Toward the design of ultrahigh-entropy alloys via mining six million texts

Text mining (TM) is an artificial intelligence method to analyze and discover scientific knowledge in literature and has been applied across domains including materials science, political science, and public health. In materials, given the vast growth in publications on high- and medium-entropy alloys (HEAs/MEAs), TM presents an opportunity for automated knowledge extraction. Prior TM approaches using word embeddings capture latent relationships (e.g., structural associations like FCC/BCC) but are limited to identifying materials already present in corpora, constraining discovery of novel alloys. A central challenge in HEA design is selecting mutually soluble, similar elements. The authors propose a "context-similar elements" concept that leverages word-embedding models trained on 6.4 million abstracts to quantify similarity between chemical elements based on how researchers use them in alloy design contexts. This context-sensitive similarity goes beyond simple elemental properties by encoding collective research experience and usage, enabling more efficient exploration of complex, high-component alloy spaces and assisting even less experienced researchers by replacing manual literature review with model-driven guidance.

The paper situates its work within TM and word-embedding literature, noting the success of algorithms such as skip-gram, GloVe, and BERT in extracting semantic relationships from text. Prior work demonstrated that unsupervised word embeddings can capture latent knowledge from materials literature and that named entity recognition is critical for large-scale extraction. In HEAs/MEAs, over 10,000 papers exist, providing a rich yet noisy corpus. TM models trained on domain-focused corpora can be more predictive since irrelevant text reduces keyword co-occurrence signal. Despite these advances, TM methods traditionally only retrieve known materials from corpora and lack mechanisms to propose materials absent from the training text, motivating the context similarity approach to extend TM into generative alloy design.

Corpora and preprocessing: The authors compiled 6.4 million abstracts via Elsevier Scopus API using ISSN lists (with additional manual inclusion of key metallic journals). Abstracts were stored in JSON with metadata. To emphasize metals, abstracts from key metallic journals were duplicated (effectively double weighting). Entity normalization addressed HEA naming variability: alloy constituent symbols were alphabetized (e.g., CoCrFeMnNi = CrCoFeMnNi = NiMnFeCrCo) and formats standardized across variants (e.g., hyphenated or with subscripts). Word-embedding model: A skip-gram neural network with one hidden layer was trained. Words were represented as one-hot vectors; the model maximized the probability of context words within a window size of 8 (left and right). Training used 30 epochs with 200-dimensional embeddings, with tests at 300 dims and window size 10 showing similar rankings but higher cost. Transfer learning was used: first train on all available scientific texts, then fine-tune on metallic materials abstracts. The trained model encodes corpus information in the hidden-layer matrix M; word vectors v_i are obtained as M·w_i. Cosine similarity between vectors quantifies context similarity S_ij. Context similarity concepts for alloy design: Two usage modes were defined: (1) seed-element mode—start with a target element A and select the most similar elements by cosine similarity to form HEA candidates; (2) equal-participation mode—given a set of M elements, compute the average pairwise cosine similarity across all element pairs in an N-component alloy and rank candidates by the mean S. Temporal models: To evaluate predictive ability ahead of discovery, yearly models were trained using publications from a given year (except 2003, which used all abstracts up to 2003), then used to rank candidate alloys for those years. Thermodynamic screening: A previously proposed thermodynamic criterion γ was computed, comparing the approximate Gibbs free energy of the N-component system to those of all binaries. The criterion γ ≥ 1 indicates a likely single-phase solid solution. The lattice misfit parameter δ was also used as an auxiliary descriptor to improve solid-solution prediction. ICME integration and high-throughput screening: For ultrahigh-entropy alloy (6- and 7-component) design limited to 30 transition metals, a three-step filter was applied: (i) S > 0.6; (ii) γ > 1; (iii) density ρ < 7.8 g/cm³. This reduced 2.6 million candidates to 494. Mechanical property modeling: Solid-solution strengthening σ_y at 300 K was estimated for top FCC candidates using the Varvenne et al. model for solid-solution strengthening in FCC HEAs. DFT calculations: VASP (v5.4.4) with GGA-PBE was used to obtain optimal volumes in FCC for pure elements (V, Cr, Mn, Fe, Co, Ni, Cu). PAW potentials, 400 eV plane-wave cutoff, 10×10×10 k-mesh, 4-atom supercells, volume-only relaxations to 10^-5 eV energy convergence were employed. Knowledge graph: An alloy knowledge graph (alloyKG) standardized alloy names by alphabetical ordering of elements and linked alloys to DOIs and authors, enabling retrieval of prior reports despite variable naming.

