Chain melting simulations in dense potassium from a machine-learned atomic potential

Compressed potassium forms a host-guest structure above 19 GPa, K-III, where the host and guest structures have incommensurate lattice constants c_h and c_g [1]. The guest structure comprises a linear set of chains with long range order at low temperature and has been observed to de-correlate or "melt" upon heating, while the host structure remains solid [2]. We studied this motion and onset of inter- and intra-chain de-correlation, which lead to the disappearance of guest structure diffraction peaks, by ab initio molecular dynamics (AIMD) using approximants of the incommensurate crystal. The system sizes that can be simulated with AIMD are limited, so a forcefield was trained on the AIMD data set. The trained potential was used to produce potassium's PT phase diagram up to 60 GPa and 1000 K, correctly predicting the stable solid phases, the chain melt and the full melting line [3].
[1] G. Woolman et al., Phys. Rev. Materials 2, 053604 (2018).
[2] E.E. McBride et al., Phys. Rev. B 91, 144111 (2015).
[3] H. Zong et al., submitted.