A seven-facet polyhedron captures the composition-only formation-energy landscape of inorganic solids

This work demonstrates that the convex hull of formation energies for solid compounds involving elements from hydrogen to uranium admits a remarkably simple description over the 92-dimensional space of chemical compositions, despite the enormous combinatorial complexity of possible atomic structures. By training an interpretable max-affine model directly on near-hull formation energies from the Materials Project density-functional theory (DFT) database, we find that the hull can be reconstructed to DFT accuracy using a polyhedron with only seven facets. These facets define seven chemically coherent materials classes, with just seven coefficients per element sufficing to capture the dominant energetic trends across composition space. Remarkably, this compact, composition-only representation generalizes far beyond bulk formation energies. Without retraining or structural input, the same model reproduces trends in DFT-calculated defect formation energies, captures experimentally observed elemental mixing correlations in high-entropy materials, and enables the construction and optimization of Pourbaix diagrams for electrochemical stability. Together, these results show that many materials properties governed by energy differences can be expressed as simple linear combinations of a small set of interpretable, element-specific parameters. The result is a bonding-geometry-free thermodynamic framework that unifies stability, defects, mixing, and electrochemistry, and enables rapid, interpretable screening across vast chemical spaces.

💡 Research Summary

The authors investigate whether the convex hull of formation energies for inorganic solids can be described by a simple, interpretable function of composition alone. Using the Materials Project database, they first train a shallow three‑layer neural network (512 units per layer, tanh activation) on 86,925 compounds whose DFT formation energies lie within 0.25 eV/atom of the convex hull. Despite its modest size, the network achieves a mean absolute error (MAE) of 0.044 eV/atom, indicating that the near‑hull energy landscape is far less complex than the underlying structural space would suggest.

Motivated by this observation, the authors introduce a max‑affine model of the form
E(x)=max₍f=1…F₎ (∑₍s=1₎⁹² a_{f,s} x_s + b_f),
where x is the 92‑dimensional elemental composition vector, a_{f,s} and b_f are learned coefficients, and F is the number of facets. Training on 51,068 bulk formation energies within 0.025 eV/atom of the hull, they evaluate MAE as a function of F. Accuracy saturates at F = 7, yielding an MAE of 0.076 eV/atom—well within typical DFT uncertainty—while using only 645 linear parameters (7 × 92 + 7).

Each of the seven facets corresponds to a chemically coherent class of materials. By grouping elements into four valence categories (s‑block, d/f‑block, early p‑block, late p‑block) and measuring how frequently an element appears in binaries with each category on a given facet, the authors construct facet‑resolved periodic‑table maps. Facet 1 captures lanthanide/actinide intermetallics (d/f–d/f or d/f–s bonding); facet 2 highlights classic alkali ionic salts; facet 3 contains transition‑metal chalcogenides and pnictides; facet 4 groups mixed ionic‑covalent oxides, nitrides, and fluorides; facet 5 is dominated by early transition‑metal intermetallics; facet 6 features alkali/alkaline‑earth with early p‑block elements (Zintl‑like semiconductors); and facet 7 reflects metallic alkali‑transition‑metal and alkali‑alkali binaries. The active facet for any composition x is f* = argmax_f (∑s a{f,s} x_s + b_f). The coefficients a_{f*,s} serve as “compositional chemical potentials” µ̃_s for that facet, providing a physically meaningful interpretation: all compounds on the same facet share identical µ̃ values and therefore exhibit similar trends in any property governed by energy differences.

The authors demonstrate the broad applicability of this composition‑only framework without any retraining. First, they predict point‑defect formation energies (substitutional, vacancy, interstitial) across diverse hosts using only the facet‑specific µ̃ values, reproducing DFT defect energies with errors ≈0.1 eV. Second, they capture experimentally observed spatial correlations in high‑entropy phosphide nanoparticles, showing that the facet‑derived bonding patterns align with measured mixing behavior. Third, they construct Pourbaix diagrams for electrochemical stability by combining µ̃ values with aqueous ion potentials, obtaining stability regions that match full DFT‑based Pourbaix calculations.

In summary, the paper reveals that the 92‑dimensional composition space of inorganic solids can be compressed into a seven‑facet polyhedron, each facet representing a distinct chemical class. This compact, interpretable representation achieves DFT‑level accuracy for bulk formation energies and generalizes to defect energetics, high‑entropy mixing, and electrochemical stability. By eliminating the need for explicit structural information, the approach offers a fast, transparent screening tool for vast chemical spaces, potentially accelerating the discovery and design of new materials where energy differences dominate performance.