Table 3 Performance of ten machine-learning models for predicting 28-day UHPC strength under leave-one-out cross-validation (LOOCV).

From: Monitoring of ultra-high performance concrete manufacturing for reproducible quality and waste reduction

Model

Abbr.

\(R^2\) (%)

MAE (MPa)

MAPE (%)

CS28

FS28

CS28

FS28

CS28

FS28

Multiple Linear Regression

MLR

74.54

73.27

4.91

1.51

4.45

9.41

Partial Least Squares

PLS

74.89

73.22

4.88

1.52

4.43

9.43

Kernel Ridge Regression

KRR

72.99

76.57

5.08

1.36

4.62

8.53

K-Nearest Neighbors

KNN

56.09

60.69

6.47

1.76

6.10

10.33

Support Vector Regression

SVR

72.29

77.91

5.16

1.33

4.66

8.31

Decision Tree

DT

57.26

72.90

6.43

1.46

5.87

9.03

Random Forest

RF

68.35

77.58

5.36

1.30

4.87

8.10

Gradient Boosting

GB

67.88

77.10

5.41

1.31

4.91

8.24

Extreme Gradient Boosting

XGB

68.58

77.28

5.35

1.29

4.84

8.18

Gaussian Process Regression

GPR

71.19

76.88

5.26

1.34

4.78

8.46

  1. Metrics are computed from held-out predictions across all folds: \(R^2\) (percent), mean absolute error (MAE, MPa), and mean absolute percentage error (MAPE, percent) for compressive strength (CS28) and flexural strength (FS28). Higher \(R^2\) and lower MAE / MAPE indicate better performance. Because all training data are drawn from a single UHPC recipe with controlled perturbations, models are local to that mix. Best values per column are shown in bold.
Back to article page