Machine Learning Assisted Property Prediction and Optimization of Magnetic Compositionally Complex Alloys

Abstract

Soft magnetic alloys, indispensable in the electrification, automation, and intelligence of human society, are widely used in power electronics, such as the magnetic cores of transformers and motors. Emerging magnetic compositionally complex alloys (CCAs), such as metallic glasses (MGs) and high-entropy alloys (HEAs), exhibit significant commercial potential due to their exceptional magnetic and mechanical properties. With various alloying elements, the possible composition space of magnetic CCAs is enormous, which is challenging for the traditional trial-and-error strategy to develop magnetic CCAs with desired properties. This study aims to accelerate the design of magnetic CCAs by developing data-driven alloy design methods.

Firstly, machine learning models were trained to predict magnetic properties and glass-forming ability of MGs, specifically focusing on saturation magnetic flux density (B_s) and critical diameter (D_max). Datasets were established by collecting experimental data from published works on magnetic MGs. A general feature space was proposed and proved adaptive for training machine learning models to predict B_s and D_max. The predictive performance of the trained models surpassed that of traditional estimation methods based on physical knowledge. Furthermore, incorporating domain knowledge into feature design and selection substantially reduced the number of features without significantly decreasing prediction accuracy. In addition, the feature importance given by the trained machine learning model indicated a potential correlation between glass-forming ability and B_s.

Secondly, a multi-stage optimization strategy based on machine learning was proposed to accelerate the design of magnetic MGs with high B_s. The vast composition space was considerably narrowed through machine learning phase prediction models. Mixing entropy constraint was added as a user preference for further narrowing the composition space and designing high-entropy MGs. Utility functions based on the exploitation and exploration strategy were designed to find the global optimization solutions, i.e., alloying compositions. Experiments in the Fe-Co-Ni-Si-B alloy system were conducted for concept validation, and several new traditional Fe-based MGs and high-entropy MGs with high B_s were designed.

Lastly, a data-driven method was proposed to accelerate the multi-objective optimization of magnetic high-entropy alloys with a decent combination of magnetic and mechanical properties. Machine learning models were trained to predict saturated magnetization and hardness of magnetic high-entropy alloys. A voting strategy was proposed for feature selection to reduce model complexity and avoid overfitting. Multi-objective optimization algorithms were designed to search for optimal alloying compositions. The advantage of the multi-objective evolutionary method in directly identifying a set of optimal compositions with desired multiple properties was highlighted by comparison with traditional methods. As a proof of concept, several new high-performance magnetic Fe-Co-Ni-Al-Si high-entropy alloys with excellent combination performance were fabricated as suggested by the multi-objective optimization methods.

This study showed the great potential of machine learning in property prediction and optimization of magnetic CCAs. The data-driven method proposed in this work can also be applied to other alloy systems for property-oriented alloy design, significantly contributing to their accelerated design and engineering applications.

Date of Award	16 May 2024
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Guancun SHAN (External Supervisor) & Chan Hung SHEK (Supervisor)