Scholarlink

<p>In the metallurgical field, the surface tension of CaF<sub>2</sub>-based slag systems is critical for smelting process control and optimization, yet accurate prediction remains hindered by data sparsity and nonlinear coupling among components. We propose PI-TL-ResNet, a physics-guided transfer-learning ResNet that integrates physical mechanism constraints, deep residual architecture, and transfer-learning strategies. Pre-training on a large multi-component slag dataset and fine-tuning on a smaller target-component set mitigates data scarcity and improves generalization. Compared with physically based models, generic neural networks, and mainstream regression methods, PI-TL-ResNet delivers superior accuracy and stability, with most predictions within ± 10 mN/m—meeting industrial tolerance. For cases exceeding this tolerance, interval-constrained soft labels (±10 mN/m) are applied to redefine outputs within physically admissible neighborhoods. This approach smooths the learning targets near boundaries, suppresses overfitting to isolated points, and stabilizes predictions in low-density regions, thus enhancing model robustness without compromising precision. This study establishes a high-precision, generalizable, and physically consistent framework for predicting metallurgical properties, offering strong potential for industrial deployment.</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

Mechanism-Guided Transfer Learning: Mechanism-Guided Residual Networks for Surface Tension Prediction in Electroslag Remelting CaF2-Based Slag

  • 期刊论文

  • 发表日期：

    2025年12月15日

查看全文
  • Xi Chen,
  • Yanwu Dong,
  • Zhouhua Jiang,
  • Ao Wang,
  • Yuxiao Liu

摘要

In the metallurgical field, the surface tension of CaF2-based slag systems is critical for smelting process control and optimization, yet accurate prediction remains hindered by data sparsity and nonlinear coupling among components. We propose PI-TL-ResNet, a physics-guided transfer-learning ResNet that integrates physical mechanism constraints, deep residual architecture, and transfer-learning strategies. Pre-training on a large multi-component slag dataset and fine-tuning on a smaller target-component set mitigates data scarcity and improves generalization. Compared with physically based models, generic neural networks, and mainstream regression methods, PI-TL-ResNet delivers superior accuracy and stability, with most predictions within ± 10 mN/m—meeting industrial tolerance. For cases exceeding this tolerance, interval-constrained soft labels (±10 mN/m) are applied to redefine outputs within physically admissible neighborhoods. This approach smooths the learning targets near boundaries, suppresses overfitting to isolated points, and stabilizes predictions in low-density regions, thus enhancing model robustness without compromising precision. This study establishes a high-precision, generalizable, and physically consistent framework for predicting metallurgical properties, offering strong potential for industrial deployment.