Scholarlink

<p>To make an accurate prediction on the liquidus temperature of stainless steel, an integrated model with two layers ensemble architecture on basis of XGBoost, RF and GBRT machine learning (ML) algorithms was established. The prediction precision on validation dataset is 10 to 20 pct higher than the individual models regardless of training database or feature engineering, and the novel combination on the basic algorithms should be the explanation. Two types of feature engineering were built up according to common mathematical operation and specialized metallurgical principles. The prediction accuracy is further enhanced by around 10 pct for each model when additional metallurgical features are included, and the items of empirical equation, Ni<sub>eq</sub>/Cr<sub>eq</sub>, Ni<sub>eq</sub>, Mn×Ni, Mn<sup>2</sup>, Mn present the significant influence. The models driven by thermodynamic and measurement dual-database exhibit the best prediction capacity, and the precision is additionally improved by at least 5 pct compared to the single database trained models. The most precise prediction on the validation dataset by in the present work is with 84 pct data below ±&#xa0;5&#xa0;°C deviation and 100 pct data below ±&#xa0;10&#xa0;°C deviation, much higher than the result in reported work and empirical equation or thermodynamic calculation. It is regarded as competent in the determination on liquidus of stainless steel for the plant parameters.</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

Machine Learning Prediction on Liquidus Temperature for Stainless Steel by Thermodynamics and Measurement Dual-Database-Driven Model

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Chenguang Su,
  • Huashun Chen,
  • Peng Lan,
  • Liang Zhang,
  • Cheng Wang

摘要

To make an accurate prediction on the liquidus temperature of stainless steel, an integrated model with two layers ensemble architecture on basis of XGBoost, RF and GBRT machine learning (ML) algorithms was established. The prediction precision on validation dataset is 10 to 20 pct higher than the individual models regardless of training database or feature engineering, and the novel combination on the basic algorithms should be the explanation. Two types of feature engineering were built up according to common mathematical operation and specialized metallurgical principles. The prediction accuracy is further enhanced by around 10 pct for each model when additional metallurgical features are included, and the items of empirical equation, Nieq/Creq, Nieq, Mn×Ni, Mn2, Mn present the significant influence. The models driven by thermodynamic and measurement dual-database exhibit the best prediction capacity, and the precision is additionally improved by at least 5 pct compared to the single database trained models. The most precise prediction on the validation dataset by in the present work is with 84 pct data below ± 5 °C deviation and 100 pct data below ± 10 °C deviation, much higher than the result in reported work and empirical equation or thermodynamic calculation. It is regarded as competent in the determination on liquidus of stainless steel for the plant parameters.