Complex composition materials, such as high-entropy alloys (HEAs), have recently attracted a lot of interest from the research community, as their vast composition–property search space has the potential to find new materials with exotic physical properties, such as high mechanical strength. However, this vast searching space also poses challenges for effective designs. Machine learning (ML) can be used as an alternative for analyzing the complex HEA composition–property data; yet, the available data for ML training is sparse. While active learning — in which a ML surrogate model is updated iteratively with new data from a closed-loop experimental feedback — can potentially address this disadvantage, the common algorithms used in active learning, such as Gaussian process (GP) and Bayesian optimization (BO), are too expensive to be scaled to high-dimensional data for HEA design. In a recent work, Dierk Raabe and colleagues reported on a possible solution to address these challenges and applied active learning to HEA design.
The active learning framework consists of three main steps: candidate composition generation, physics-informed screening, and experimental feedback. In the first two steps, instead of using the conventional GP/BO process, the authors use generative models (GM) to project the high-dimensional composition data into an interpretable latent space for sampling and selecting candidates. These candidates are used in a two-stage ensemble regression, together with a rank–order strategy, to fine-select compositions for experimental synthesis (third step). The new experimental results are then iteratively fed into the first step as part of the active learning process. Interestingly, the authors demonstrated that this framework can help find two HEAs with record-low thermal expansion coefficients. It is worth mentioning that, compared with other methods, such as neural network models, for reducing data dimension in active learning, the latent space generated from GM models can provide physical knowledge for further analysis. Overall, the main ideas put forth in the proposed framework can not only be extended to other complex materials systems, such as superconductors and high-entropy perovskites, but also inspire new developments for other autonomous systems, such as spectroscopy.
This is a preview of subscription content, access via your institution