Table 5 Performance metrics for UQ models for yield strength and ultimate tensile strength, given as average value ± standard deviation (in subscript). Best performance metrics are highlighted in bold.
From: Empirical validation of size effects in sub-sized tensile specimens for nuclear structural materials
Model | Yield strength | Ultimate tensile strength | ||||
|---|---|---|---|---|---|---|
R2 ↑ | RMSE ↓ | Coverage (%) ↑ | R2 ↑ | RMSE ↓ | Coverage (%) ↑ | |
Quantile regression | 0.909± 0.04 | 62.9± 13.5 | 83.8± 7.3 | 0.830± 0.10 | 54.3± 15.6 | 86.1± 4.1 |
Natural grad boosting | 0.966± 0.01 | 38.4± 5.6 | 90.2± 4.1 | 0.813± 0.05 | 58.4± 10.3 | 91.9± 8.2 |
Gaussian process regression | 0.921± 0.04 | 58.3± 13.7 | 90.7± 4.2 | 0.855± 0.05 | 51.1± 8.5 | 92.4± 5.9 |
Deep ensemble | 0.956± 0.03 | 41.4± 15.9 | 53.9± 6.5 | 0.844± 0.05 | 52.8± 6.8 | 60.1± 11.1 |
MC dropout | 0.914± 0.04 | 61.6± 12.9 | 85.3± 6.1 | 0.844± 0.04 | 53.0± 8.0 | 90.9± 2.7 |
Bayesian NNs—VarInf | 0.891± 0.04 | 69.5± 13.8 | 91.7± 4.8 | 0.712± 0.08 | 72.5± 12.2 | 93.8± 5.1 |
Bayesian NNs—MCMC | 0.965± 0.02 | 38.0± 9.7 | 94.6± 2.8 | 0.839± 0.04 | 53.9± 6.7 | 95.2± 3.7 |
