Sb-2021-08-31-all-icsd

The current structure dataset comprises 13881 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.

Sb-2021-08-31-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-27	0.018 / 0.005	36.817 / 0.2178	–	–
pair-28	0.027 / 0.007	33.949 / 0.2094	–	–
pair-31	0.050 / 0.007	32.345 / 0.2077	–	–
pair-32	0.051 / 0.010	29.729 / 0.2041	–	–
pair-33	0.069 / 0.012	26.112 / 0.1967	–	–
pair-37	0.136 / 0.018	23.481 / 0.1858	–	–
pair-38	0.171 / 0.022	22.884 / 0.1840	–	–
gtinv-300	0.228 / 0.023	12.265 / 0.1118	–	–
gtinv-235	0.281 / 0.023	11.321 / 0.1047	–	–
gtinv-175	0.376 / 0.031	11.266 / 0.1041	–	–
gtinv-240	0.393 / 0.030	10.932 / 0.1033	–	–
gtinv-312	0.472 / 0.036	10.097 / 0.1010	–	–
gtinv-190	0.523 / 0.038	9.6394 / 0.0955	predictions	mlp.lammps input log
gtinv-255	0.573 / 0.040	9.4514 / 0.0949	predictions	mlp.lammps input log
gtinv-195	0.748 / 0.053	8.5058 / 0.0934	predictions	mlp.lammps input log
gtinv-260	0.787 / 0.053	8.3115 / 0.0922	predictions	mlp.lammps input log
gtinv-265	0.983 / 0.070	7.9484 / 0.0928	predictions	mlp.lammps input log
gtinv-205	1.307 / 0.088	7.7768 / 0.0950	predictions	mlp.lammps input log
gtinv-270	1.323 / 0.091	7.5011 / 0.0942	predictions	mlp.lammps input log
gtinv-313	1.590 / 0.102	7.4776 / 0.0919	predictions	mlp.lammps input log
gtinv-191	1.600 / 0.102	7.2004 / 0.0873	predictions	mlp.lammps input log
gtinv-339	1.704 / 0.073	6.3525 / 0.0853	predictions	mlp.lammps input log
gtinv-342	1.775 / 0.083	5.7879 / 0.0832	predictions	mlp.lammps input log
gtinv-340	2.596 / 0.116	5.3475 / 0.0815	predictions	mlp.lammps input log
gtinv-343	2.986 / 0.131	4.8043 / 0.0781	predictions	mlp.lammps input log
gtinv-351	3.420 / 0.142	4.7053 / 0.0758	predictions	mlp.lammps input log
gtinv-354	3.661 / 0.154	4.4449 / 0.0749	predictions	mlp.lammps input log
gtinv-349	3.839 / 0.178	4.2193 / 0.0800	predictions	mlp.lammps input log
gtinv-350	6.441 / 0.340	4.1443 / 0.0780	predictions	mlp.lammps input log
gtinv-294	17.107 / 0.825	4.1010 / 0.0718	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.