Determining the Phase Diagram of a Machine-Learned Carbon Potential via Nested Sampling

We detail how the many-body potential energy landscape of a machine-learned (ML) interatomic potential for carbon [1] can be explored by utilising the nested sampling algorithm,[2] allowing for the determination of carbon's phase diagram up to high pressures. The ML potential in question is the GAP-20 model, which was developed in recent years using the gaussian approximation potential methodology to describe the properties of bulk crystalline and amorphous carbon phases with the accuracy of electronic structure methods. With the nested sampling algorithm the 3N-dimensional potential energy surface of the GAP-20 potential can be automatically (and efficiently) sampled at constant pressure, providing a set of configurations with which to test the potential's thermodynamic capabilities. As a point of comparison the phase diagrams of other interatomic potentials - including the ReaxFF, EDIP and Tersoff potentials - are also studied. Despite having been trained on optimised structures at only zero pressure, GAP-20 provides an accurate description of carbon's macroscopic properties up to approximately 200 GPa. Our results also highlight areas for future improvement of the potential.<br/><br/>[1] Rowe P, Deringer VL, Gasparotto P, Csányi G, Michaelides A. An accurate and transferable machine learning potential for carbon. J Chem Phys. 2020 Jul 21;153(3):034702.<br/><br/>[2] Pártay LB, Bartók AP, Csányi G. Efficient Sampling of Atomic Configurational Spaces. J Phys Chem B. 2010 Aug 19;114(32):10502–12.