Machine learning reveals a temperature-dependent defect state missed at absolute zero

Removing one atom from a crystal can change how the material handles electrical charge. Calculating that defect at absolute zero, however, may miss a charge state that becomes favourable when the atoms vibrate. A study published in npj Computational Materials on 23 October uses machine learning within an atomistic calculation to examine that temperature dependence.[1]

The researchers studied vacancies in magnesium oxide, lithium fluoride and caesium tin bromide, or CsSnBr3. A vacancy is a missing atom; its possible charge states help determine how it interacts with electrons and holes. The energy boundaries at which the preferred charge state changes are known as charge-transition levels.[1][2]

A static zero-kelvin calculation compares defect energies without the finite-temperature contribution of atomic motion. Once the crystal is warm, different configurations have different vibrational free energies. Their relative stability can shift, and a charge state that was never the most favourable choice in the static calculation may become favoured within a particular range of electronic conditions.[1][2]

The team used machine-learned interatomic potentials to make the required sampling more efficient. These models approximate atomic energies and forces. Thermodynamic integration then estimates free-energy differences between charge states. AI’s role here is a numerical surrogate within a physics calculation, rather than a language model proposing a material from a verbal description.[1][2]

The published threshold and the older summary differ

For CsSnBr3, the journal article reports that a neutral vacancy charge state becomes thermodynamically stable above 40 kelvin, opening a range of Fermi levels absent at zero kelvin. The Fermi level describes the electronic conditions under which the competing charge states are compared. The finding concerns which state is most stable in the calculation; it does not mean a previously nonexistent particle or defect suddenly appears at that temperature.[1]

There is a version discrepancy in the public source record. The authors’ laboratory publication page describes the threshold as above 60 kelvin, while the journal’s early accepted article says above 40 kelvin. This report uses the journal figure and keeps the difference explicit. The journal identifies its text as peer-reviewed accepted research still subject to edits before the final version of record.[1][2]

The calculations also find temperature-dependent shifts in the transition levels of magnesium oxide and lithium fluoride. Together, the results show two ways a static calculation can be incomplete: it can misplace the boundary between preferred charge states, or miss a range in which another charge state becomes stable.[1][2]

That matters for materials screening because charge states influence carrier trapping, recombination and transport. A prediction made at absolute zero may answer a different question from the one relevant at an operating temperature. The study supplies a computational way to examine that difference; it does not report a fabricated semiconductor device or a measured improvement in its performance.[1][2]

The contribution is consequently quite specific. Machine learning helps researchers include more of the thermal physics in a calculation they already use. The resulting change in the preferred charge state is a reason to examine temperature explicitly when evaluating these defects, rather than treating a zero-temperature diagram as a complete description.[1][2]

Report an error or suggest a correction