The design of new materials, especially complex alloys like high-entropy alloys (HEAs), traditionally relies on extensive literature reviews. However, the exponential growth of scientific publications makes this process increasingly challenging and time-consuming. Text mining (TM), a powerful artificial intelligence technique, offers a potential solution for automating this process. Existing TM methods, however, suffer from a significant limitation: they primarily identify materials already present within the training corpus, hindering the discovery of truly novel materials. This paper addresses this limitation by introducing a novel concept: "context similarity." Instead of relying solely on the explicit presence of materials in the literature, the approach focuses on identifying chemical elements that frequently appear together in the context of alloy design. This captures the implicit knowledge and experience embedded within scientific publications, effectively expanding the search space beyond materials explicitly described. The study utilizes a large corpus of 6.4 million scientific abstracts to build a TM model that captures these contextual relationships between chemical elements, ultimately accelerating the design of ultrahigh-entropy alloys and multi-component materials. The high-throughput screening process is then further refined through the application of Integrated Computational Materials Engineering (ICME) methods.

Text mining (TM) has emerged as a powerful tool in various scientific fields, including materials science, political science, and public health. In materials science, TM shows promise for automating materials discovery, particularly in the context of high- and medium-entropy alloys (HEAs and MEAs). Existing TM methods often utilize word embedding algorithms that represent words as vectors. The cosine similarity between vectors measures the semantic similarity of words. This approach has been effective in identifying similar alloys, but its reliance on existing data limits its ability to predict truly novel materials. For instance, increasing the frequency of a specific alloy in the training data substantially improves its ranking in similarity searches. This demonstrates the predictive power of such models, but also highlights the limitation of only finding alloys already represented in the corpus. This study aims to overcome this limitation through the introduction of "context similarity," a new approach that leverages the contextual information present in the literature to identify promising chemical element combinations for the design of novel HEAs.

This study employs a skip-gram word embedding model, a type of neural network, trained on a corpus of 6.4 million materials-related abstracts, including a weighted emphasis on abstracts concerning metallic materials. The model represents words (in this case, chemical elements) as vectors in a high-dimensional space, where semantically similar words have vectors closer to each other. The cosine similarity between these vectors quantifies the "context similarity" between elements. The model addresses the challenge of inconsistent alloy naming conventions by alphabetizing the elements in each alloy. The researchers address potential bias from the overrepresentation of metallic materials by using a transfer learning approach, training the model on all available text and then fine-tuning it with the metallic materials data. Accurate extraction of named entities (alloy names) is crucial and requires special handling to account for variations in the representation of alloys. To design HEAs, two different methods were employed using the generated word vectors. The first method starts with a preferred element and identifies its most similar elements based on cosine similarity. The second method considers all elements equally, averaging the pairwise cosine similarity of all elements in a candidate alloy. Both methods yield rankings of promising alloy candidates. The study also employed a thermodynamics-based rule (γ ≥ 1), previously published by the authors, and calculated via density functional theory, to further refine the selection of likely solid solution HEAs. This parameter assesses the relative stability of a multicomponent alloy compared to its constituent binary systems. In addition, a knowledge graph (KG) was developed for a more efficient search for existing HEAs, standardizing alloy naming to prevent redundant alloy discovery and synthesis efforts. The authors further integrated their TM-based approach with existing Integrated Computational Materials Engineering (ICME) methods such as calculations of mass density and solid solution strengthening, to further screen and refine their predictions. This hybrid approach combines data-driven discovery with physics-based simulations to provide a more comprehensive alloy design workflow. The process begins with a vast pool of 2.6 million potential six- and seven-component alloys which is systematically reduced by filtering criteria based on context similarity (S>0.6), thermodynamic stability (γ >1), and mass density (ρ < 7.8 g/cm³), ultimately resulting in a shortlist of 494 promising candidates.

The "context similarity" approach successfully identified known HEAs like the Cantor (CoCrFeMnNi) and Senkov (TiZrNbHfTa) alloys as top candidates, even before their experimental discovery. The model accurately predicted the rise in importance of the Cantor alloy by reflecting the increased number of publications focused on it and its subsystems. The method effectively screened for promising six- and seven-component lightweight HEAs, identifying approximately 500 candidates from a pool of 2.6 million potential alloys using a three-step filtering process based on context similarity, thermodynamic stability, and density. The context similarity (S) was found to be strongly correlated with the previously developed thermodynamics-based parameter (γ), indicating a consistent measure of solid solution formation. The analysis of body-centered-cubic (BCC) HEAs demonstrated that the Senkov alloy consistently ranked highly across models trained on data from different years, showcasing the predictive power of the approach. Similarly, analysis of face-centered-cubic (FCC) HEAs showed that the Cantor alloy was ranked among the top candidates years before its experimental discovery. Solid solution strengthening calculations using a model from Varvenne et al. demonstrated that the identified alloys possess favorable mechanical properties. This combination of TM with established ICME methods allows for a closed-loop materials design approach.

This study demonstrates that incorporating "context similarity" into text mining overcomes the inherent limitations of traditional TM methods that are confined to the existing data. The successful identification of established HEAs (Cantor and Senkov alloys) before their discovery highlights the predictive power of the approach and its ability to explore the vast compositional space of HEAs beyond those already known. The strong correlation between the context similarity (S) and the thermodynamic parameter (γ) provides further validation, suggesting that the TM model captures relevant physical and chemical relationships. The integration of the TM-based pre-screening with ICME methods provides a powerful, multi-faceted approach to alloy design, bridging the gap between data-driven discovery and physics-based simulations. The ability to generate a shortlist of promising six- and seven-component alloys, which expands beyond the traditional focus on five-component HEAs, demonstrates the scalability and potential of this approach for designing ultrahigh-entropy alloys. Future research could focus on exploring additional filtering criteria within the ICME framework, enhancing the predictive power of the method and refining the identification of optimal compositions.

This paper presents a novel text-mining based method for designing high-component high-entropy alloys by utilizing the concept of "context similarity." This approach overcomes the limitation of traditional methods by discovering alloys not explicitly present in the training data. The successful prediction of established HEAs and the identification of numerous promising new candidates demonstrate the significant potential of this method for accelerating the discovery and design of advanced materials. Future work could incorporate additional data sources, explore more sophisticated TM techniques, and further integrate the method with advanced ICME simulations to refine the process and expand its application to a broader range of materials.

While the study demonstrates the effectiveness of the proposed approach, there are limitations to consider. The accuracy of the model relies heavily on the quality and completeness of the training corpus. The weights assigned to different parts of the corpus might require further optimization. The thermodynamics-based rule used for filtering may not be universally applicable to all alloy systems. Finally, experimental validation is crucial to confirm the predicted properties of the identified HEA candidates.