- Context-similar elements form clusters in embedding space reflecting how researchers use elements; e.g., Al groups near Mn and Cr in latent space despite PTE separation. - The method recovers known FCC/BCC HEAs: it identified Cantor (CoCrFeMnNi) and Senkov (TiZrNbHfTa) alloys as top candidates using only literature available prior to their discoveries. Cantor ranked second before 2004; Senkov ranked among top three across years and first in 2011, 2016, 2018. - The averaged context similarity S strongly and positively correlates with the thermodynamic γ parameter for both BCC and FCC sets, indicating higher S implies greater likelihood of single-phase solid solution; combining γ with lattice misfit δ shows only a small fraction of 3–5 component alloys are true solid solutions even among high-S candidates. - BCC design (Ti, Zr, Nb, Mo, Hf, Ta, W; 21 five-component combos): Top predicted alloys include TiZrNbHfTa (Senkov), TiZrNbMoHf, TiZrMoHfTa, ZrNbMoHfTa, TiNbMoHfTa, TiZrNbMoTa; Senkov consistently ranks top in later years. - FCC design (V–Cu third-period transition metals): Besides Cantor, MnFeCoNiCu and CrFeCoNiCu were highly ranked and have been synthesized; S–γ correlation confirmed for FCC set. - Mechanical property predictions: For top FCC candidates, predicted σ_y at 300 K shows CrFeCoNiCu with the largest strengthening effect (reported as −290 MPa), followed by CrMnFeCoCu; Cantor predicted ~146 MPa, consistent with experimental ~125 MPa. - High-throughput ultrahigh-entropy screening: From 2.6 million 6- and 7-component candidates (30 transition metals), filtering by S > 0.6 (to ~0.5M), γ > 1 (to ~36k), and ρ < 7.8 g/cm³ yielded 494 promising lightweight HEAs. • Top by S: six-component TiCrFeCoNiMo; seven-component TiCrFeCoNiCuZn; six-component TiFeCoNiCuZn. • Top by γ: six-component VCrMnFeCoNi; VCrMnFeCoCu; VCrMnFeNiCu. • Lowest density: six-component ScTiZnZrAgCd; TiVCrMnFeZn; TiVCrMnCuZn. - Hyperparameter tests (window size 8→10, 200→300 dims) produced stable rankings (e.g., Senkov remained top) with higher training cost, indicating robustness of the 200-dim, window-8 model.

The context similarity approach overcomes a core limitation of traditional TM by enabling proposal of alloys not explicitly present in the corpus. By leveraging solid-solution behavior typical of HEAs and focusing on element-level similarity independent of composition initially, the method efficiently narrows the vast compositional space. The strong correlation between S and the thermodynamic γ parameter suggests that the embedding-derived similarity captures physically meaningful tendencies toward single-phase solid solutions. Integration with ICME tools (thermodynamics, kinetics, structure-property modeling) enables a hybrid design pipeline: large-scale, automatic literature-informed pre-screening followed by physics-based filtering to select candidates with targeted properties and to avoid duplicates using a standardized alloy knowledge graph. Temporal analyses show the method tracks field evolution (e.g., rise of Cantor-related elements in Fe’s similarity neighborhood post-2014), and hyperparameter studies indicate stable, computationally efficient configurations suffice for alloy design. Overall, the approach facilitates closed-loop materials discovery, expanding to higher component counts and assisting users with varying domain expertise.

The study introduces a context similarity concept using word embeddings trained on 6.4 million abstracts to guide HEA element selection and alloy design. It successfully retrospectively identifies hallmark HEAs (Cantor and Senkov) ahead of their discoveries, establishes a strong linkage between text-derived similarity and a thermodynamics-based solid-solution criterion, and demonstrates an integrated TM–ICME workflow for high-throughput screening. Applying this to 6- and 7-component lightweight HEAs reduced 2.6 million candidates to 494 promising alloys and provided concrete top-ranked suggestions across similarity, thermodynamics, and density criteria. The approach offers a general, scalable tool for discovering ultrahigh-entropy and multicomponent materials. Future work may extend to broader property targets, incorporate additional physics-informed descriptors, expand corpora and entity normalization, and experimentally validate and iterate on prioritized candidates within a closed-loop design framework.

- Dependence on literature: Embedding-based similarity reflects how elements are used in published research; biases or gaps in the literature may skew similarity and candidate selection. - Abstract-only training: Relying primarily on abstracts can miss detailed compositional or processing nuances present in full texts. - Naming standardization: Despite alphabetization, non-standard HEA naming remains a challenge broadly; comprehensive standardization is needed for universal retrieval. - Predictive scope: High S correlates with γ, but many high-S candidates are still multi-phase; additional descriptors (e.g., δ, kinetics) are required to refine to single-phase alloys. - Domain weighting and transfer learning: While metals-focused weighting improves relevance, optimal weighting is heuristic and may omit beneficial cross-domain signals. - Data access constraints: Large-scale abstract retrieval depends on API quotas and access policies, which may limit reproducibility timelines.