An active machine learning approach for optimal design of magnesium alloys using Bayesian optimisation.

Area of Science:

• Materials Science
• Computational Materials Science
• Alloy Design

Background:

• Magnesium (Mg) alloys are crucial for applications requiring lightweight materials with specific mechanical properties.
• Designing Mg alloys with targeted performance often involves complex, multi-objective optimization challenges.
• Existing methods may struggle with prediction accuracy and data dependency in alloy design.

Purpose of the Study:

• To develop and validate a multi-objective Bayesian optimization workflow for the efficient design of Mg alloys.
• To enhance the prediction accuracy of alloy properties by incorporating uncertainty quantification.
• To create a user-friendly tool for deploying optimal Mg-alloy design strategies.

Main Methods:

• A multi-objective Bayesian optimization framework was implemented.
• A probabilistic Gaussian process regressor model was trained using an active learning loop.
• The upper confidence bound acquisition function was employed to balance exploration and exploitation.
• Regret analysis was used to validate the sequential design strategy's performance.

Main Results:

• The workflow successfully identified optimal Mg-alloy compositions by iteratively refining predictions.
• The approach mitigated prediction errors by considering both predicted values and their uncertainties.
• The developed method demonstrated improved efficiency in exploring the design space compared to traditional methods.

Conclusions:

• The presented Bayesian optimization workflow offers an effective strategy for targeted Mg-alloy design.
• The integration of uncertainty quantification enhances the reliability of alloy property predictions.
• A web-based tool with a GUI facilitates the practical application of this optimal Mg-alloy design approach.