Table 2.
Computational cost.
| Model | (2a) 2nd order IFC | (3a) velocity operator | (2b) 3rd order IFC | (3b) linewidths | ||
|---|---|---|---|---|---|---|
| method (# of displacements) | kCPUh | kCPUh | displacements | kCPUh | kCPUh | |
| 108(D) | DFPT 2 × 2 × 2 q mesh | 12 | 0.00005 (5 × 5 × 5 q mesh) | 12348 (3 NN) | 9 | 0.2 (q = 0, 2 cm−1) |
| ’ | ’ | ’ | ’ | 18972 (6 NN) | 13 | 81 (5 × 5 × 5, thm) |
| ’ | ’ | ’ | ’ | 29592 (10 NN) | 21 | 0.2 (q = 0, 2 cm−1) |
| ’ | ’ | ’ | ’ | 53352 (20 NN) | 37 | 0.2 (q = 0, 2 cm−1) |
| 144(D) | DFPT 2 × 2 × 2 q mesh | 44 | 0.0001 (5 × 5 × 5 q mesh) | 26028 (6 NN) | 24 | 0.5 (q = 0, 2 cm−1) |
| 192(D) | DFPT 2 × 2 × 2 q mesh | 102 | 0.0005 (3 × 3 × 3 q mesh) | 33300 (6 NN) | 50 | 2.1 (q = 0, 2 cm−1) |
| 192(D) | BKS supercell 2 × 2 × 2 (1152) | 0.003 | 0.0005 (3 × 3 × 3 q mesh) | 32472 (6 NN) | 0.1 | 2.1 (q = 0, 2 cm−1) |
| 192(D) | GAP supercell 2 × 2 × 2 (1152) | 0.007 | 0.0005 (3 × 3 × 3 q mesh) | 33372 (6 NN) | 0.2 | 2.1 (q = 0, 2 cm−1) |
| 192(G) | GAP supercell 2 × 2 × 2 (1152) | 0.007 | 0.0005 (3 × 3 × 3 q mesh) | 34272 (6 NN) | 0.2 | 2.1 (q = 0, 2 cm−1) |
| ’ | ’ | ’ | ’ | ’ | ’ | 1.2 (q = 0, 0.4 cm−1) |
| 1536(G) | GAP supercell 1 × 1 × 1 (9216) | 0.007 | 0.1 (3 × 3 × 3 q mesh) | – | – | – |
| 5184(G) | GAP supercell 1 × 1 × 1 (31104) | 1.5 | 2.1 (3 × 3 × 3 q mesh) | – | – | – |
The calculations for the 108(D) and 144(D) models were performed on the Skylake nodes of the SCITAS High Performance Computing facility at the École Polytechnique Fédérale de Lausanne. The calculations for the 192(D), 192(G), 1536(G) and 5184(G) models were performed on the Icelake nodes of the Cambridge Service for Data-Driven Discovery. DFT and DFPT calculations were performed using Quantum ESPRESSO115,142, BKS- and GAP-based calculations were performed using LAMMPS132 (using the quip133–135 interface for GAP). Velocity operator and linewidths were computed using Phono3py91. In column 2b, the cutoff used for the third-order force constants (in nearest neighbor units) is reported in parentheses. In column 3b, ‘thm’ refers to the tetrahedron method91, 2 cm−1 or 0.4 cm−1 denote the Gaussian smearing used for the Dirac delta appearing in the linewidth expression. The third-order force constants for the 1536- and 5184-atom models have not been computed, since the subsequent linewidth calculations would have had a prohibitively large computational cost (see Fig. 17). ‘kCPUh’ means 1000 CPU (core) hours.