Einstein-Roscoe regression for the slag viscosity prediction problem in steelmaking.

Area of Science:

• Materials Science
• Chemical Engineering
• Machine Learning

Background:

• Classical machine learning models often lack interpretability and fail to extrapolate beyond training data.
• Natural sciences require robust, interpretable models that elucidate underlying mechanisms and predict phenomena in unseen domains.
• Viscosity prediction in steelmaking is crucial but challenging due to complex material interactions.

Purpose of the Study:

• To develop an interpretable machine learning model for viscosity prediction in steelmaking.
• To enable extrapolation to unseen domains, providing robust predictions beyond the training data.
• To incorporate transfer learning for efficient parameter estimation using auxiliary measurements.

Main Methods:

• Proposing Einstein-Roscoe regression (ERR) to learn coefficients of the Einstein-Roscoe equation.
• Employing a Gaussian process-based transfer learning framework to utilize auxiliary measurements.
• Validating the approach using viscosity measurements in high-temperature slag suspension systems.

Main Results:

• ERR demonstrated superior performance in extrapolation tasks compared to other machine learning methods.
• The model achieved favorable interpolation results, comparable to existing approaches.
• Transfer learning using room-temperature data improved accuracy for high-temperature viscosity predictions.

Conclusions:

• Einstein-Roscoe regression offers an interpretable and extrapolative alternative to classical machine learning for materials science.
• Transfer learning enhances model accuracy and efficiency, particularly when dealing with expensive or unavailable measurements.
• The proposed method provides a robust framework for viscosity prediction in challenging industrial applications like steelmaking.