Closed-Loop Materials Discovery for Multi-Principal Element Alloys

Advanced functional materials have widespread applications in energy, health, and space exploration, but their discovery is limited by a large combinatorial space that is expensive to characterize experimentally or computationally. One such material class is multi-principal element alloys (MPEAs), which are alloys consisting of three or more elements in significant amounts. These alloys are desirable for their excellent mechanical properties such as high hardness, strength, and thermal stability, but synthesizing and testing them is costly. To address this problem, we used PAL 2.0, an algorithm we previously developed that utilizes a physics-based prior in concert with Bayesian optimization to efficiently search the material space. We demonstrated the effectiveness of this method through closed-loop materials discovery of high-hardness MPEAs with experimental collaborators at the Johns Hopkins University Applied Physics Laboratory. Based on known MPEA data, we used PAL 2.0 to recommend a list of high-hardness alloys, some of which the experimentalists then manufactured and tested. Using these results, we retrained PAL 2.0’s models and derived new recommendations, thereby continuing the cycle. To expand the search space beyond confirmed MPEA compositions, we developed an alloy generation algorithm based on k-means clustering to further identify novel and previously undiscovered alloy compositions with high hardness. Based on recommendations from PAL 2.0, we manufactured thirteen new MPEAs. While our initial 350-alloy dataset contained only four alloys with high hardness (Vickers hardness value > 1000), we were able to identify five more high-hardness alloys within just three feedback cycles. Among those five alloys, two had exceptionally high Vickers hardness values of 1263 and 1269. This demonstrates that, using a closed-loop approach, PAL 2.0 can greatly accelerate the discovery of new materials with optimal properties from a complex, high-dimensional material space.

Figure 1: Manufactured MPEA recommendations. Vickers hardness values are shown for each recommended MPEA that was manufactured. The color and shape of each data point denote the closed-loop cycle during which the corresponding alloy was manufactured. Two manufactured alloys with particularly low hardness were omitted.