Magnesiothermic Reduction of Silica: A Machine Learning Study

Area of Science:

• Materials Science
• Chemical Engineering
• Computational Materials Science

Background:

• Magnesiothermic reduction of silica is crucial for producing silicon materials.
• Traditional models struggle to capture the complex kinetics and product encapsulation observed experimentally.
• Existing thermochemical software is insufficient for accurately describing the reaction pathways.

Purpose of the Study:

• To develop a more accurate model for the magnesiothermic reduction of silica.
• To apply a machine learning approach to describe the complex reaction kinetics.
• To predict the influence of process parameters on reduction products.

Main Methods:

• Experimental studies of magnesiothermic reduction with varying Mg/SiO2 ratios, temperatures, and times.
• Development of a physics-informed Gaussian process machine (GPM) using hybrid datasets.
• Incorporation of experimental data and thermochemical equilibrium calculations as boundary conditions for the GPM.

Main Results:

• The developed GPM achieved a high regression score of 0.9665 on hybrid data.
• The GPM successfully predicted the effects of Mg-SiO2 ratios, temperatures, and reaction times on reduction products.
• Experimental validation confirmed the GPM's accuracy for interpolating observed data.

Conclusions:

• A physics-informed Gaussian process machine (GPM) provides an effective framework for modeling complex magnesiothermic reduction reactions.
• The GPM overcomes limitations of traditional models and thermochemical software.
• This approach enables accurate prediction of reaction products, guiding future material design and process optimization.