Table 2 Performance comparison of models across all crystal systems

From: Dielectric tensor prediction for inorganic materials using latent information from preferred potential

Models

DTNet

DTNet-simple

PaiNN

M3GNet

MatTen

Task

Electronic (ε∞)

Cubic

0.570 ± 0.325

0.714 ± 0.449

1.811 ± 0.900

0.971 ± 0.498

1.012 ± 0.340

Hexagonal

0.412 ± 0.272

0.430 ± 0.260

0.893 ± 0.537

0.469 ± 0.310

0.548 ± 0.274

Monoclinic

0.314 ± 0.045

0.319 ± 0.047

0.515 ± 0.115

0.382 ± 0.070

0.447 ± 0.025

Orthorhombic

0.401 ± 0.071

0.391 ± 0.088

0.749 ± 0.200

0.420 ± 0.059

0.523 ± 0.086

Tetragonal

0.268 ± 0.095

0.268 ± 0.095

0.665 ± 0.224

0.456 ± 0.191

0.415 ± 0.092

Triclinic

0.405 ± 0.179

0.404 ± 0.170

0.722 ± 0.282

0.431 ± 0.079

0.564 ± 0.214

Trigonal

0.285 ± 0.122

0.311 ± 0.133

0.737 ± 0.327

0.371 ± 0.103

0.494 ± 0.191

All

0.368 ± 0.063

0.391 ± 0.063

0.836 ± 0.125

0.489 ± 0.085

0.554 ± 0.049

Task

Ionic (ε0)

Cubic

1.400 ± 0.368

1.383 ± 0.329

3.217 ± 1.025

1.901 ± 0.321

1.819 ± 0.621

Hexagonal

1.068 ± 0.201

1.028 ± 0.121

2.168 ± 0.479

1.359 ± 0.206

1.572 ± 0.308

Monoclinic

1.679 ± 0.147

1.711 ± 0.130

2.351 ± 0.272

1.827 ± 0.175

1.982 ± 0.178

Orthorhombic

1.898 ± 0.074

1.915 ± 0.114

3.120 ± 0.171

1.957 ± 0.150

2.156 ± 0.173

Tetragonal

1.910 ± 0.305

1.867 ± 0.348

3.14 ± 0.398

2.058 ± 0.192

2.092 ± 0.265

Triclinic

1.926 ± 0.179

1.901 ± 0.144

2.523 ± 0.229

1.783 ± 0.267

2.172 ± 0.201

Trigonal

1.494 ± 0.253

1.571 ± 0.315

2.745 ± 0.600

1.742 ± 0.230

1.769 ± 0.333

All

1.677 ± 0.050

1.685 ± 0.049

2.823 ± 0.180

1.862 ± 0.131

1.969 ± 0.052

Task

Total (ε)

Cubic

1.808 ± 0.615

1.898 ± 0.621

4.834 ± 1.443

2.627 ± 0.702

2.548 ± 0.845

Hexagonal

1.409 ± 0.199

1.419 ± 0.271

3.151 ± 0.498

1.671 ± 0.557

1.954 ± 0.266

Monoclinic

1.854 ± 0.098

1.890 ± 0.148

2.784 ± 0.344

2.047 ± 0.160

2.267 ± 0.184

Orthorhombic

2.141 ± 0.089

2.178 ± 0.068

3.535 ± 0.152

2.228 ± 0.153

2.461 ± 0.180

Tetragonal

2.018 ± 0.373

2.036 ± 0.284

3.325 ± 0.544

2.403 ± 0.345

2.351 ± 0.434

Triclinic

2.142 ± 0.105

2.218 ± 0.181

3.103 ± 0.624

1.906 ± 0.284

2.488 ± 0.136

Trigonal

1.739 ± 0.269

1.624 ± 0.280

3.265 ± 0.709

1.984 ± 0.259

2.040 ± 0.371

All

1.911 ± 0.094

1.927 ± 0.104

3.427 ± 0.265

2.185 ± 0.114

2.320 ± 0.153

  1. Models are trained separately for 3 tasks: the electronic component, ionic component and total dielectric constants. The results are shown by the MAEten and standard deviation of 5 runs using different random split. The 5 trained DTNet models will be employed as ensemble models for virtual screening in the next section.
Back to article page