Machine learning-driven closed-loop discovery of hard multiple principal element alloys

Multi-principal element alloys (MPEAs) form a unique class of alloys that typically consist of three or more principal elements in near-equimolar ratios. These alloys are often sought-after due to their exceptional mechanical properties, such as high hardness, strength, and thermal stability, which make them highly desirable for various applications. However, a significant challenge in the discovery and design of MPEAs lies in the vast and complex compositional space they occupy, which is both high-dimensional and sparsely explored. Traditional methods for identifying MPEAs with desirable properties tend to rely heavily on trial-and-error experimentation, which is time-consuming and inefficient. In this work, we apply an active learning approach, PAL 2.0, utilizing a Bayesian optimization approach as a means to significantly accelerate the discovery of MPEAs with particularly high hardness. For this accelerated discovery, PAL 2.0 can be operated as a closed-loop framework, closely integrating physics-based Gaussian process models with experimental validation. Our methodology enables the model to intelligently navigate the compositional space and make informed decisions about the most promising alloys to synthesize and test. Based on recommendations made by PAL 2.0, we successfully synthesized 20 new MPEAs through a rapid arc-melting process. Among these 20, we identified two new alloys with exceptionally high Vickers hardness values of 1269 and 1263. While the original training dataset had only three MPEAs with hardness above 1000, our method recognized five additional compositions with a hardness over 1000, thereby doubling the number of very hard MPEAs. The most striking discovery is the appearance of silicon and tantalum together in the alloys, an "out of distribution" combination not seen in any high hardness alloy within the original training dataset. This is highly sought occurrence indicating when an AI system is capable of suggesting solutions that it was not trained to handle. This study demonstrates the power of PAL 2.0 as a fast, efficient, and scalable tool for the discovery of materials with optimal properties, significantly reducing the parameter space for promising new materials candidates. This work not only accelerated the development of high-performance MPEAs but, since PAL's use is materials-agnostic, it offers a pathway to explore other complex, high-dimensional material spaces, paving the way for creative advancements in materials science.