A bold new idea in inorganic chemistry could change how we think about solid materials: a recently discovered homologous series of barium compounds forms a seemingly endless sequence of related structures with predictable unit cells. The team behind the finding suggests this kind of structure–composition relationship could make machine learning models more capable of discovering new inorganic materials.
In chemistry, homologous series are common in organic systems. They describe sequences of compounds that share a repeating unit and can be captured by a general formula—think of straight-chain alkanes and alkenes. Such orderly sequences also appear, though less often, in solid-state inorganic materials. Notable examples include non-stoichiometric titanium oxides and the two-dimensional halide perovskites used in solar cells.
The Northwestern University team, led by inorganic chemist Mercouri Kanatzidis, began with the compound barium antimony telluride (BaSbTe3). They gradually replaced portions of tellurium with sulfur, a sister element in Group 16. At first glance, one might expect sulfur and tellurium to mix randomly within the crystal lattice, forming a straightforward solid solution. Yet the researchers showed, on theoretical grounds, that sulfur’s higher electronegativity leads it to favor more electron-rich sites in the lattice. This substitution triggers secondary changes in the structure: as sulfur content rises, the remaining telluride becomes more electron-poor. In other words, a simple solid solution doesn’t form. Instead, the material rearranges itself to create a new member of the homologous series. Kanatzidis describes this as a remarkable departure from conventional expectations—the system discovers an alternative, ordered arrangement rather than a random mixture.
Building the series
The team synthesized ten members of this homologous series, observing increasingly complex structures along the way. The final member, BaSbSTe2, exhibits a charge density wave, a type of electronic instability. Materials with charge density waves often become superconductors at high temperatures or under low pressure, which excites researchers about the potential to rationally predict new superconductors—something not yet possible with current techniques.
Implications for machine learning
Kanatzidis notes that while machine learning is increasingly used to design new materials, it excels most at predicting materials within established structure types. As a result, AI has enjoyed more success in fields with well-developed chemical principles—like organic chemistry—than in the discovery of genuinely novel solid-state materials. Findings such as this new phase homology could provide rich, informative training data that improve AI models’ predictive power for inorganic systems.
Independent assessments highlight the work’s significance. Materials chemist Leslie Schoop of Princeton University calls the study “very solid” and emphasizes the need to probe the new materials for properties worthy of deeper investigation. Schoop has voiced concern about autonomous materials discovery, but she also believes that the kind of structure–composition relationships revealed by Kanatzidis and colleagues could help AI tackle challenges in inorganic chemistry. “We need to start integrating this kind of reasoning—one that leads to real material breakthroughs—into AI algorithms,” she argues.
Bottom line
What makes this development exciting is not just a single new compound, but a framework: a predictable, expandable sequence where changing one parameter nudges the structure along a known path. This opens up new possibilities for understanding how composition shapes structure and, in turn, how those structures influence properties. It also provides a practical path for training AI to recognize and exploit analogous relationships in other inorganic systems.
Question for readers
Could this approach accelerate the discovery of practical superconductors or other technologically important materials? What safeguards or checks would you propose to ensure AI-guided discoveries translate into real, usable breakthroughs rather than simply elegant theoretical constructs?
