanding. While B₂O₃ dominates the Glass Transition Temperature (T_g), M_g content is primarily governed by SiO₂, and followed by B₂O₃. Na₂O is a dominant contributor to both Density and Refractive Index. The primary contributors to the thermal expansion coefficient (TEC) are temperature-range dependent. Na₂O followed by K₂O dominates at 20-300 ℃, whereas SiO₂ followed by Na₂O dominates at 20-400 ℃.

The use of hyperparameter tuning method proves critical, especially for hyperparameter-sensitive models. The Bayesian optimization can obtain the most substantial improvements for the lower-accuracy SVM model, achieving a maximum R² improvement of 76.56%. In contrast, its impact on the high-performing ensemble algorithm RF is marginal, indicating that alternative, potentially more explorative methods like Genetic Algorithms (GA) can be more suitable for such models. This highlights a necessity for model-specific hyperparameter optimization strategies.

Conclusions

This study proposed and validated a comprehensive optimization framework for glass property prediction, addressing key limitations in data preprocessing and hyperparameter tuning prevalent in the existing literature. The systematic evaluation demonstrated that preprocessing choices, particularly the sequence of operations and imputation method under high missing data scenarios, had profound and quantifiable impacts on both predictive accuracy (MSE, MAE, R²) and the reliability of interpretability frameworks like SHAP. Among the evaluated algorithms, RF and XGB consistently outperformed others. Critically, the SHAP analysis provided insights into composition-property relationships, thus revealing temperature-dependent dominance for TEC and challenging assumptions regarding M_g's primary drivers. Finally, the hyperparameter optimization experiments established that tuning strategy efficacy could be highly model-dependent. The Bayesian optimization offered significant gains for sensitive models like SVM but was less effective for robust ensembles like RF, necessitating tailored approaches. This framework could provide a validated, end-to-end methodology for enhancing the accuracy, reliability, and interpretability of predictive models in glass science and potentially other material domains.

Key words

machine learning algorithm comparison / property prediction / hyperparameter tuning / data preprocessing

引用本文

王思雨, 孙元成, 余睿, 杜秀蓉, 周利生, 刘星, 陈志威. 预处理方法和机器学习算法对玻璃材料性能预测的影响：比较研究和优化策略. 硅酸盐学报, 2025 , 53 (10) : 2870 -2881 . DOI: 10.14062/j.issn.0454-5648.20250283

Siyu WANG, Yuancheng SUN, Rui YU, Xiurong DU, Lisheng ZHOU, Xing LIU, Zhiwei CHEN. Impacts of Preprocessing Methods and Machine Learning Algorithms on Glass Material Property Prediction: Comparative Study and Optimization Strategies[J]. Journal of the Chinese Ceramic Society, 2025 , 53 (10) : 2870 -2881 . DOI: 10.14062/j.issn.0454-5648.20250283

基金

收起

CNBM关键技术研发计划(2023SJYL03)
中国建材总院自立科研项目(ZT-20)

参考文献

收起

文献

收起

参考文献引证文献

排序方式：

2025年第53卷第10期

PDF下载

引用本文

BibTeX

文章信息

doi: 10.14062/j.issn.0454-5648.20250283

接收时间：2025-04-14
首发时间：2026-10-21
出版时间：2025-10-20

补充材料

相关文章

文章信息

作者

出版历史

收稿日期：2025-04-14
修回日期：2025-05-17

基金

CNBM关键技术研发计划(2023SJYL03)

中国建材总院自立科研项目(ZT-20)

作者信息

^1．中国建筑材料科学研究总院有限公司，北京 100024

^2．武汉理工大学材料与微电子学院，武汉 430070

^3．武汉理工大学材料科学与工程学院，武汉 430070

通讯作者:

孙元成（1984—），男，博士，正高级工程师。

参考文献

分享链接

https://castjournals.cast.org.cn/joweb/gsyxb/CN/10.14062/j.issn.0454-5648.20250283

分享至

全文二维码

扫描看全文

引用本文

BibTeX

本文的引用情况

2种不同金属材料的力学参数

科 Family	属数 Number of genus	种数 Number of species	占总种数比例 Percentage of total species (%)	属 Genus	种数 Number of species	占总种数比例 Percentage of total species (%)
鹅膏菌科Amanitaceae	2	11	5.26	鹅膏菌属 Amanita	10	4.78
小菇科 Mycenaceae	2	12	5.74	丝盖伞属 Inocybe	5	2.39
多孔菌科 Polyporaceae	8	14	6.70	蜡蘑属 Laccaria	5	2.39
红菇科 Russulaceae	3	23	11.00	小皮伞属 Marasmius	6	2.87
				小菇属 Mycena	11	5.26
				光柄菇属 Pluteus	5	2.39
				红菇属 Russula	17	8.13
				栓菌属 Trametes	5	2.39

关闭全屏

BibTeX
EndNote
RefWorks
TxT