Abstract
As global energy demands rise, the advancement of new energy technologies increasingly relies on the development of metals that can endure extreme pressures, temperatures, and fluxes of energetic particles and photons, as well as aggressive chemical reactions. One way to assist in the design and manufacturing of metals for the future is by learning from their past. Here we track the progress of metallic materials for extreme environments in the past 35 years using the text mining method, which allows us to discover patterns from a large scale of literature in the field. Specifically, we leverage transfer learning and dynamic word embeddings. Approximately one million relevant abstracts ranging from 1989 to 2023 were collected from the Web of Science. The literature was then mapped to a 200-dimensional vector space, generating time-series word embeddings across six time periods. Subsequent orthogonal Procrustes analysis was employed to align and compare vectors across these periods, overcoming challenges posed by training randomness and the non-uniqueness of singular value decomposition. This enabled the comparison of the semantic evolution of terms related to metals under extreme conditions. The model's performance was evaluated using inputs categorized into materials, properties, and applications, demonstrating its ability to identify relevant metallic materials to the three input categories. The study also revealed the temporal changes in keyword associations, indicating shifts in research focus or industrial interest towards high-performance alloys for applications in aerospace and biomedical engineering, among others. This showcases the model's capability to track the progress in metallic materials for extreme environments over time.