Scholarlink

<p>The physicochemical properties of slags play a critical role in most steelmaking processes. However, available datasets on slag properties are typically small, and the predictive accuracy of existing machine learning algorithms is often limited under such conditions. In this study, we developed self-constructed high-dimensional spatial regionalized learning models that effectively overcome the low accuracy of conventional machine learning in small-sample scenarios. The model not only improves predictive performance but also quantifies the specific contributions of each component at every point in the compositional space to the targeted physicochemical properties. To demonstrate its capability, we applied the models to predict the viscosity and electrical conductivity of the CaF<sub>2</sub>–Al<sub>2</sub>O<sub>3</sub>–CaO–MgO–SiO<sub>2</sub> slag system. For viscosity, where a relatively large dataset was available, the model accurately predicted 98 pct of the data using only 2 pct of the dataset for training. For electrical conductivity, where the dataset was extremely limited, we employed a sliding-window cross-validation approach. In both cases, the predictive accuracy of our model was consistently and significantly higher than that of conventional machine learning methods. Based on the current results, the proposed model shows higher prediction accuracy compared with traditional machine learning methods when composition is used as the input variable under a fixed temperature condition.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Self-Constructed High-Dimensional Spatial Regionalized Learning Models for Predicting Key Physical Properties of Slag: A Case Study on CaF2–Al2O3–CaO–MgO–SiO2 Slag System

  • 期刊论文

  • 发表日期：

    2026年4月1日

查看全文
  • Yi-Ming Li,
  • Zhou-Hua Jiang,
  • Ao Wang,
  • Shan Cui,
  • Jia-Long Tian

摘要

The physicochemical properties of slags play a critical role in most steelmaking processes. However, available datasets on slag properties are typically small, and the predictive accuracy of existing machine learning algorithms is often limited under such conditions. In this study, we developed self-constructed high-dimensional spatial regionalized learning models that effectively overcome the low accuracy of conventional machine learning in small-sample scenarios. The model not only improves predictive performance but also quantifies the specific contributions of each component at every point in the compositional space to the targeted physicochemical properties. To demonstrate its capability, we applied the models to predict the viscosity and electrical conductivity of the CaF2–Al2O3–CaO–MgO–SiO2 slag system. For viscosity, where a relatively large dataset was available, the model accurately predicted 98 pct of the data using only 2 pct of the dataset for training. For electrical conductivity, where the dataset was extremely limited, we employed a sliding-window cross-validation approach. In both cases, the predictive accuracy of our model was consistently and significantly higher than that of conventional machine learning methods. Based on the current results, the proposed model shows higher prediction accuracy compared with traditional machine learning methods when composition is used as the input variable under a fixed temperature condition.