Machine Learning as a Tool for Refining Diffusion Sintering Models of Al2O3 Ceramics

Abstract

The primary objective of this research is to evaluate the predictive accuracy of classical deterministic models versus machine learning algorithms in determining the final grain size of alumina ceramics. This study aims to identify the specific limitations of the traditional Coble model within complex powder systems and establish high-precision alternatives for microstructural control. Furthermore, the research explores the potential for energy consumption optimization by refining the thermal and kinetic parameters of the sintering process through advanced data analysis.

This study conducts a comprehensive comparative analysis between the analytical Coble model and the Random Forest ensemble learning algorithm. The initial stage involved testing the classical model with standard literature constants (n=3, Q=395 kJ/mol), which revealed significant systematic errors and overestimations. Subsequently, the deterministic model was optimized using the least squares method, adjusting physical parameters to n=2 and Q=451 kJ/mol to better suit the experimental data. In parallel, a machine learning framework was developed and trained on experimental datasets, utilizing Mean Absolute Error (MAE) and R-squared (R²) metrics for performance evaluation. A critical component of the work was the Feature Importance analysis, which quantified the relative impact of temperature, sintering time, and MgO dopant concentration on grain growth kinetics. The developed software tool was tested for its ability to simulate densification saturation points, providing a basis for digital twin integration in ceramic manufacturing. The methodology bridges the gap between theoretical material science and practical industrial applications.

The results demonstrate that the Random Forest model significantly outperforms deterministic approaches, reducing prediction error by 25 times (MAE=1.65 μm) with an R² of 0.91. Feature Importance analysis revealed that sintering time (weight 0.485) exerts a more substantial influence on final microstructure than temperature (0.475) or MgO additives (0.040) for the studied dataset. This kinetic dominance suggests that peak sintering temperatures can be replaced by extended isothermal holding, potentially lowering energy consumption by 10–15%. The proposed ML-based approach provides a robust foundation for the development of digital twins, ensuring precise microstructural control while minimizing R&D costs.