Ir-2021-08-05-all-icsd

The current structure dataset comprises 12652 structures generated from unique ICSD prototype structures composed of single elements with zero oxidation state. A more detailed procedure is found in Phys. Rev. B 99, 214108 (2019). The procedure to estimate interatomic potentials from the dataset is found in Phys. Rev. B 99, 214108 (2019) and Phys. Rev. B 102, 174104 (2020).

Improvement from **-dataset-10000-all-icsd

• More robust for structures with a small interatomic distance

• More robust for structures with a large interatomic distance

• More complex potential models are included.

• MLPs are estimated without using DFT stress tensors.

• MLPs are estimated by using small regression weights for energetically unstable structures.

Predictions using Pareto optimal MLPs

The cohesive energy and volume are obtained by performing a local structure optimization from the DFT equilibrium structure. In addition, the DFT equilibrium structure is obtained by optimizing a prototype structure included in ICSD, and the prototype is used as the structure legend in the figure. Therefore, the structure type of the converged structure is sometimes different from that shown in the legend even if the potential energy surface predicted by MLP is almost the same as the true one.

The other properties predicted by each Pareto optimal MLP are available from column Predictions in the following table.

Ir-2021-08-05-all-icsd shows large prediction errors. They should be carefully used. Such an MLP is often accurate for reasonable structures, but it is not accurate for unrealistic structures.

Pareto optimals

Name	Time [ms] (1core/36cores)	RMSE [meV/atom]/[eV/A]	Predictions	Files
pair-14	0.017 / 0.038	104.03 / 0.3926	–	–
pair-27	0.018 / 0.005	74.307 / 0.3350	–	–
pair-31	0.050 / 0.007	49.742 / 0.3162	–	–
pair-37	0.136 / 0.018	40.373 / 0.2937	–	–
gtinv-300	0.228 / 0.023	23.995 / 0.2160	–	–
gtinv-235	0.281 / 0.023	20.237 / 0.1993	–	–
gtinv-312	0.472 / 0.036	18.103 / 0.1985	–	–
gtinv-190	0.523 / 0.038	16.122 / 0.1864	–	–
gtinv-255	0.573 / 0.040	14.634 / 0.1788	–	–
gtinv-301	0.743 / 0.062	12.319 / 0.1695	–	–
gtinv-236	0.804 / 0.059	10.381 / 0.1576	–	–
gtinv-172	1.152 / 0.075	8.5767 / 0.1473	predictions	mlp.lammps input log
gtinv-237	1.177 / 0.073	8.4029 / 0.1472	predictions	mlp.lammps input log
gtinv-191	1.600 / 0.102	8.1721 / 0.1441	predictions	mlp.lammps input log
gtinv-256	1.780 / 0.109	7.9978 / 0.1446	predictions	mlp.lammps input log
gtinv-174	1.973 / 0.134	6.5691 / 0.1373	predictions	mlp.lammps input log
gtinv-239	2.057 / 0.133	6.4737 / 0.1372	predictions	mlp.lammps input log
gtinv-193	3.465 / 0.217	6.3446 / 0.1381	predictions	mlp.lammps input log
gtinv-258	3.556 / 0.220	6.1271 / 0.1378	predictions	mlp.lammps input log
gtinv-194	3.908 / 0.240	5.3103 / 0.1254	predictions	mlp.lammps input log
gtinv-259	4.078 / 0.244	5.1756 / 0.1253	predictions	mlp.lammps input log
gtinv-332	13.113 / 0.653	4.9751 / 0.1333	predictions	mlp.lammps input log
gtinv-223	13.159 / 0.653	4.8309 / 0.1293	predictions	mlp.lammps input log
gtinv-288	13.824 / 0.659	4.7399 / 0.1290	predictions	mlp.lammps input log
gtinv-224	14.066 / 0.687	3.7923 / 0.1180	predictions	mlp.lammps input log
gtinv-289	14.825 / 0.712	3.7658 / 0.1179	predictions	mlp.lammps input log

Column “Time” shows the time required to compute the energy and forces for 1 MD step and 1 atom, which is estimated from a simulation of 10 runs for a structure with 284 atoms using a workstation with Intel(R) Xeon(R) CPU E5-2695 v4 @ 2.10GHz. Note that the MLPs should be carefully used for extreme structures. The MLPs often return meaningless values for them.

• All Pareto optimal MLPs are available here.