Composition/structure directed search for new chalcogenide compounds

This work presents a simple scheme for finding new crystalline compounds by adapting structure types from neighbor atoms compounds. The approach is demonstrated for the selenide and sulfide families of binary compounds. It predicts ten new compounds that are not currently included in the inorganic crystal structure database (ICSD). The compounds primarily originated from a small search domain that includes near neighbors. Comparison with extended searches that include structures from binary systems of more remote atoms in the periodic table demonstrate the relative efficiency of near neighbor screening. This points at the possibility of using similar directed searches as a heuristic rule for efficiently finding new stable compounds in additional compound families.

💡 Research Summary

The paper introduces a pragmatic, composition‑structure directed search strategy for discovering new crystalline compounds, focusing on binary sulfide (A‑S) and selenide (A‑Se) families. Recognizing that the Inorganic Crystal Structure Database (ICSD) contains a vast number of entries but only about 13 000 distinct structure prototypes, the authors argue that exhaustive enumeration of all prototypes is computationally prohibitive. Instead, they propose to restrict the search space by borrowing structure types from chemically “neighboring” systems—those that share the same A‑element but differ in the chalcogen (S ↔ Se) or even the oxide (O) counterpart.

The methodology proceeds in three hierarchical stages. In the minimal search domain, the authors enrich the convex‑hull calculations for a given A‑S (or A‑Se) system with missing prototypes from the corresponding A‑Se (or A‑S) and A‑O binaries. This step already yields two previously unknown stable compounds (Ti₃S₄ and Cu₃S₂) and ten metastable candidates within 0.03 eV/atom of the hull. The second stage expands the domain to include “near‑neighbor” A‑elements—those residing in the same or adjacent columns of the periodic table. Adding 204 such prototypes leads to five additional stable compounds (Ta₃Se₄, CuSe, ZrSe, Zr₃Se₄, Zr₂Se₃) and 37 more metastable phases. The third stage, a full‑periodic‑table extension, incorporates all possible A‑elements for three representative systems (TiS, CuS, TaSe), adding 436 prototypes. This exhaustive search produces only one extra stable compound (Ti₅S₈) and 13 metastable structures, demonstrating diminishing returns as the search domain widens.

All calculations employ VASP with the SCAN meta‑GGA functional, PAW pseudopotentials, and stringent convergence criteria (energy change <10⁻⁵ eV, k‑point density up to 6000 NKPRA). Formation enthalpies are derived from fully relaxed total energies, and convex hulls are constructed using SciPy scripts. Structures within 0.03 eV/atom above the hull are considered potentially synthesizable under typical metallurgical temperatures (~300 K).

A key insight is the markedly higher overlap (≈33 %) between sulfide and selenide structure types compared with the modest overlap (≈10 %) between sulfide/selenide and oxide prototypes. Consequently, borrowing structures from the chemically analogous chalcogen system is far more productive than from oxides. Moreover, the success rate per DFT calculation drops dramatically from ~2 % in the minimal domain to <0.3 % in the full‑periodic‑table domain, underscoring the efficiency of the neighbor‑based heuristic.

The authors also test an alternative extension where the chalcogen itself is replaced by other elements (e.g., using Ti‑As prototypes for TiS). This yields a single new compound (Ti₃S) and confirms that the farther the chemical substitution, the less likely a stable structure will emerge.

In discussion, the paper emphasizes that the “structure‑composition proximity” principle provides a rational, low‑cost pathway to explore vast compositional spaces without resorting to blind high‑throughput searches. By focusing on prototypes that have already proven viable in chemically similar systems, researchers can dramatically reduce the number of required DFT evaluations while still uncovering a substantial fraction of the discoverable compounds. The approach is readily generalizable to other families such as carbides, nitrides, phosphides, or halides, and could be integrated with machine‑learning models to further prioritize promising candidates.

Overall, the study demonstrates that a carefully curated, neighbor‑centric prototype library can serve as an effective launchpad for materials discovery, balancing computational tractability with a high yield of novel, potentially synthesizable compounds.