This work combines metallurgical thermodynamics and kinetics to train a high-accuracy and low-cost deep neural network (DNN) prediction model for hydrogen content in vacuum oxygen decarburization (VOD) process. The mean absolute error of the trained DNN model was 0.0258 ppm, and the R² and root mean square errors were 0.9973 and 0.0329, respectively. The sensitivity of the model to single-feature and multi-feature coupling is tested, and the current VOD process is optimized through the model. It was found that by increasing the consumption of argon to 100 m³, controlling the initial hydrogen content below 8 ppm, and reducing the vacuum to 70 Pa, the final hydrogen content will be reduced to 0.99 ppm.

本工作结合冶金热力学和动力学，训练了一种高精度、低成本的深度神经网络（DNN）预测模型，用于真空氧脱碳（VOD）工艺中的氢含量。训练后的 DNN 模型的平均绝对误差为 0.0258 ppm，R ²和均方根误差分别为 0.9973 和 0.0329。测试了模型对单特征和多特征耦合的敏感性，通过模型优化了当前的VOD流程。研究发现，通过将氩气消耗量增加到100 m ³，将初始氢含量控制在8 ppm以下，将真空度降低到70 Pa，最终氢含量将降低到0.99 ppm。