New machine learning approach for high-entropy alloy discovery

Using a new physics-informed machine learning approach, researchers discovered two new high-entropy alloys with extremely low thermal expansion, a new study reports. The approach could represent a powerful new tool for materials discovery and design. To fuel the rapid development of new technologies, researchers and engineers are ceaselessly searching for novel materials with specific properties, including alloys. Conventional alloys generally consist of a single principal metal element combined with smaller fractions of other elements. More recently, researchers have begun looking for alloys that combine similar amounts of multiple principal elements. These alloys – also called high-entropy alloys (HEAs) – greatly expand the search space of alloys for material design. However, given the vast amount of element combinations possible, discovering those with valuable properties is extremely challenging and simply cannot be managed by conventional alloy design methods. Here, Ziyuan Rao and colleagues present a machine learning approach to screen this practically infinite design space to identify Invar alloys – valuable alloys with extremely low thermal expansion. Even using sparse data, Rao et al. show that their physics-informed artificial neural network can learn and predict the complex relationship between elements and their collective properties in an alloy. Depending on the training data used, specific properties of an alloy, such as its thermal expansion coefficient, can be targeted and predicted. Using this approach, the authors identified 17 new Invar HEAs out of millions of possible candidates. Experimental testing of two of the compositionally complex high-entropy Invar alloys demonstrated extremely low thermal expansion coefficients (2 x 10^-6 per degree kelvin at 200 kelvin), which is much lower than the current known record of HEAs. “With the accumulation of experimental datasets, development of the optimization models for artificial neural networks, and a better understanding of the physics underlying the relationships between composition processing, microstructure, and property, a universal virtual laboratory may one day become a reality,” write Qing-Miao Hu and Rui Yang in a related Perspective.