Composition Engineering on the Local Structure and Viscosity of the CaO-SiO2-Al2O3-P2O5-FeO Slag by Machine Learning Methods

Ziyu Lyu, Chao Gu, Ziyang Lyu, Yanping BaoCrystals 2022 · DOI

TL;DR

This paper tackles the challenge of accurately predicting the viscosity of complex industrial slags, which is crucial for optimizing steelmaking processes. By combining molecular dynamics simulations with machine learning, specifically Gradient Boosting Decision Trees, the researchers developed a highly accurate model that can guide slag composition for desired viscosity properties.

What Problem Does This Paper Address?

High-temperature viscosity measurements of multicomponent industrial slags are notoriously difficult, expensive, and prone to inaccuracies due to experimental challenges. This lack of reliable viscosity data hinders the precise engineering of slag compositions for optimal performance in iron and steel production, impacting fluidity, mass transfer, and separation efficiency. Existing semi-empirical models often have limited applicability or data coverage for these complex systems.

Methods

The study employs a multi-pronged approach:

  • Molecular Dynamics (MD) Simulations: The researchers used MD simulations to generate viscosity data for various CaO-SiO₂-FeO-Al₂O₃-P₂O₅ slag compositions. This involved building atomic models, defining interatomic potentials (Born-Mayer-Huggins and Lennard-Jones), and simulating the system under relevant conditions (1873 K). The Reverse Non-Equilibrium Molecular Dynamics (RNEMD) method was specifically used to calculate shear viscosity.
  • Semi-Empirical Models: Several established semi-empirical models (NPL, Urbain, Kondratiev & Jak) were used to predict slag viscosity and compare their performance against MD results.
  • Machine Learning (ML): Four ML algorithms were investigated for viscosity prediction: Artificial Neural Network (ANN), Random Forest (RF), Support Vector Machine (SVM), and Gradient Boosting Decision Tree (GBDT). The GBDT model, in particular, was optimized and used to develop a predictive tool.

Datasets

The primary dataset for training and evaluating the ML models was generated through:

  • Molecular Dynamics Simulations: 64 sets of MD simulations were conducted for the CaO-SiO₂-FeO-Al₂O₃-P₂O₅ slag system, with varying Al₂O₃ (0-30%) and P₂O₅ (0-30%) content, while keeping CaO/SiO₂ ratio at 1.5 and FeO at 30%.
  • Semi-Empirical Model Predictions: Viscosity predictions from the NPL, Urbain, and Kondratiev & Jak models served as a basis for comparison and applicability assessment.

The paper does not introduce a new benchmark dataset in the traditional sense but rather generates a valuable synthetic dataset through MD simulations for a specific slag system.

Key Insights and Ideas

  • MD as a Viable Data Generator: Molecular dynamics simulations proved to be a powerful tool for generating reliable viscosity data for complex slag systems where experimental data is scarce.
  • Structural Insights Drive Viscosity: The study highlights how the local structure, particularly the types and connectivity of oxygen atoms (bridging vs. non-bridging), significantly influences slag viscosity. Increased Al₂O₃ and P₂O₅ content generally leads to more polymerized structures (higher bridging oxygen) and thus higher viscosity.
  • GBDT Outperforms Other ML Models: Gradient Boosting Decision Trees emerged as the most accurate ML method for predicting slag viscosity in this study, demonstrating its effectiveness in capturing the complex relationships between composition and viscosity.
  • Urbain Model's Relative Reliability: Among the semi-empirical models, the Urbain model showed the best correlation with structural insights and MD predictions, suggesting its relative applicability to this slag system.
  • Visualizing Composition-Viscosity Relationships: The development of a color-map illustrating the impact of Al₂O₃ and P₂O₅ on viscosity provides a practical tool for engineers to design slag compositions for specific viscosity targets.

Key Results

  • GBDT Accuracy: The Gradient Boosting Decision Tree (GBDT) model achieved the lowest Root Mean Squared Error (RMSE) of 0.000424539, indicating superior predictive performance for slag viscosity.
  • Structural Correlations: A strong positive correlation was found between bridging oxygen (BO) content and Al₂O₃ content, and a weaker positive correlation with P₂O₅ content. Conversely, non-bridging oxygen (NBO) and free oxygen (FO) content decreased with increasing Al₂O₃ and P₂O₅.
  • Model Applicability: The Urbain model was identified as the most applicable semi-empirical model for predicting the viscosity of the CaO-SiO₂-FeO-Al₂O₃-P₂O₅ slag system, based on Pearson correlation coefficient analysis.
  • Composition-Viscosity Map: A detailed color-map and iso-viscosity lines were generated, allowing for the prediction of viscosity based on Al₂O₃ and P₂O₅ content, aiding in composition engineering.

Limitations and Future Work

  • Solid Fraction Ignored: The MD simulations considered the slag as a liquid state, neglecting any potential solid fraction, which might introduce some inaccuracies.
  • Limited Experimental Validation: The study acknowledges the absence of experimental data for direct validation of the MD and ML predictions, relying instead on comparisons between different modeling approaches.
  • Focus on Specific Components: While Al₂O₃ and P₂O₅ were the primary focus for ML prediction, the influence of other components and temperature variations on viscosity could be further explored.

Future work could involve experimental validation of the developed ML model and the generated viscosity map. Further exploration into the impact of other slag components and temperature on viscosity, as well as the integration of more advanced ML architectures, could enhance the predictive capabilities.

Summary updated

Abstract

Due to the high cost and low accuracy of high-temperature tests, the viscosity data for multicomponent slag systems is difficult to be obtained precisely. Therefore, it is important to fulfill the viscosity database of the multicomponent slag systems via reasonable methods with lower costs. In this study, a viscosity prediction method based on the machine learning method was proposed for the CaO-SiO2-FeO-Al2O3-P2O5 quinary slag system. To provide valid data for the machine learning model, the viscosity predicted by the molecular dynamic method and multiple semi-empirical models were compared to verify the applicability of these methods to the slag system. Different machine learning models were also developed. The results showed that the prediction results from the gradient boosting decision tree method were the most accurate for the CaO-SiO2-FeO-Al2O3-P2O5 quinary slag system. Based on this method, a color-map concerning the numerical effect of Al2O3 and P2O5 contents and slag viscosity is provided, which also provides assistance for the composition engineering to fulfill a certain demand on the viscosity design.
BibTeX
@inproceedings{lyu2022composition,
  author = {Ziyu Lyu and Chao Gu and Ziyang Lyu and Yanping Bao},
  title = {{Composition Engineering on the Local Structure and Viscosity of the CaO-SiO2-Al2O3-P2O5-FeO Slag by Machine Learning Methods}},
  booktitle = {Crystals},
  year = {2022},
  doi = {10.3390/cryst12101338},
}
Figure