We present a fine-tuned machine learning interatomic potential (MLIP) for all inorganic halide perovskites $\left(A{A}^{\prime }B{B}^{\prime }{X}_6\right)$, which we use to investigate stability trends across 195 968 charge-balanced compositions. The MLIP accurately predicts the relative stability of cubic and distorted phases, capturing octahedral tilting effects that often favor lower-energy or nonperovskite ground states. For the entire set of examined materials, only 2% of compositions are metastable $(\mathrm{\Delta }{\mathrm{H}}_d< 100 \mathrm{meV}/\text{atom})$ in the high-symmetry cubic phase. To rationalize this result, we analyze 23 730 ${\mathrm{Cs}{\mathrm{A}}^{\prime }\mathrm{B}{\mathrm{B}}^{\prime }\mathrm{X}}_6$ compositions and show that stability in the cubic phase correlates with the size mismatch between $A$ and $A^{\prime }$ cations. Because the MLIP yields DFT-quality optimized structures, single-point DFT can be used efficiently to compute energies and band gaps. Assuming the cubic structure instead of the ground state structure, which typically contains various degrees of octahedral distortions, can lead to changes in the band-gap energy of up to 0.50 eV. The MLIP enables vibrational free-energy calculations via phonon density of states, allowing temperature-dependent phase stability predictions. These reveal several compounds that are predicted to transition from the (distorted) ground state phase (at $T=0\mathrm{K}$) to high-symmetry cubic phases at elevated temperatures. Overall, this work provides a tool and dataset of DFT quality $A{A}^{\prime }B{B}^{\prime }{X}_6$ structures for analyzing key $\mathrm{structure}-\mathrm{property}$ relationships and supports the discovery of new perovskite compositions with more realistic stability and band-gap estimates than considering the cubic phase alone.

Close

PrevNext

View figure in article

• Crystalline systems
• Metal halides
• Density functional theory
• Machine learning