Reveal the high-fidelity phase selection rules of high-entropy alloys: Combining CALPHAD and machine learning research

In recent years, high-entropy alloys (HEA) have attracted the attention of materials science and related fields due to their special mechanical and physical properties, such as high strength, thermal stability, wear resistance and radiation resistance. The excellent properties of high-entropy alloys are mainly attributed to its unique microstructure. Therefore, in the direction of its microstructure, its phase selection rule has always been a key issue in the research of high-entropy alloys. Although extensive experiments and calculations have been carried out over the years, this rule is still difficult to implement. At this stage, empirical methods are usually used to study the phase selection rule of HEA. However, due to the limited composition space and insufficient data points, it is necessary to use more Parameters to develop accurate HEA selection rules.

In recent years, machine learning (ML) methods have developed rapidly, which is expected to explore the forming process and stability of high-entropy alloys. Although certain results have been achieved on this basis, there are still some shortcomings. First, the machine learning model is trained on data sets with limited composition space. These data sets generally contain only a few hundred experimental data on the alloy microstructure of casting or forging. Secondly, the machine learning model is often a “black box”, and simple and effective rules are often not found in the design of high-entropy alloys. Therefore, for high-entropy alloy design, it is particularly necessary to obtain high-fidelity phase selection rules from machine learning. Based on this, in December 2020, the Institute of High-Performance Computing of the Singapore Institute of Science, Science and Technology generated more than eight elements including Al, Co, Cr, Cu, Fe, Mn, Ni, and Ti through phase diagram calculation technology. 300,000 data points and 15 description parameters were selected. Finally, the machine learning model was trained through the maximum gradient method, and the five most critical parameters were obtained as shown in Figure 1.

Additive manufacturing technology in
aviation
aerospace
,national defense,
car
and
Medical treatment
Continuous development in various applications such as implantation. Selective laser melting technology (SLM) and laser metal deposition (LMD) technology are the two main methods widely explored in additive manufacturing technology. However, due to the “step effect” and powder adhesion on the surface, the surface quality of the metal parts formed by SLM and LMD is very poor, and further surface treatment is required to be applied. However, traditional mechanical grinding and chemical polishing have many shortcomings: mechanical polishing is difficult to achieve on complex surfaces, and electrochemical polishing is a non-selective process. As a new type of surface treatment method, laser polishing has the advantages of short process duration, high repeatability, and no abrasives. However, for the surface of parts with large ripples formed by LMD, the small melting zone of laser polishing often cannot directly remove the surface ripples and Big bulge. The two-step laser surface treatment method proposed by Hong Shen can solve this problem well.

The overall prediction accuracy of the model is displayed in the lower right corner of the matrix. It can be seen that the accuracy of the training data set is as high as 99.95%, and the accuracy of the test data set is also 99.92%. Such a high accuracy rate indicates machine learning based on these five features. The model has a very good effect, which provides more effective rules and design tools for guiding the design of single-phase FCC and BCC high-entropy alloys and finding new single-phase alloy materials.

references:

BHANVADIA AA, FARLEY RT, NOH Y, et al. High-resolution stereolithography using a static liquid constrained interface [J]. Commun Mater, 2021, 2(1):

(Editor in charge: admin)