Table 5 UQ statistics for the formation energetics of surfaces in a BCC crystal (c.f., Fig. 8)

From: Uncertainty quantification for misspecified machine learned interatomic potentials

Structure

MLE Error

Std

Low Error

High Error

100

−1.19 × 100

1.32 × 102

−4.06 × 102

4.15 × 102

110

8.88 × 101

3.61 × 102

−2.72 × 102

1.19 × 103

111

1.35 × 102

3.02 × 102

−5.41 × 102

9.17 × 102

210

5.72 × 101

2.71 × 102

−7.48 × 102

7.09 × 102

211

9.14 × 101

2.95 × 102

−1.04 × 103

4.14 × 102

310

1.19 × 102

2.74 × 102

−7.76 × 102

6.82 × 102

311

8.27 × 101

2.57 × 102

−6.92 × 102

7.65 × 102

320

4.45 × 101

2.47 × 102

−6.26 × 102

8.31 × 102

321

7.56 × 101

2.43 × 102

−5.21 × 102

9.37 × 102

322

9.52 × 100

2.37 × 102

−6.39 × 102

8.18 × 102

332

2.93 × 101

2.30 × 102

−6.16 × 102

8.41 × 102

\({\rm{Mean}}\,\frac{{\rm{Error}}}{{\rm{Std}}}\)

0.24

\({\rm{Mean}}\,\frac{{\rm{Error}}}{{\rm{High}}-{\rm{Low}}}\)

0.045

  1. MLE error corresponds to the errors of the MLE relative to the DFT reference, Std is the standard deviation of the predictions of \({\pi }_{{\mathcal{H}}}^{* }\), Low Error is the difference between the smallest prediction in \({\pi }_{{\mathcal{H}}}^{* }\) and the DFT reference, and High Error is the difference between the largest prediction in \({\pi }_{{\mathcal{H}}}^{* }\) and the DFT reference. A negative Low Error and a positive High Error indicate that the predictions from the UQ ensemble bracket the reference value.