Table 1 Comparison of the mean absolute errors (MAEs) on predicting 50 ns MD simulated properties between different approaches.

From: Accelerating amorphous polymer electrolyte screening by learning to reduce errors in molecular dynamics simulated properties

Method

σ

σ*

DLi

\({D}_{{{{{{{{\rm{Li}}}}}}}}}^{* }\)

DTFSI

\({D}_{{{{{{{{\rm{TFSI}}}}}}}}}^{* }\)

DPoly

\({D}_{{{{{{{{\rm{Poly}}}}}}}}}^{* }\)

5 ns (direct)

0.528

0.278

0.503

0.419

0.455

0.249

0.612

0.528

5 ns (linear)

0.152

0.275

0.148

0.247

0.096

0.297

0.072

0.110

GCN CV

0.093 ± 0.017

0.186 ± 0.053

0.106 ± 0.016

0.209 ± 0.050

0.101 ± 0.020

0.181 ± 0.028

0.072 ± 0.019

0.114 ± 0.030

GCN average

0.076

0.182

0.080

0.202

0.075

0.171

0.056

0.104

  1. The first row denotes the MAE between 5 and 50 ns simulated properties. For each property, interpolation and extrapolation performance are represented by labels without and with the * symbol. Uncertainties are the standard deviations of MAEs from tenfold cross-validation (CV).
Back to article page