Scholarlink

<p>Accurate prediction of the temperature field is crucial for controlling the duplex microstructure in additive manufacturing of 2205 duplex stainless steel. Conventional numerical methods are limited by high computational cost, data-driven approaches often lack physical interpretability, and conventional physics-informed neural networks show restricted spatiotemporal modeling capability. A physics-informed encoder–decoder network with skip connections and convolutional gated recurrent units is developed to predict transient temperature evolution. The framework combines an encoder–decoder architecture based on convolutional neural networks with skip connections, while the encoder further incorporates convolutional gated recurrent unit layers to enhance spatiotemporal feature learning. Physical laws of heat transfer are incorporated into the loss function to maintain consistency with thermodynamic principles. The model attains a mean absolute percentage error of 1.37% for single-step prediction and maintains errors between 2% and 5.3% for 27-step forecasting. Based on online monitoring data, it achieves a 5.4 s ahead temperature prediction with only 0.2&#xa0;s computational latency. Ablation analysis indicates that skip connections and physical constraints contribute significantly to prediction accuracy, convergence, and stability.</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 temperature field prediction in 2205 duplex stainless steel additive manufacturing via physics-informed spatiotemporal network

  • 期刊论文

  • 发表日期：

    2026年5月27日

查看全文
  • Deng-Kui Ao,
  • Yong-Cheng Lin,
  • Mao-Lin Li,
  • Guan Liu,
  • Jia-Yu Yang,
  • Ning-Fu Zeng,
  • Miao Wan,
  • Gang Xiao,
  • Sai-Feng Peng

摘要

Accurate prediction of the temperature field is crucial for controlling the duplex microstructure in additive manufacturing of 2205 duplex stainless steel. Conventional numerical methods are limited by high computational cost, data-driven approaches often lack physical interpretability, and conventional physics-informed neural networks show restricted spatiotemporal modeling capability. A physics-informed encoder–decoder network with skip connections and convolutional gated recurrent units is developed to predict transient temperature evolution. The framework combines an encoder–decoder architecture based on convolutional neural networks with skip connections, while the encoder further incorporates convolutional gated recurrent unit layers to enhance spatiotemporal feature learning. Physical laws of heat transfer are incorporated into the loss function to maintain consistency with thermodynamic principles. The model attains a mean absolute percentage error of 1.37% for single-step prediction and maintains errors between 2% and 5.3% for 27-step forecasting. Based on online monitoring data, it achieves a 5.4 s ahead temperature prediction with only 0.2 s computational latency. Ablation analysis indicates that skip connections and physical constraints contribute significantly to prediction accuracy, convergence, and stability.