Scholarlink

<p>Accurate and rapid control of slag viscosity is pivotal for ensuring stable production and maximizing efficiency in contemporary blast furnace ironmaking, especially given the increasing reliance on low-grade iron ores. Purely data-driven models suffer from the critical limitation of “black-box” characteristics and a deficit in physical consistency. To surmount these challenges, this study proposes a novel method for high-fidelity, dynamic blast furnace slag viscosity prediction based on the deep integration of a self-adaptive physics-informed neural network (SAPINN) and the classic Urbain phenomenological model. Taking operational parameters as inputs, the Urbain model, calculated from chemical composition, is introduced into the loss function as a mechanism-derived physical constraint. A self-adaptive dynamic optimization framework meticulously adjusts the constraint weight via an annealing growth strategy, robustly integrating data-driven and physical mechanism approaches. Experimental results demonstrate that SAPINN significantly outperforms established baseline methods, including RNN and LSTM, achieving a superior coefficient of determination <i>R</i><sup>2</sup> of 0.96. The model also exhibits outstanding robustness across a wide temperature range of 1763–1803&#xa0;K. This study conclusively confirms that embedding physical models as prior knowledge substantially increases predictive accuracy and interpretability, providing a new approach for slag viscosity prediction.</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

Real-time Dynamic Prediction of Blast Furnace Slag Viscosity Based on an Adaptive Physical Information Neural Network

  • 期刊论文

  • 发表日期：

    2026年6月1日

查看全文
  • Bin Zhu,
  • Jindong Zhou,
  • Nanpeng Zou,
  • Chenyu Xu,
  • Wei Xiong,
  • Xuegong Bi

摘要

Accurate and rapid control of slag viscosity is pivotal for ensuring stable production and maximizing efficiency in contemporary blast furnace ironmaking, especially given the increasing reliance on low-grade iron ores. Purely data-driven models suffer from the critical limitation of “black-box” characteristics and a deficit in physical consistency. To surmount these challenges, this study proposes a novel method for high-fidelity, dynamic blast furnace slag viscosity prediction based on the deep integration of a self-adaptive physics-informed neural network (SAPINN) and the classic Urbain phenomenological model. Taking operational parameters as inputs, the Urbain model, calculated from chemical composition, is introduced into the loss function as a mechanism-derived physical constraint. A self-adaptive dynamic optimization framework meticulously adjusts the constraint weight via an annealing growth strategy, robustly integrating data-driven and physical mechanism approaches. Experimental results demonstrate that SAPINN significantly outperforms established baseline methods, including RNN and LSTM, achieving a superior coefficient of determination R2 of 0.96. The model also exhibits outstanding robustness across a wide temperature range of 1763–1803 K. This study conclusively confirms that embedding physical models as prior knowledge substantially increases predictive accuracy and interpretability, providing a new approach for slag viscosity prediction.