Skip to content

Paving the Way for High-Entropy Alloy Machine Learning

Fri, 13 August, 2021

Induction Melting Paves the Way for High-Entropy Alloy Machine Learning

The high-entropy alloy (HEA) concept was based on the idea that high mixing entropy can promote formation of stable single-phase microstructures. In more recent years, the concept has been broadened to “Compositionally complex alloys” (CCA), with combinations of minimum three elements in ranges of 5 to 35%. These CCAs can offer an infinite pool of new alloys with unexplored properties.

As exploring all these new alloys experimentally is impossible, the FORGE project will train a machine learning (ML) model with the ultimate goal to develop a machine learning based decision support system to help material scientists and engineers discover novel CCAs with targeted properties with a minimum number of trial and errors (see previous post for more info blogpost)

Train and feed the machine

Despite the increasing number of papers reporting on more service performance properties for CCA materials, the availability of this data is still very limited and scattered when compared to more fundamental properties, such as hardness. Therefore, a combined approach of machine learning with experimental screening is being used in the project for the specific service performances of interest.

The objective of this experimental approach is also twofold. A first model was trained, based on mainly hardness data from literature. In parallel, three pools of 6-8 elements each were selected for different performance targets under study in this project (CO2 corrosion, H2 embrittlement, wear and corrosion/sintering). Several thousand combinations of elements out of these three pools were calculated via the machine learning model and ranked according to their hardness. Out of these, 30 new alloys were selected by the different partners to cover a wide range of compositions and hardness values.

These alloys are now being synthetized by means of small-scale casting at OCAS. Hardness evaluation is being compared with the machine learning model predictions to further test the model on new alloys.

In a second step, the actual performance targets in the project, such as corrosion, H2 embrittlement and wear/mechanical damage, as well as microstructural characteristics of the alloys will be evaluated and collected in a database to further train the ML model. To construct this database with advanced properties, two complementary CCA preparation routes routes will be applied. As mentioned above, partner OCAS is carrying out small-scale induction melting while partner EMPA will apply combinatorial PVD. Induction melting allows the screening of a variety of different elements; combinations of three to six elements are being tested from a pool of 14 different ones.

Figure 1. Induction melting
Figure 1. Induction melting

Pioneering unknown alloy processing at lab-scale

Although OCAS has a long tradition in induction melting and casting ferrous and several non-ferrous alloys, melting CCAs adds to the challenge. In fact, elements are not only present in large quantities in CCAs but several of them, such as Ti, Cr, V having melting points around 1900°C and elements such as Nb, Mo, Ta, W having melting points in the range of 2500 up to 3600°C, also possess high or even ultra-high melting points.

Moreover, some elements such as Ti, are very reactive to most crucibles. To avoid time-consuming iterations in a high-throughput flow, anticipating the selection of raw materials or pre-alloys and crucibles is crucial.

Currently, 30 alloys have been synthetized, the majority via induction melting in quantities of approximately 0.5 kg each. The graph below highlights their success rate versus their complexity in induction melting. As one alloy failed upon casting via induction due to too strong interaction with the crucible (red), four other alloys were cast via arc melting (open circles). Four alloys (orange) broke upon solidification, most likely due to their inherent brittleness, not being able to accommodate the thermal strains upon cooling.

Figure 2. Success rate upon casting versus complexity
Figure 2. Success rate upon casting versus complexity

High-throughput rolling allows fast sampling and enables advanced characterisation

In a second step, OCAS has rolled the casts into sheets, via a dedicated high-throughput flow for small casts. This offers the possibility to break up the cast structure and homogenise the microstructure. Moreover, these sheets allow fast sampling, as required for further testing, by means of spark erosion or waterjet cutting. On these samples, the more advanced properties such as corrosion, H2 embrittlement and wear/mechanical damage are currently being evaluated.

Hardness in the as-cast condition was already measured at OCAS for 19 alloys and the results were compared to the ML model predictions. The agreement depends strongly on representation of the alloy family in the database used to train the ML model. A good agreement was found for the largely represented alloys, whereas alloys with a very limited representation showed large deviations.

Despite the huge number of available hardness values in literature for HEA and CCA, the need for a fast way to provide experimental data remains to support databases and machine learning models.

Figure 3. Hardness data versus predictions
Figure 3. Hardness data versus predictions

Article courtesy: OCAS

G-DXBWG8050M