Fig. 6

Predicted versus measured values on the held-out test set. (a,c,e) Fragmentation (\(P_{80}\)) for ANN, RF, and the ANN–RF ensemble; (b,d,f) PPV for the same models. The 1:1 line aids calibration assessment. The ensemble combines ANN’s ability to learn smooth multi-way interactions with RF’s robustness on split-friendly variables, yielding the tightest scatter about the identity line and the lowest error.