Fig. 1 Shapes and dimensions of tensile a and fatigue b specimens (units in mm), metallographic c and 3D graphite morphology d samples

Fig. 2 Fatigue life prediction process: search for fatigue life related features a, data feature selection based on decision tree algorithm b, fatigue life prediction of GCI by BPNN, RF and XGBoost algorithms c

Fig. 3 Schematic diagrams of the artificial neural network a and random forest b

Fig. 4 Schematic diagram of the XGBoost model structure

Fig. 5 2D morphologies of graphite in GCI observed by OM a, SEM b and XRT c

Fig. 6 Spatial structure and morphology of graphite of GCI a graphite on the surface, b 3D morphology of graphite with the sizes less than 351 µm and c 3D morphology of graphite with the sizes about 25 µm to 125 µm

Fig. 7 GCI microstructure characteristics: a graphite in the matrix, OM, b graphite shape, SEM

Fig. 8 Comparison of feature importance under decision tree algorithm

Fig. 9 Heatmap of Pearson correlation coefficients: feature correlation coefficients (upper triangle) and scatter plot (bottom triangle)

Fig. 10 ML evaluation of GCI with dataset ratio (the numbers 1, 2 and 3 indicate the 1st, 2nd and 3rd splits in the training set)

Fig. 11 ML model predictive capability based on the BPNN a, RF b and XGBoost c algorithms

Fig. 12 Comparison of the predictive performances (R2, MAE and RMSE) of different ML (BPNN, RF and XGBoost) models for GCI

Fig. 13 Evaluating the importance of features on BPNN a, RF b and XGBoost c models

Fig. 14 Predicted results based on the Basquin relation: testing data a, BPNN b, RF c, XGBoost d, comparison of the three predicted results with the testing results e, fitted curves error relationship f